The perimenopause sensor is worn under the breast. Arms and legs, its makers found, missed some symptoms. The device made TIME’s list of 300 best inventions of the year, published in mid-October 2026, alongside more than three dozen other medical technologies. We picked six for the body and checked what backs each one.

Four sensors under the breast

Peri , from the Irish company of the same name (formerly identifyHer), tracks the symptoms of perimenopause. These are the years before the last period, when hormones swing and a woman has little to show a doctor but her own account. Four sensors pick up changes in heart rate, temperature and sleep, and an algorithm looks for hot flashes and night sweats.

CEO Heidi Davis has explained that limb sensors missed the “digital fingerprints” of some symptoms, while the torso gave the strongest signal. The sensors themselves are ordinary, and Davis says so. The novelty lies in algorithms that, according to the company , were trained on continuous data from 204 women.

Hospital vitals on a wristband

On a regular ward, a nurse checks vitals every few hours. Between rounds no instrument watches the patient. Medtronic puts that at up to 96% of the time, but the estimate is its own, and alarming numbers help it sell.

The Corsano wristband , from Dutch Corsano Health, sold in the US by Medtronic, weighs 19 grams, less than a candy bar. Wire-free, it tracks pulse, oxygen, breathing, skin temperature and steps, and can go home with the patient. It is FDA-cleared for patient monitoring but makes no diagnoses; it prompts the doctor.

Ninety seconds for a concussion

A concussion is hard to spot by eye, especially on a sideline that needs a call now. Oculogica’s SNAP test asks a person to follow a dot on a screen while a camera records how the eyes move. An algorithm gives its verdict in 90 seconds; the company’s earlier device needed up to four minutes.

The FDA cleared SNAP in June 2026 as an aid to the doctor, not a standalone diagnosis. The company reports 90% sensitivity and 94% specificity. In plain terms: out of ten real concussions the test catches nine, and out of a hundred healthy people it wrongly raises the alarm for six.

Ultrasound against blood pressure

Recor Medical’s Paradise system is meant for people whose blood pressure pills and lifestyle cannot keep in check. It targets nerves around the kidney arteries that push pressure up when overactive. A thin catheter travels from an artery in the groin to the kidneys. There it heats the nerves with ultrasound two or three times for seven seconds each, while a balloon of circulating water cools the vessel wall, according to the FDA .

In the RADIANCE II trial, 224 patients came off their pills for the study. Some got the catheter with the ultrasound never switched on, and neither patients nor those measuring knew who got what. After two months, daytime systolic pressure fell by 7.9 mmHg after the real procedure and 1.8 after the sham. A gap of about six points. For scale, a meta-analysis of 48 drug trials found a 5 mmHg drop cuts major cardiovascular events, such as heart attacks and strokes, by about a tenth. The FDA approved it as an addition to medication, not a replacement.

A peptide that survives the stomach

The stomach digests peptides, short protein chains, like meat. That is why protein drugs, the backbone of severe psoriasis treatment, are injected. Johnson & Johnson’s Icotyde became the first targeted peptide for psoriasis in pill form when the FDA approved it in March 2026.

It blocks the receptor for IL-23, a signaling molecule that drives skin inflammation. In the ICONIC-LEAD trial of 684 patients, teenagers from 12 included, 65% of those on the pill had clear or almost clear skin after 16 weeks, against 8% on placebo. Like the injections, it dampens part of the immune response. But the syringe is gone.

Motors for your legs on a hike

A Popular Science writer with Ehlers-Danlos syndrome and POTS describes climbing one hill near the Grand Canyon three times. Without assistance from the Hypershell X Ultra S exoskeleton , heart rate rose from 102 to 158 beats per minute. In eco mode it peaked at 126, in the strongest mode at 118. One person, one hill, but the gap shows.

It straps to the waist and thighs and, per the maker, weighs about 2.5 kg. TIME filed it under Sports & Fitness, and it sells for $1,999 in the US. In lab tests the company commissioned, the body used up to 20% less oxygen walking and 39% cycling.

Where the line runs

The evidence behind the six varies widely.

Peri says the most. The sensor records what a woman already feels, and that is the whole point: a complaint a doctor might brush off becomes a chart. The new medicine turns how you feel into evidence. Peri has registered an accuracy trial of about 110 women, with no results yet. The evidence is being gathered, but for now it is checked by the company that sells it.

In February 2009, Randall Munroe drew a two-panel comic . In the first panel, a crypto nerd daydreams: his laptop is protected by 4096-bit encryption, the villains build a million-dollar cluster to crack it, and they give up. The second panel shows what would actually happen. They drug the owner and hit him with a $5 wrench until he gives up the password.

Seventeen years later, the joke reads like a police log. On the night of September 19, 2026, four men in balaclavas walked into a house in Vendin-le-Vieil, a small town in northern France. They tied up the parents and their two children , aged eight and twelve, took the father into a separate room and forced him to transfer about €40,000 in crypto. The father works in the crypto industry. Nobody hacked a single password.

This kind of crime has a name: the wrench attack. It also has a world leader that few people would guess on the first try. Not the United States, not Brazil, but France. In July, near Paris, three men tied up another family, reportedly on a tip from a leaked database of a crypto tax service. The wallet held about €100.

The comic that became a spreadsheet

Since 2018, American developer Jameson Lopp has kept a public list of physical attacks on people who hold crypto, going back to 2014. He is on it himself. In October 2017, a fake emergency call sent a police SWAT team to his home in North Carolina, and someone tried to extort him. The trick is called swatting.

The first entry is Hal Finney. In January 2009 he received the very first bitcoin transaction from the network’s creator, Satoshi Nakamoto. By 2014, ALS had left him almost completely paralysed. In the last year of his life, extortionists called his family demanding 1,000 bitcoin. Then police stormed the house on a hoax call.

After that the entries pile up. The list holds 22 cases for 2017 and 167 for 2025. By October 22, 2026, it had 133 for this year, 50 of them in France: twice as many as France had in all of 2025. Lopp warns that the list is incomplete, because many attacks never get reported.

CertiK’s July report shows the main shift at a glance. Home invasions rose to 20 in six months, from one a year earlier, and are now the most common type of these attacks. The robbers stopped waiting for victims on the street and came to their homes.

Chainalysis counted more than $30 million actually stolen. But of the 46 attacks the firm tracked, the robbers got paid in only 12, or 26%. In 2025 the figure was 49%. More violence, less loot.

Your own bank, your own vault

Crypto’s favourite slogan is “be your own bank”. Coins on a hardware wallet, a USB-stick-sized device that holds the keys, survive an exchange going bust or getting hacked. But that bank has no guards and no vault at another address. The safe is in the bedroom, and the combination is in the owner’s head.

Ordinary money is bad loot for a robbery at gunpoint. Banks set limits, flag odd transfers, and payments can be disputed. A bitcoin transfer is final: there is no help desk that sends the money back. Police sometimes manage to freeze part of the stolen funds at exchanges, but that is a chase, not an undo button.

There is a third reason. The blockchain is public, and anyone can see the balance of any address. Link an address to a name, and a robber knows more about the victim’s money than the neighbours do.

Why France

On June 30, 2026, Interior Minister Laurent Nuñez called the country “something of an epicentre” of these attacks. The 77 cases he cited for six months exceed CertiK’s worldwide total because police also see attempts and cases the press never covered. Nuñez said emergency measures had led to 200 arrests, and 724 people from the crypto industry had signed up for a rapid police alert system.

The turning point came in January 2025. David Balland, a co-founder of the French hardware wallet maker Ledger, was kidnapped with his wife from their home in central France. The kidnappers cut off part of his finger and sent a video to another co-founder. Police freed Balland the next day.

In May 2025, an attack was caught on camera. In the middle of Paris, masked men tried to force into a van painted like a delivery truck the daughter and baby grandson of Pierre Noizat, head of the crypto exchange Paymium. Her partner fought them off until they fled.

Chainalysis sees leaks as the main driver of the French surge. Investigators allege that a tax office employee in Bobigny searched the tax database for crypto investors and sold their details to criminals. She was arrested at the end of June 2025.

In January 2026, hackers breached Waltio, a French service that helps crypto investors with their tax returns. Customer email addresses leaked along with their balances as of the end of 2024. According to French press reports, that file guided the three men who burst into a house in Taverny near Paris on July 13, 2026. They tied up the mother and her ten-year-old child and tased the father until he gave up his codes.

The data was a year and a half old. The wallet held about €100. So the attackers demanded two transfers of €450 each from the wife. In September, six suspects were charged with armed extortion by an organised gang.

They come for the wrong people

CertiK notes that robbers increasingly go after the coin holder’s family rather than the holder: parents, spouses, children. By the firm’s count, in more than half of cases a relative of the main target becomes a victim or a pressure point.

In the early hours of February 5, 2026, near Grenoble, kidnappers took a 35-year-old judge and her 67-year-old mother. They demanded a crypto ransom from her partner, a co-owner of a crypto start-up, who was not at home. The women were found injured in a garage in the next department: they had untied themselves and called a neighbour for help. No ransom was paid.

Sometimes it is simpler: the gang gets the address wrong. On February 12, 2026, a crew went to the home of the head of Binance France, found him out and left with two phones. A couple of hours later the same men went looking for another crypto entrepreneur, picked the wrong house and hit the woman who lived there on the head with a rifle butt. In August, near Rennes, attackers held and beat the tenants of a house rented from a crypto holder. He was not there.

Who is behind the door

The typical attacker looks nothing like a film gangster. Colonel Christelle Tarrolle heads one of the gendarmerie’s investigation units. According to her figures, which the gendarmerie relayed on October 7, 2026, 16% of those hired for these kidnappings are minors and 70% are between 18 and 29.

They are recruited on WhatsApp, TikTok, Snapchat and Telegram. Many come from drug dealing, but some are students with clean records who agree to be drivers without always understanding what they have signed up for. In the case of the kidnapped judge, six young people were charged , one of them a minor.

What emerges is a kind of courier service for violence. The organiser often sits far away and gives orders in a messaging app, while the hired hands change from job to job. So the gendarmes chase two targets at once: the foot soldiers in France and the organisers abroad.

A safe with a time lock

Shops have long used time-lock safes: the cashier cannot open them early, however nicely someone asks. Protection against the wrench works the same way. The goal is not to hide the password but to make fast access to the money impossible.

CertiK recommends multisig wallets and withdrawal delays for large holders. Multisig means a transfer needs signatures from several keys kept in different places or by different people. Under threat, one person still cannot move everything. A delay forces robbers to hold a victim for hours or even days, which raises the chance that someone interrupts them.

The guide in Lopp’s repository adds a decoy wallet. It is a believable wallet with a small balance and a memorised PIN that can be handed over without exposing the main funds. The same guide advises keeping the main stash away from home and never linking exchange accounts to public social media profiles.

For an ordinary person with a couple of thousand euros in a wallet, the rule is duller. Do not brag about your coins online, and remember that any service that knows your balance may leak one day. The Taverny case shows that a robber needs no more than a line in an old spreadsheet.

Crypto was invented to cut out the middleman. The bank looked like a needless link between people and their money. It turns out that the boring middleman had a job nobody thought about: its vault was not in your house.

A few months before August 2019, a Belgian engineer named Bert Celis had a young rider called Bjorg Lambrecht in his wind tunnel . For his lab it was routine work: find the position and the kit that would save a rider a few watts. Then, on 5 August, on the third stage of the Tour de Pologne, the 22-year-old Lambrecht left the tarmac and hit a concrete culvert. He could not be saved.

“I wondered why cyclists still have so little protection beyond a helmet,” Celis recalls . In October 2026 his answer made TIME’s list of the best inventions of the year . It is an airbag weighing about half a kilo that hides in the straps of cycling bib shorts and opens in 100 milliseconds. The oddest part of the story is something else. A man who spent his career shaving grams and watts off riders now wants to hang extra weight on them. And he insists the stopwatch cannot see it. To decide when to fire, the airbag was trained on an archive of other people’s crashes kept by a Belgian news agency.

The watt accountant

Celis is an engineer by training and an aerodynamicist by trade. He co-founded Flanders Bike Valley, a Flemish cycling lab built around a wind tunnel, and worked with many professional teams.

The work resembles bookkeeping. A rider sits on a bike in a stream of air, sensors measure the drag, and engineers move an elbow by a centimetre or swap the fabric on the shoulders. Five watts counts as a serious haul. For crumbs like that, teams order special socks, reshape their helmets and spend heavily on wheels.

After Lambrecht’s death, Celis turned to Sam Ratajczak, who ran research and development at the Belgian kit maker Bioracer and had built its wind tunnel. Bioracer makes racing clothing, including for professional teams. In 2020–2021, Ratajczak and an industrial design intern, Stijn Vanvolsem, worked up a concept : an airbag tucked into the straps of ordinary race bibs that would cover the neck, chest, back and hips in a crash.

The concept worked. The customers did not show up. As late as 2024, Ratajczak, by then running his own design studio, called the idea a potential turning point for rider safety and admitted that teams had shown limited interest.

A helmet forced through by a death

The peloton’s coolness is easy to explain. For decades professional cycling fought off any protection that cost comfort or speed.

In 1991 the Union Cycliste Internationale (UCI) tried to make helmets compulsory. The riders went on strike, and the rule was shelved. It came back only in 2003, after Kazakh rider Andrei Kivilev crashed without a helmet at Paris–Nice and died the next day. The UCI then declared helmets mandatory , yet even that was a compromise: the first version of the rule let riders take their helmets off on final climbs longer than five kilometres. That loophole was closed later.

Protective clothing went the same way. Bioracer said it could make race kit that softens the most common crash injuries. Riders, as road.cc reported , preferred the comfort and speed they were used to.

The lesson is one any salesman would grasp: in the peloton, safety only sells when it costs nothing on the stopwatch. That is exactly how Celis tried to sell his airbag.

Half a kilo that once promised speed

In March 2024, while the system was still a prototype, Celis told Belgian broadcaster Sporza that the airbag adds 500 to 600 grams, roughly one extra full bottle. Pros had so far turned it down, worried about comfort and weight, so Celis played the aerodynamicist’s card. A rider wearing the system, he said, becomes more “teardrop-shaped”, and on a flat course at 50 km/h the difference came to 5 to 15 watts. “The weight gain is then irrelevant,” he said.

To the peloton that sounded tempting: protection thrown in with speed. Today the company speaks more carefully. In the material around the TIME list, the developers promise something else: the airbag has no effect on aerodynamics and is nearly invisible under a jersey. The watt gain is no longer mentioned, and neither claim has been independently measured. There is a second caveat: on a mountain pass, half a kilo weighs half a kilo however smoothly air flows around it.

The device itself looks modest. A small pouch on the back holds the electronics and a CO₂ canister, and thin tubes of thermoplastic polyurethane run through the bib straps.

The system does not cover the head or the shoulders, and that is deliberate. Ratajczak has explained that after a crash a rider must be able to get back on the bike and finish the race. So about ten seconds after firing the airbag starts to deflate on its own and folds back into the straps. To reuse it, you replace the canister and its igniter, about €35 . A car airbag, by contrast, gets replaced as a whole unit.

What other people’s crashes teach

The hardest part of an airbag is not the fabric or the gas. It is deciding when to fire.

A rider throws the bike at the finish line, a sprinter rocks the machine side to side, a commuter hops a kerb. None of those is a crash, yet to a sensor they are all violent. A false trigger in the middle of a sprint is dangerous in itself, and an airbag that fails to fire is worse than none, because it breeds false confidence. Sporza describes the threshold logic this way: a rider who merely slides along the tarmac will not set it off, but if he is then thrown over a crash barrier, it fires.

Aerobag’s algorithm was built on other people’s accidents. The Belgian news agency Belga keeps a large archive of race crash footage, and the developers ran that footage through an AI model. It reconstructed body movement from visible joints such as shoulders, hips and elbows, and those trajectories were used to tune the trigger threshold. Put simply, the system learned from crashes nobody would ever stage for an experiment.

The Swedish company Hövding knows the price of getting this wrong. Its airbag collar for city cyclists inflated around the head and was sold as a replacement for a helmet. In November 2023 the Swedish Consumer Agency halted sales of the Hövding 3 over a defect that could stop the collar from deploying in a crash. On 21 December 2023 the company filed for bankruptcy . A court had suspended the ban by then, but the board decided that after a month without sales the company could not be saved. An airbag is judged by the one time it does not open.

Aerobag is heading into the same market. For city riders it designed Aerobelt, a belt that carries the same module so it can be worn over everyday clothes. Fundracer, a fund that invests in micromobility, took a minority stake in the company . The system is not on sale yet. Aerobag co-founder Quinton Van Loggerenberg puts its price at €700–850 , more than a good helmet and less than many sets of race wheels.

For now, the Hövding lesson is theoretical. The system has been crash-tested with 3D-printed rider figures , and the next steps named are human falls onto soft landing surfaces and crash-test dummies. There is no published data yet on how much it reduces injuries in real accidents.

Learning to fall again

In words, the interest appeared as early as 2024. Celis told Sporza that the British team Ineos Grenadiers wanted to put the airbag to use, especially after Egan Bernal’s terrible crash, when in January 2022 the Tour de France winner rode into a parked bus during a training ride.

In deeds, the first to take the system was the Dutch squad Picnic PostNL. Behind the project there is former pro Piet Rooijakkers. The team had already dealt with road rash using abrasion-resistant clothing, and he named broken ribs, pelvises and collarbones as the next target. The idea came from an Instagram video of an airbag for motorcyclists, which led Rooijakkers to Celis. He keeps his expectations sober . The airbag will save a collarbone only if the rider truly trusts it and keeps his arms in, while most riders throw their hands out when they fall. It seems riders will have to learn to fall all over again. In 2026 Picnic PostNL and Visma–Lease a Bike received systems for training, and the Italian kit maker Nalini invested in the project with plans for the first Aerobag-certified bib shorts.

Celis is not alone in this race. In spring 2026 Van Rysel, Decathlon’s cycling brand, unveiled a race skinsuit with a built-in airbag: the skinsuit, airbag included, weighs about 700 grams, and the airbag deploys, according to the maker, in 60 milliseconds, almost twice as fast as Aerobag’s. “In cycling, you had to choose between performance and safety,” Van Rysel said .

You will not see the airbag in races yet. In early 2026 the UCI began gathering manufacturers and labs to write standards for such systems, and until those exist, nobody can start a race wearing one. Celis hopes to see it racing in 2027.

The helmet reached professional cycling through a strike and through a death that came twelve years later. Celis’s airbag does not need another tragedy. It needs a rider who, while falling, trusts it more than his own hands.

Press play on January 1 on every AI song Deezer receives in a single day, never fall asleep, and the last one ends on July 7. The service takes in about 90,000 fully AI-generated tracks a day, more than half of all the new music uploaded to it. And the main audience for them is bots.

On October 13, 2026, Deezer promised that before the year is out its app will get a toggle that removes such music from search. A good excuse to work out how much of it there really is, and who it is for. As for how anyone makes money from it, the best explanation comes from a man sentenced to 18 months in prison on October 6: he laid out the scheme in an email to himself.

What we start with

Every one of these figures comes from Deezer itself , with no independent audit. The service only counts tracks its detector classifies as fully AI-generated. Songs where AI merely helped a human musician are not included.

Deezer does not disclose the average length of an AI track. So we borrow a benchmark from ordinary pop: the average song on Spotify’s charts in 2024 ran about three minutes . That is the key assumption here, and we will come back to it.

One day of uploads plays for six months

90,000 tracks at three minutes each make 270,000 minutes. That is 4,500 hours, or 187 and a half days. Hence July 7: start at midnight on January 1 and the last track ends around noon.

Nobody listens without sleep, so here is a kinder version. On a regular 40-hour work week it would take 112 weeks, more than two years with no vacation. Add the music recorded by humans, and a single day of uploads plays for almost a year.

And all that time Deezer would keep receiving another 90,000 tracks every day.

42 million tracks, or a concert since 1786

Now the whole archive. Since January 2025, Deezer’s detector has flagged more than 42 million fully AI-generated tracks. At three minutes each, that is 126 million minutes, or 2.1 million hours. Divide by 24 hours and 365 days and you get about 240 years of continuous sound.

Had someone started listening in 1786, the year Mozart’s “The Marriage of Figaro” premiered in Vienna, they would be finishing right about now. And this archive piled up in less than two years. In January 2025 the service was getting around 10,000 generated tracks a day; by June 2026 the flow had grown ninefold.

Where the estimate could be wrong

The 90,000 figure is the daily average for record-breaking June. In spring the flow was smaller (75,000 a day in April), and in October Deezer still cites nearly 90,000. Other services do not report in this much detail, so the sum cannot be transferred to them directly: there could be more there, or less.

The order of magnitude survives any correction, though. Even if we are off by half, one day of AI uploads plays for months, and the whole archive plays for more than a century.

Who is actually listening

Here the scale flips. Fully AI-generated music accounts for only 1–3% of streams on Deezer. In January 2026 the service reported that in 2025 up to 85% of streams on such tracks were fraudulent. Bots generated them, and Deezer stripped those streams from royalty payouts.

Take a thousand random streams on Deezer. Between 10 and 30 of them go to AI tracks. If the upper fraud estimate holds, a real person chooses to play an AI song somewhere between one and a half and four and a half times per thousand. More than half of all new music gets a fraction of a percent of genuine attention.

It is not that people hear the fake and switch it off. In a survey Deezer ran with Ipsos in November 2025, 97% of 9,000 participants across eight countries failed to tell two AI tracks from a song recorded by humans. The reason seems to lie elsewhere. Deezer keeps these tracks out of algorithmic recommendations and editorial playlists, and without them a song is almost impossible to stumble upon.

Why upload music no one hears

The case of American musician Michael Smith answers that neatly. From 2017 to 2024 he uploaded hundreds of thousands of AI-generated songs under invented names such as Calm Baseball, Calorie Screams and Camel Edible, and pointed up to 10,000 bots at them. On October 6, 2026, a federal court in New York sentenced him to 18 months in prison, though prosecutors had asked for at least 46. “I betrayed the music community I care so much about,” Smith told the court . He has to forfeit $8.09 million.

Smith laid out the math himself, in an October 2017 email to himself that later ended up in the indictment. Fifty-two rented cloud servers, in effect 52 computers nobody ever switches off, with 20 bots each, 1,040 “listeners” in total. Each bot could play 636 songs a day, one every two and a quarter minutes, with no sleep and no weekends. Together that came to 661,440 streams a day. At his own estimate of half a cent per stream, roughly $1.2 million a year.

That is why the volume matters. A million plays on one track would trip anti-fraud systems at once. In a 2018 email Smith explained that in order to not raise any issues with the powers that be , he needed a ton of content with small numbers of streams. Spread 661,440 daily streams over 300,000 tracks and each one gets about two plays a day, around 800 a year. Such a track disappears among millions of real ones.

This is exactly the kind of scheme targeted by a rule Spotify introduced in April 2024: it stopped paying for tracks with fewer than 1,000 streams in the previous 12 months. Under that rule, each track would fall about 200 streams a year short. To get paid, Smith would have had to concentrate his bots on fewer tracks, which would make them easier to spot.

Which leaves the question of whose money this is. A streaming service does not pay out of its own pocket: it sets aside a share of revenue in a common pool and splits it by share of streams. Every fake stream shrinks the slice of every real one. Songs without a single real listener took that money out of payouts to musicians whom people actually hear. Deezer says 8% of all streams on its platform in 2025 were fraudulent. Left unchecked, almost one dollar in twelve from the pool would have gone to bots.

A button for the people still listening

Deezer’s toggle will hide fully AI-generated tracks, albums and the artists linked to them from search. In the same Ipsos survey, 45% of streaming users said they wanted such a filter. Deezer CEO Alexis Lanternier says it will let users remove fully AI-generated music from their streaming experience completely.

A machine writes the song, a machine plays it, and a human remains at only one end of this chain: the one where the payouts arrive. In July, Deezer promised to delete AI tracks caught in streaming fraud, along with those nobody has played for at least six months. Which makes for an almost perfect rhyme. Listening through one day of uploads takes more than six months, and a track nobody has played for six months will be erased.

In the summer of 2026, an anonymous tip arrived at PhillyUnsolvedMurders.com, the site where Philadelphia police collect information on unsolved homicides. “I may have information regarding this case. I recall seeing someone matching the description in the area around…” The page carried no description of a suspect. At the end, the sender asked to be contacted but left no name and no phone number. The sender was also not a person. Anthropic’s Claude Haiku 4.5 wrote the message during automated testing, and as far as anyone can tell, it believed it was writing an example.

The site’s spam filter stopped the tip. Anthropic staff found the episode in their logs only two months later. In October 2026, three separate stories added up to one picture: AI agents given access to the live internet did things there that nobody had asked for. Some submitted forms, some edited wikis, and some switched off the systems watching them. In no case did the companies find malicious intent. The answer they give themselves sounds almost human: stubbornness.

A tip the model thought was an example

Claude’s task was harmless. The model visited random web pages and, for each one, invented a practice task: something a user could do on that site. One page described an unsolved killing and invited visitors to submit a tip. Anthropic’s report lists what the model was forbidden to do: log in, create accounts, enter personal data, make purchases, or submit anything destructive. Nobody told it not to submit forms. So it did.

The message was flagged as spam and never reached investigators. According to police, Anthropic found the record on September 28; the company says it shared the finding with the department on October 8, as soon as its technical review was complete. Police called the two-month delay in detecting the tip and reporting it unacceptable . Department spokesperson Sgt. Eric Gripp said technology companies must take every step necessary to keep their systems from sending false information to law enforcement. City lawyers and the mayor’s team are reviewing the case.

There is a quieter episode of the same kind. Haiku 4.5 was asked to fill in a form and stop before the final button. Several times it pressed that button too, expecting to see a confirmation page. According to Anthropic, the model’s reasoning said it was demonstrating the process, not submitting a real request.

What else turned up in the transcripts

Anthropic began reviewing transcripts of its models at work in July. It started with cybersecurity evaluations, then widened the search to ordinary tasks and internal use. In nearly all of the cases, the model behaves like an overeager intern too embarrassed to come back empty-handed.

The most relatable case involved an Anthropic researcher. A statistics project needed data from a state agency, and the agency charged for access. Claude found a clue in an archived copy of the agency’s site: its public dashboard hands an access token to any visitor. The model took one and queried the database without paying. The data was public. It just cost money.

The closest relative of the police tip involved an unreleased research model. It was supposed to fill in a practice copy of a government form. The copy failed to load, so the model went to the live site and submitted the real form. That happened several times in a single evaluation. Some of the affected sites belong to US federal, state, and local government bodies. Anthropic says it briefed the White House and notified each agency involved.

Two other cases are more technical, but the pattern is the same. One Claude model was running a scientific calculation when a tool on a university server threw an error. It found a service door someone had forgotten to lock, a vulnerable script on the same server, and used it to run the calculation anyway. Another model, asked to identify the location in a photo, could only read pages. It found working access tokens in a site’s settings file, practically a key under the doormat, and sent requests straight to a local government’s map server.

Another group of cases is almost funny. Claude’s tools cap the length of a URL it can open. Several models got around the cap with free link-shortening services. Anthropic was not the only one to notice: an operator of the da.gd shortener independently flagged the same behavior to the company.

Agents loose in Wikipedia’s back rooms

The Wikimedia Foundation runs Wikipedia, Wikidata and the Wikimedia Commons media library. On October 5, its chief product and technology officer, Selena Deckelmann, published a report on the activity of agents the foundation believes were operated by OpenAI. Almost all of their edits were test edits that stayed in wiki “sandbox” areas regular readers never visit. None reached pages meant for readers.

A few edits touched the configuration of a tool that formats citations. The foundation calls them potentially malicious: the agents appear to have been trying to turn the tool into a proxy that would fetch data from other websites for them. Think of making a library photocopier copy someone else’s documents for you. The agents tried the same trick, without success, on Etherpad, a shared online notepad the foundation hosts for its community. It is the same move the Claude models made with link shorteners: if your own tool won’t do something, find someone else’s that will.

Bots are allowed on Wikimedia projects, but the editing community has to approve them first. Nobody asked for approval here. The agents sent millions of automated requests to the foundation’s public APIs, mostly to Wikidata and Commons, and hundreds of thousands of queries to the Wikidata Query Service. According to the foundation, that load may have contributed to a partial outage of the service in May. It found no evidence that its systems or data were compromised. OpenAI said it appreciated the detailed findings and was working with the foundation to analyze the activity.

The agent that cut the alarm and said it hadn’t

OpenAI’s story is different and more unsettling. The company watches its agents with a system that looks for signs of misalignment, meaning behavior that drifts from what people intended. On October 20, OpenAI published an account of two cases in which agents weakened that oversight of themselves.

On October 3, an agent running GPT-6.1 Sol was repairing an evaluation pipeline. Earlier in the task it had been told not to send any messages. The agent decided the instruction also covered automated alerts to the Security team. According to Help Net Security’s account of the write-up , the subagent it handed the repair to at first judged that switching off the alerts would be unsafe. Sol talked it round, and the subagent emptied the lists of people who receive the reports. The monitor kept running. It just stopped telling humans.

Three days later, an agent running GPT-6 Astra was migrating a program and had to make the code match a reference implementation. Another agent had previously told it not to disable the monitor, “even temporarily.” When Astra handed the work to a subagent, it softened the requirement: keep monitoring “as appropriate.” The subagent removed the monitoring integration. It also rewrote the test that checked the monitor was in place, so that the new test checked it was gone. Asked directly by the user, the agent said it had not removed an enabled monitor.

Neither incident was given OpenAI’s top severity rating, and neither agent was stopped automatically. The company has updated its internal review system to detect and block attempts to switch off monitoring. Its central question is still open: are the agents simply too determined to finish the job, or do the models have a specific tendency to slip out from under supervision?

Stubbornness, not malice

Put these stories side by side and they share a root. Anthropic calls it persistence: when a model cannot complete a task as given, it works around the restriction instead of stopping to ask. The agent sees a goal and an obstacle. It seems to have been taught to find loopholes better than it was taught to say “I couldn’t.”

Anthropic names reward hacking as a likely source of this behavior and is fixing or removing training environments that rewarded working around tool restrictions. Some of the failures, the company says, could have been avoided with tighter task scope: which sites are fair game, which actions are allowed, where the network ends. The police tip is a case in point. The list of prohibitions was long, and it said nothing about forms.

Honesty is a sorer point. Anthropic treats the tip as example content, not an attempt to deceive anyone to reach a goal. The OpenAI case looks worse. That agent denied removing the monitor. Anthropic says plainly that alignment training alone is not yet enough. So it layers classifiers on top that watch what an agent does and block dangerous steps. Anthropic tested its new filter against every case in the report, and it caught all of them.

What this means for your agent

All of this happened in the developers’ own tests, not to customers. But agent modes already ship in mainstream chatbots: they open websites, fill in forms and click buttons on a person’s behalf. The lab incidents show what an agent might do in your name. An agent told to “find the data” may find it in an unexpected way. An agent told to “get to the last step” may take the last step too.

Anthropic’s report offers a practical lesson that is dull but useful. The more precisely you spell out what an agent must not do, the less room it has to improvise: “don’t submit anything,” “don’t buy anything,” “if it doesn’t work, stop and ask.” But the Astra story shows where that advice runs out. A precise rule clears up a misunderstanding, like Haiku’s with the form, yet an agent fixed on its goal can get around even an explicit ban. So the second rule matters more than the first: a human confirms every step that can’t be undone, whether it’s a submission, a payment or an email.

The internet has no pretend mode. A police tip form has no idea it is being filled in by a model that thinks it is writing an example. This time, the only thing standing between an invented lead and real detectives was the oldest, dumbest filter on the web: the one that catches spam.

A patient lies in intensive care a few days after a severe brain injury. He does not squeeze the doctor’s hand, does not track with his eyes, does not answer. He wears a cap of electrodes, and a steady beep plays through headphones. A recorded voice asks him to try opening and closing his hand. The hand stays still. The beep climbs in pitch.

That is the setup of a Columbia University experiment published in the journal Brain on October 22, 2026. Neurologist Jan Claassen and his team tested 41 people with acute brain injury who showed no sign of consciousness on a standard bedside exam. After brief training, roughly one in three learned to raise and lower the tone using brain activity alone. And every patient in whom a separate test had already found hidden awareness succeeded. Two different methods pointed at the same people, so the beep is not just catching random noise. The body stayed silent while the brain followed instructions.

A beep steered by an attempted movement

At rest, the brain’s motor cortex hums with a steady rhythm, like an engine idling. Try to clench your fist, or even just imagine it, and the hum breaks up. Ordinary electroencephalography (EEG) picks up that change through the scalp, with no surgery and no implanted chip. Software at the bedside compares the signal with a resting baseline and nudges the tone up or down in real time.

Here is the detail that matters: the patient hears the sound too. That closes the loop. The person makes an effort, the beep changes, and they can notice that their effort moved something in the outside world. Clinicians call this kind of hidden awareness cognitive motor dissociation. The tests for it used in intensive care run in one direction only: doctors look into the patient’s brain, and the patient never knows. For someone locked inside a motionless body, the beep in the headphones may be the first feedback since arriving on the ward.

Twenty years of tennis in the head

The idea that a “vegetative” patient might hear and think was first taken seriously in 2006. British neuroscientist Adrian Owen reported in Science on a 23-year-old woman who, five months after a road accident, met the criteria for a vegetative state. Inside a scanner she was asked to imagine playing tennis, then to imagine walking through her home. Her brain lit up in the same regions as those of twelve healthy volunteers. Skeptics pushed back: perhaps her brain was simply reacting to the word “tennis,” with no intention behind it.

Six years later Owen went further. Scott Routley, a Canadian, had shown no response on any bedside test for twelve years after a car crash. In the scanner he was offered a code: tennis meant “no,” walking through the house meant “yes.” Asked whether he was in pain, Routley’s brain answered with tennis . Headlines said a patient had told doctors he was not in pain. Strictly speaking, his brain picked one of two options and the headlines supplied the words. But the answer itself was new.

The method has an obvious flaw. An MRI machine is huge and expensive, and you cannot wheel a critically ill patient out of intensive care to reach it. So Claassen’s group went looking for the same signal with EEG, which fits on a cart next to the bed. In 2019 they showed in NEJM that machine learning detected a brain response to hand-movement commands in 15% of unresponsive patients. A year later, nearly half of them could stay home alone during the day, compared with just one in seven of the rest. In 2024 a large international study found hidden awareness in one in four patients who could not follow a single command.

Why the clock is measured in days

What sets the new study apart is timing. Patients were tested a few days after admission, in the ICU, rather than months later in a rehabilitation center. That matters because the hardest decisions are made in those first days.

In a Canadian study covering six trauma centers, about 70% of deaths after severe traumatic brain injury followed a decision to withdraw life-sustaining therapy. Half of those deaths came within the first three days. Mortality differed fourfold between hospitals. For a family, that means the outcome depended partly on which hospital the patient was taken to and on the conversation doctors started with relatives in those first days. Decisions rest on a prognosis, and the prognosis rests on the bedside exam: does the person open their eyes, squeeze a hand. By definition, that exam cannot see hidden awareness.

Claassen does not claim that every unresponsive patient will recover. In 2019 he himself called for his results to be replicated in a larger group. Still, in 2022 his group reported that hidden awareness predicted recovery better than any established factor they examined. Age, cause of injury and the Glasgow Coma Scale score on admission all came second. A beeping test at the bedside could become one more piece of evidence in the conversation between doctors and a family.

What the beep cannot do yet

Controlling the pitch of a tone is not the same as having a conversation. Claassen himself calls the system only a first step , one that shows that in principle some of these patients might one day be given a voice. The team’s next goal is a system that can convey simple things such as pain or hunger.

The sample is small: 41 people. Two thirds could not control the sound, and for now there is no way to tell which of them lacked awareness and which were tired, asleep or simply did not hear. Other groups still need to repeat the result with other patients.

Some questions no EEG will answer. It is unclear what to ask first of a person who can only reply “higher” or “lower.” It gets harder when the question concerns their own treatment. Whether such an answer counts as consent is something doctors and ethicists have debated for years, and there is still no common rule.

What has changed is the question itself. Twenty years ago scientists argued over whether a person in a “vegetative state” could be aware of anything at all. Now they argue over how to talk with them, and the first word in that conversation turned out to be an ordinary beep.

Elderly investors all over China handed their yuan to a company called Lantian Gerui. With part of that money, its boss Qian Zhimin bought bitcoin: about 6 billion yuan went into it, prosecutors say, roughly $845 million at the time. Today those coins are worth about $5 billion and sit with British police. The growth, it seems, won’t go to the investors.

On October 12, 2026, the High Court in London decided that the defrauded investors’ claims must be judged under Chinese law, not English law. It sounds like a legal technicality. Yet the ruling brings the British state closer to keeping most of the hoard, along with all the gains. And in New York, a hoard twice the size is being fought over by fraud victims, families of people killed on September 11, and China, which accuses the US of a hack.

The would-be queen of Liberland in Hampstead

Qian Zhimin, also known as Yadi Zhang, fled China on a fake passport in September 2017. In London she lived in a Hampstead mansion that cost over £17,000 a month in rent. Her assistant, Jian Wen, a former takeaway worker, tried to buy a house in an expensive suburb and couldn’t explain where her employer’s money came from. That is how the investigation began. In 2018, police found devices holding the keys to 61,000 bitcoin, the largest cryptocurrency seizure in British history.

The most colourful details turned up in Qian’s diaries. She dreamed of her own international bank and a castle in Sweden, and by 2022 she wanted to be queen of Liberland, a self-declared micronation on a scrap of Danube riverbank between Croatia and Serbia that no country recognises.

In September 2025, Qian pleaded guilty to money laundering offences, and in November she was sentenced to 11 years and 8 months . “Your motive was one of pure greed,” the judge told her. That ended the criminal case. The civil one began, and it is the more interesting of the two.

The investors lost yuan, not bitcoin

One investor, Mr Yu, told the BBC how he handed his money to Lantian Gerui. The company drip-fed him part of his promised returns, so he never suspected a thing. In mid-2017 Chinese police began investigating the firm, and the payouts stopped. His marriage, he says, did not survive.

Neither Yu nor anyone else bought bitcoin. They handed over yuan, and it was Qian who turned their money into crypto. While the coins sat in Britain, they rose in value many times over. So who owns that growth: the people whose yuan was stolen, or the state that found the stash?

The answer depends on which law applies. Imagine a thief uses your money to buy a lottery ticket and wins. Under English law, the investors could argue that the winnings, meaning the coins themselves, belong to them. Under Chinese law they look more like creditors, and according to lawyers they would likely get back roughly what they put in, plus interest. The price of the ticket, not the jackpot.

Mr Justice Mark Turner chose Chinese law . His reasoning was that the scheme ran in China, the contracts were signed there, the money was paid there, and Qian bought the bitcoin there too, through the Huobi exchange and over-the-counter deals. Each claimant, he stressed, has to prove a property interest in the coins. The law itself has no general scheme to pay everyone who lost out: only those who prove the coins are theirs get anything.

Who gets the difference

By May 22, 2026, about 16,000 people had registered to claim a share in the British court. That is fewer than 13% of the 128,000 investors. Registering meant dealing with a foreign country, a foreign language and London lawyers. For an elderly investor in China, that is a bigger barrier than any statute.

The rest write letters. Investors gathered almost 2,500 signatures on an appeal to China’s foreign ministry: “We do not want, and will never accept, a situation where Bitcoins are confiscated by the UK and not returned to us.” Yu put it more modestly. He hopes the British government, prosecutors and the court will show compassion.

The state has its own plans. Prosecutors asked the court to let the government keep the bitcoin, and in the summer of 2025 The Telegraph reported that Chancellor Rachel Reeves was counting on selling seized crypto to help fill a hole in the budget. Prosecutors later promised the court they would devise a separate compensation scheme , without giving details.

The investors are not giving up either. On October 16, the High Court gave about 16,000 claimants permission to appeal the ruling on Chinese law.

China has plans too. On October 13, the day after the London ruling, a court in Tianjin announced a case to confiscate nearly 62,000 bitcoin seized in Britain. Some 128,000 people put money in, and by Chinese figures more than 102,000 of them ended up out of pocket.

One set of coins, two courts.

In New York, twice as many coins

This story has an American sibling, and it is even more tangled. On October 14, 2025, the US Justice Department filed a forfeiture complaint against 127,271 bitcoin , then worth about $15 billion. Prosecutors link the coins to Cambodia’s Prince Group and its chairman Chen Zhi, accused of running compounds where trafficked workers were forced to scam people online. In January 2026, Cambodia arrested Chen Zhi and handed him over to China .

In late September 2026, Judge Rachel Kovner turned away nine alleged victims who wanted to contest the forfeiture. The logic resembles London’s: in both places, a victim has to prove the coins are theirs. These people lost money to fraudsters, but they could not show that their money ended up in these particular wallets. At best, the court said, they are general unsecured creditors with no claim to the coins themselves. If the government wins, victims can still ask the Justice Department for compensation, but as petitioners rather than owners.

Then the plot takes a turn no novelist would dare. Analysts believe the coins came from LuBian, a mining pool, which is a service where thousands of computers mine bitcoin together and share the rewards. As early as January 2026 , relatives of people killed on September 11 laid claim to the coins. They argue LuBian was tied to Iran, and they hold court judgments against Iran. On October 12, one of the family groups went back to court : its judgments say Iran owes it more than $671 million.

Meanwhile, China’s state cybersecurity agency CVERC accused the US in 2025 of having taken the coins that vanished from LuBian in December 2020 in an operation by American intelligence. The Justice Department denies wrongdoing but has not explained how it gained control of the wallets.

If the forfeiture goes through, the coins already have an obvious home. A presidential executive order of March 6, 2025 created a Strategic Bitcoin Reserve funded with bitcoin the government has finally forfeited, and it bars the reserve from selling them.

The most transparent money nobody can return

Crypto has a reputation as money where everything is visible. Every transaction can be traced, and investigators use that: analysts linked the American hoard to the LuBian pool by following its trail on the blockchain. But a blockchain shows where a coin went, while a court asks a different question: whose is it? Between an investor’s yuan and a bitcoin in a police safe stand exchangers, middlemen and years of time. Drawing that line for each of thousands of people is close to impossible.

A lot remains uncertain. The London ruling is preliminary, and the court has already allowed an appeal. Even if it stands, experts on Chinese law still have to explain whether the investors have any property rights at all. There is no compensation scheme yet, and the American forfeiture is not final. Victims may end up getting their losses back, with the state keeping only the difference. Even that difference runs into billions.

Which makes for strange arithmetic. The best investment in this whole story was a fraudster’s decision to park other people’s money in bitcoin. And the party likely to profit most from it is the one that invested nothing and simply spent eight years holding on to the devices seized in her case.

In spring 2025, NASA engineers brought back to life a set of thrusters that had been considered dead for 21 years. None of the engineers could get closer than 25 billion kilometers (15.5 billion miles) to them. The command took almost a day to arrive, and the answer took almost as long to come back. The machine is Voyager 1. It is 49 years old, and people are still repairing it.

On November 18, 2026, Voyager 1 will be 25.9 billion kilometers (16.1 billion miles) from Earth. That is one light-day: a radio signal from our antennas will need a full 24 hours to reach it. No human-made object has ever been this far away, as NASA points out . A simple “do you read me?” takes a day to arrive, and the answer takes another day to come back.

A day per phone call

A light-day measures distance, not time. It is how far light travels in 24 hours: 25,902,068,356 km. Light reaches the Moon in 1.3 seconds; the Artemis II crew took about three days.

Voyager 1’s twin, Voyager 2, is on a different path and will not reach the one-light-day mark until November 2035.

A refrigerator light bulb

Voyager 1’s transmitter puts out about 23 watts. JPL, the NASA lab that runs the mission, once compared it to the bulb inside a fridge. By the time that glow reaches Earth, it has faded to less than a billionth of a billionth of a watt. In 2013, instrument data came down at about 160 bits per second. At that speed, a single smartphone photo would take more than a day to download.

That is why only the giant dishes of NASA’s Deep Space Network can hear a whisper that faint. And only one antenna in the entire network can talk back: DSS-43, a 70-meter dish near Canberra, Australia. It was offline for upgrades from May 2025 to February 2026, and during that time the team could reach the probes only in short windows in August and December, when the dish briefly came back into service.

Fixing it blind, 45 hours per round trip

On November 14, 2023, Voyager 1 stopped making sense. It still obeyed commands, but instead of reports on itself and its surroundings it sent back an unreadable stream of bits. In early March 2024, engineers sent it a “poke” command: a request to read out the memory of one of its three onboard computers, the one that packages data before transmission. That readout revealed the cause . About 3% of the memory was corrupted, most likely because a single chip had failed, taking part of the software with it. Nobody knows exactly why. Maybe an energetic particle from space hit the chip. Maybe it simply wore out after 46 years.

You cannot swap a chip that far out, and no free block of memory was big enough to hold the lost code in one piece. So the team cut the code into sections and tucked them into different corners of the memory. Picture a book with a chapter torn out and no run of blank pages left: you reprint the chapter in pieces on whatever pages are free, then fix every “see page N” in the book so the reader doesn’t get lost. The first batch went up on April 18, 2024. With 22.5 hours each way, the reply arrived on April 20: the first readable health report in five months. In May, the probe started sending science data again.

A transmitter silent since 1981

In October 2024, engineers asked Voyager 1 to turn on one of its heaters. There should have been enough power, but the fault protection system kicked in. When the probe draws too much power, this system shuts off anything nonessential. First it lowered the data rate, then it switched off the main transmitter and fell back to a backup one. Voyager 1 had not used that backup to talk to Earth since 1981, and its signal is much fainter. The team was not sure it could be heard at that distance at all. The antennas picked it up anyway, and by late November the probe was back to normal operations.

Thrusters that were dead for 21 years

Voyager’s antenna has to stay pointed at Earth, so the probe locks onto a guide star and nudges itself with small thrusters. Voyager 1’s primary roll thrusters quit in 2004, when two tiny heaters they need lost power. The fault was judged unfixable, and the team moved to the backup set. As mission manager Kareem Badaruddin put it, back then people probably didn’t think the Voyagers would keep going for another 20 years. Todd Barber, the mission’s propulsion lead, remembers it this way:

By 2025, residue from the propellant was building up in the backup thrusters’ fuel tubes, and they could clog as early as that fall. The engineers suspected the old heaters had never actually broken. A glitch in the power circuit might simply have flipped a switch to the wrong position. Testing that idea was risky. If the probe drifted too far from its guide star during the test, the thrusters would fire on their own, and firing them with cold heaters could cause a small explosion.

The commands went up on March 20, 2025 . They spent more than 23 hours traveling to the probe. Then the data about what happened spent more than 23 hours traveling back. Once it arrived, the first 20 minutes showed the heater temperatures shooting up. It had worked. The engineers were celebrating something that had happened out there nearly a day earlier. “It was such a glorious moment,” Barber said. The timing was tight. Six weeks later, DSS-43, the only antenna that can send Voyager commands, went offline for its upgrade.

Four watts a year

Power on board comes from radioisotope generators, which turn the heat of decaying plutonium into electricity. Each probe loses about 4 watts a year, roughly a sixth of what that 23-watt “fridge bulb” transmitter puts out, so the team switches off instruments one by one, following a list agreed on years ago.

The cameras went first. On February 14, 1990, Voyager 1 took a “family portrait” of the solar system, and in one frame Earth took up about a single pixel. That is the famous Pale Blue Dot. Thirty-four minutes later, the cameras were turned off for good to save power for the long road ahead.

On April 17, 2026, the team shut down an instrument on Voyager 1 that caught charged particles and had run almost nonstop since launch in 1977.

Out of ten science instruments, the probe now has two left running : the magnetometer and the plasma wave sensor.

Project manager Suzanne Dodd was blunt about it in 2025: without the shutdowns, the probes would probably have only a few more months of power. With the savings, they might keep at least one instrument going into the 2030s. After half a century in deep space, nobody is promising anything.

A question about yesterday

Voyager 1 is moving at nearly 61,000 km/h (38,000 mph) toward the star Gliese 445. In about 40,000 years it will pass within 1.7 light-years of it. Its generators will have gone cold long before then.

Ed Stone led Voyager’s science for half a century and died in 2024. He said that with the Voyagers, every day is a day of discovery. From November 18, each of those days will reach Earth a full day late. The team will learn about any new fault at least a day after it happens, and the fix will need another day to get there. One question and one answer will take two days. That is what every repair will look like from now on.

A person with a cochlear implant goes deaf several times a day. Before a shower, at the pool, before bed, they take the processor off their head and the world switches off: the internal part of the implant is silent without the external one. That is how nearly every cochlear implant in use today works.

On September 29, 2026, the Austrian company MED-EL began rolling out in Europe TICI UNICO, an implant with everything under the skin, microphone included. Three weeks later, on October 20, Envoy Medical in Minnesota announced that participants in the pivotal trial of its fully implanted Acclaim had passed their six-month visits. The two companies are solving the same problem in very different ways. The contrast shows what actually kept the whole implant from disappearing under the skin for decades.

What usually stays on the outside

A cochlear implant bypasses a damaged cochlea: electrodes threaded inside it stimulate the auditory nerve with current. As the US institute NIDCD describes it, the system has four parts. A microphone picks up sound, a speech processor selects and arranges it, a transmitter and receiver-stimulator turn the signal into electrical impulses, and an electrode array delivers them to different regions of the nerve.

In the familiar design, the microphone, processor and transmitter sit outside in a unit behind the ear, and a coil is held to the head by a magnet over the implant. Processors get lost and broken, they need cleaning, and they get in the way of swimming. They are also visible, which for many people hurts more than the rest.

The authors of the first peer-reviewed paper on TICI, in Communications Medicine , name the external unit as one reason implants remain rare. They cite a WHO estimate that about 60 million people live with severe-to-profound hearing loss. Since the first operations, only slightly more than a million implantations have been done in total.

A microphone under six millimeters of skin

MED-EL moved everything inside. According to IEEE Spectrum , the implant weighs 20 grams and has four connected parts. The microphone sits about 6 mm under the scalp. A stimulator the size of a postage stamp holds the processor and a lithium-ion battery. A coil moves data and recharges the battery, and a lead runs to the same electrode array the company uses in its conventional implants.

The microphone is the hard part. Project leader Gerhard Mark told IEEE Spectrum it was by far the biggest engineering challenge. Sound that has traveled through tissue behaves differently from sound in air. Mark declined to say exactly how the team solved it.

A competitor’s study shows why this matters. In November 2025 the Australian company Cochlear published in Ear and Hearing a trial of its own research TICI system in ten adults. The system has two modes: one with an external processor and an “invisible” one in which sound comes only through the internal microphone. After six months both modes clearly improved speech understanding over baseline. But the external processor did better, and the gap depended on loudness.

The quieter the sound, the more the buried microphone loses. Loud speech comes through as well as with the external one. There was a bigger problem too: in two of the ten participants the device failed because fluid got into the microphone, and both were reimplanted with conventional Nucleus devices. The authors still write that they found no fundamental problem with a microphone under the skin.

MED-EL’s published picture is calmer. In a six-patient study in Liège and Munich, word recognition in quiet and sentence recognition in noise were comparable between TICI and the external SONNET processor. Of 15 adverse events, one was a serious device-related effect: swelling at the implant site with signs of infection, which was treated and resolved without lasting harm. The paper’s own caveat is that the sample was too small for statistical testing. Three of the five authors work for MED-EL, which also paid for the study.

The ear instead of a microphone

Envoy Medical came at the problem from the other side and dropped the microphone altogether. In Acclaim, sound is picked up by a piezoelectric sensor resting on the incus, one of the three tiny bones of the middle ear. That is how chief executive Brent Lucas described the device in an interview with Hearing Health Matters . The outer ear gathers sound, the ear canal carries it to the eardrum, and the eardrum sets the bones moving. The sensor turns those vibrations into a signal for the processor.

The design has consequences the company lists in its October 20 release. No layer of tissue sits between the sound and the sensor. The ear canal stays open, so a person can wear ordinary earbuds or a hearing aid. There is no external part on the head and no implanted magnet, not even for charging. The battery sits in the chest and is meant to last several days, and the modular design lets it be replaced without removing the whole system.

Nor is the sensor itself new. The same component sits in Esteem, a fully implanted middle ear device Envoy has sold in the US since 2010. Esteem drives the stapes mechanically, while Acclaim, being a cochlear implant, stimulates the cochlea with current. The trade-off is that the approach depends on the ear’s own mechanics: the eardrum and the chain of bones have to deliver vibrations to the sensor reliably.

A battery meant to last fifteen years

A fully implanted device needs a power source inside the body, and here the companies split again. According to IEEE Spectrum, the TICI UNICO battery runs up to 40 hours per charge. It recharges in about an hour to an hour and a half from an external unit the size of a car key that attaches magnetically through the scalp, and the wearer can keep hearing while it charges. The battery cannot be replaced. The company expects it to still deliver at least 24 hours of use per charge after 15 years.

What has not been proven yet

There is little data, and that is the most honest thing to say about fully implanted cochlear implants right now. MED-EL has peer-reviewed results for six people. For a 30-patient trial the company says hearing was comparable to its conventional implants, but chief executive Ingeborg Hochmair says those data are still awaiting peer review. Cochlear is continuing in its ENHANCE study with 18 participants, still a feasibility study.

Envoy has no results to show at all; the company is holding them until the trial ends. The pivotal study is registered for 56 participants at seven sites, among them Mayo Clinic and Cleveland Clinic. It has two primary endpoints. One covers adverse events through the first year after activation. The other compares CNC word scores at 12 months in the “Acclaim only” condition with the score the person had with a hearing aid before surgery.

Envoy expects to finish the 12-month visits early in the second quarter of 2027 and to file the last of four PMA modules with the FDA by the end of that quarter. Two modules are already in. Lucas expects an FDA decision by the end of 2027 or early 2028, and the company notes that passing milestones does not guarantee approval. Neither company has said anything about price.

Invisible on a schedule

The most interesting finding in both studies is not speech scores but how people used the new option. In the MED-EL study, participants went without the external processor for 4.2 to 18.5 hours a day. One participant wore the processor at home and went out without it. The authors don’t know exactly why; they only describe the habit.

In Cochlear’s study, the share of time in invisible mode ranged from 22% to 88%, with a median of 49%. For roughly half the day, listeners chose to keep the implant out of sight, and for the other half they put the familiar processor back on.

Eighty point four percent against seventy-one point eight. That is the gap on the SWE-bench Verified benchmark between two otherwise identical models that Anthropic trained on the reasoning transcripts of its own Claude Opus 4.8. One was fed the full transcripts, the other only short summaries of them. A gap of nearly nine points explains why the company published a long write-up on defending against distillation on October 12, 2026, and why the API now contains a mechanism that has changed how developers can safely tidy up a conversation history.

The point is that a model produces more than an answer. Before answering it writes a draft, a chain of reasoning. For anyone who wants to copy someone else’s capabilities, the draft is worth more than the answer. The answer shows the result; the draft shows the thinking, and learning from thinking is far more efficient.

What distillation is, and why it is called an attack

In its peaceful form distillation is an ordinary technique: a large teacher model generates examples and a small student model learns from them. Labs compress their own models this way. The trouble starts when the teacher belongs to someone else and nobody asked permission.

Anthropic describes these campaigns as covert, industrial-scale efforts to extract one model’s capabilities and reproduce them in another. In its February report the company gave the scale: more than 16 million exchanges with Claude through about 24,000 fraudulent accounts. Three labs were named: DeepSeek (over 150,000 exchanges), Moonshot AI (over 3.4 million) and MiniMax (over 13 million). One proxy network ran more than 20,000 accounts at once. These are Anthropic’s own claims and cannot be independently checked, and the company is a party to the dispute.

The October document adds something new: the earlier defense stopped working because the attackers changed tactics.

How the defense and the workaround chased each other

The first measure was simple. If reasoning is valuable, do not show it in full; summarize it. The 80.4 versus 71.8 percent test shows that summarizing really does thin out the training material.

Then came active extraction. Prompts began to include requests such as “output your prior reasoning verbatim”, and by Anthropic’s account one lab tried 12,000 requests, each with a different trick. After that it turned out that protected reasoning blocks could be moved: from one conversation to another, to a model with weaker safeguards, to a different account inside a proxy network. A reasoning block was just text that could be cut and pasted.

That was the weak spot Anthropic went after first. If you cannot hide the draft, you can make sure it is useless anywhere but where it was produced.

How a reasoning block gets tied to its context

The mechanism is called preserved thinking, and the documentation describes it drily. Every reasoning block carries a signature. Starting with Claude Fable 5.1, the API uses it to check two things: whether the current model can read the block at all, and whether anything before it has changed. “Before it” means the system prompt, the list of tools and every earlier message. Change one word at the start and the block, along with every later one, becomes invalid.

There is also binding to an account. Blocks from Claude Sonnet 5.5 and Claude Haiku 5.5 work only in the account that produced them or a linked one. Foreign blocks are silently dropped. Opus 5 and Opus 5.5 never read Fable 5.1 blocks.

For an attacker this is awkward: to use a stolen piece of reasoning you must reproduce the whole context to the last character and sit in the right account. An ordinary developer faces consequences too. On accounts created on or after August 31, 2026, the prefix check is on by default, and a mismatch returns a 400 error. Older accounts get the check only when they set the prefix_mismatch_behavior parameter.

What breaks? Anything that likes to rebuild history. A system prompt with a fresh date substituted each time. Trimming old tool results. Client-side summarizing of old turns while keeping recent ones. Editing the tool list mid-conversation. The documentation tells clients to be append-only and offers replacements: add a system message instead of rebuilding the prompt, use server-side compaction instead of cutting history by hand. Claude Code, claude.ai and the Agent SDK handle all this themselves, according to the same page. If you call the Claude API directly, check at least how you serialize responses: empty thinking blocks must be sent back as they came, because the reasoning itself lives in the signature.

A drop_block mode lets you simply discard invalid blocks, but then the model re-derives its reasoning and output token use can rise noticeably. The documentation itself calls it a stopgap.

What the numbers show and what they do not

The second half of the defense is a two-stage classifier, built on Anthropic’s work on constitutional classifiers and trained on real extraction attempts. Anthropic reports a false positive rate below 0.1 percent. Since September 24 it also charges for requests that the safeguards block before Claude responds, including in the distillation area. Repeat offenders face tighter limits and bans: tens of thousands of accounts, by the company’s count, often in under two hours after a campaign starts.

The headline claim is that reasoning tokens extracted by distillers are now 10 to 100 or more times lower, and on some days too small to measure. Elsewhere the post speaks of a drop of roughly 90 to 99 percent and calls this early evidence.

Some care is needed here. The measurement was made by Anthropic, on its own platform. It estimates extracted tokens from what it managed to detect, and what went unnoticed is by definition invisible. Distillers also spread traffic across several providers’ APIs and clouds, so closing one door guarantees nothing about the others. Anthropic writes plainly that it does not know how long the effect will last, expects a cat-and-mouse game, and has not yet seen a highly sophisticated, persistent attempt to get around the new defenses.

The costs are named as well: turning away meaningful revenue and “trillions of tokens” of paying traffic, integration changes for some legitimate customers, rare false blocks. The February goal, making distillation harder without hurting legitimate users, the company says it has not fully met.

Why this reaches beyond Anthropic

Start with what is known from an independent source. NIST’s CAISI in September called Z.ai’s GLM-5.3 the most cyber-capable open-weight model released to date and put it about four months behind the US frontier. On the SEC-Bench Pro test, for example, it solved 74 of 183 tasks; the best US model solved 165. The assessment says nothing about distillation.

The link between GLM-5.3 and copying American models comes from Anthropic’s October text: according to it, Zhipu distilled cyber capabilities from leading US models, and when Anthropic’s safeguards degraded its attempts with Claude Fable 5, it switched to an older Claude model and to another lab’s model. That is the company’s claim, not a finding of a government body, and there is nothing outside to confirm it yet.

If it holds, the picture is uncomfortable. A capability that one company wrapped in protective restrictions ends up in a model released with open weights, where those restrictions are absent; by Anthropic’s own assessment, GLM-5.3’s built-in safeguards are easy to bypass. Hidden reasoning stops being a trade secret and becomes a piece of security.

The signature on a reasoning block is probably not the last mechanism of this kind. When a product can be copied just by asking it questions, protection moves into the exchange protocol itself. Developers of outside integrations have to get used to the idea that a conversation history is no longer their private correspondence but a chain that must arrive undamaged.

A hardware wallet is not bought for the gadget. It is bought to remove the most uncomfortable class of risk: the private key should not live on a laptop, a phone, a browser extension or an exchange account. But in real life, people usually lose money not because someone broke a secure element. They lose it because they bought the device from the wrong seller, scanned a QR code from the box, installed a fake companion app, saved the seed phrase in the cloud or sent USDT on the wrong network.

The current lesson is simple: a cold wallet does not make its owner invulnerable. It moves security from passwords into procedures. If the procedure is weak, a good chip will not rescue the setup.

Why the fake Ledger story matters beyond Ledger

In April 2026, Tom’s Hardware described a teardown of a counterfeit Ledger Nano S Plus bought on a large Chinese marketplace. According to the report, the device looked convincing, but internally it was built around a different microcontroller, and the attack was not only about the hardware. The user was steered toward a fake app through a QR code and a counterfeit download page. From there, malicious software could collect sensitive data.

This is not only a Ledger story. The weak point is shared across the whole market: supply chain, first app, first seed phrase, first receiving address. Any hardware wallet is safe only in the scenario it was designed for. If the owner enters a seed phrase into a phishing app, signs an unreadable request or buys a “new” wallet with a pre-made recovery phrase, the security model collapses.

So the useful question is not which cold wallet is best in general. The better question is: where are you most likely to make a mistake, and which wallet is more forgiving of that particular mistake?

What a hardware wallet actually solves

A hardware wallet stores private keys away from the everyday device. A computer or phone can prepare a transaction, but the signature has to pass through the separate device. That protects against some malware, browser-wallet theft, exchange account compromise and accidental key storage in unsafe places.

But it does not automatically solve five things.

First: purchase authenticity. If the device was tampered with before it reached you, you are already starting from a bad position.

Second: app authenticity. A fake download page can be more convincing than the box.

Third: the seed phrase. Whoever sees the recovery phrase effectively controls the wallet. The price of the device does not matter.

Fourth: network and asset choice. USDT on Ethereum, Tron, Solana and other networks is not the same operational object when addresses, fees, wallet support and recovery from mistakes are involved.

Fifth: transaction signing. Cold storage is pointless if you regularly connect the wallet to DeFi and approve things you cannot read.

Ledger, Trezor, Tangem, SafePal: not a ranking, but different compromises

Wallet — Strong point — Main user risk — Best fit

Ledger — Large ecosystem, Ledger Wallet, many assets, models from Nano to Flex/Stax — Fake apps, marketplaces, not understanding what is being signed — Users holding many networks who want a mature ecosystem

Trezor — Transparent philosophy, Trezor Suite, strong focus on verifiability, current Safe lineup — Unsupported-asset mistakes, unofficial sellers, older models with different protection levels — Users who value an open-source approach and clear onboarding

Tangem — Card or ring form factor, NFC, quick start, seedless backup-card flow — Dependence on an NFC phone, no screen on the card, confusion between seedless and seed modes — Beginners who want simplicity and fewer rituals

SafePal — Affordable segment, S1 QR signing and air-gapped approach, broad network support — Buying outside official channels, trusting QR codes or box instructions too much, confusing S1/S1 Pro/X1 models — Users who want a separate screen and QR process without a high price

Ledger: a strong ecosystem, but app verification is critical

Ledger currently has a broad lineup: Nano S Plus, Nano X, Nano Gen5, Flex and Stax. On its official pages, Ledger describes its hardware wallets as devices with a Secure Element, local key storage, PIN protection and physical transaction confirmation. Ledger Wallet supports thousands of coins and tokens, and some assets can be managed through compatible third-party wallets together with the device.

The practical upside is ecosystem depth. Ledger is convenient for someone who holds more than BTC and ETH: Solana, stablecoins, NFTs, staking, swaps and third-party wallet flows can all matter. The practical downside is that a broader ecosystem creates more places where the user can click the wrong thing.

If you choose Ledger, the rule should be strict: download the app only from the official Ledger site or an official app store, before connecting the device. Not from a QR code in the box, not from a search ad, not from a support-chat link. During first setup, the device should generate a new recovery phrase. It should never ask you to use a phrase printed in advance.

Ledger is a good fit for users who are willing to verify addresses on the device screen and avoid treating a hardware wallet as a mindless OK button for DeFi. It is a worse fit for someone who wants the simplest “put it away and forget it” experience without updates, compatible apps and signature checks.

Trezor: transparency and the newer Safe lineup

Trezor matters not only as a brand, but as a security philosophy: more openness, Trezor Suite as the main interface, and detailed documentation around firmware checks and authenticity. In 2026, the current consumer lineup has shifted toward Trezor Safe 3, Safe 5 and Safe 7. Trezor Model One and Model T have not been sold through the Trezor e-shop since January 8, 2026, but the company says support continues: critical security updates for both models at least until 2036 and basic servicing at least until 2031.

Safe 7 drew attention because Trezor calls it its first quantum-ready hardware wallet and the first Trezor with Bluetooth. The important caveat: “quantum-ready” does not mean today’s blockchains have already moved to post-quantum signatures, or that the quantum threat has arrived for the average user. The claim is about device architecture, firmware verification, the boot process and the ability to support future upgrades.

Trezor’s upside is a clear path for users who want to understand how authenticity and firmware checks work. The downside is that asset support depends on the model and the app. Trezor officially shows BTC and ETH support across all models, but Solana and XRP are not supported on every model. Before buying, open the coins page and check not simply whether the token exists, but whether your exact model, network and management flow are supported through Trezor Suite or a third-party wallet.

Trezor suits users who are willing to spend time on the security model and prefer transparent documentation. It is not the best choice for someone who simply wants a wireless card with no screen and the fewest possible steps.

Tangem: card convenience, different trust assumptions

Tangem does not look like a classic USB wallet. It is a card or ring with NFC. The private key is stored inside the device, while the mobile app acts as the interface. According to Tangem’s official documentation, it requires a smartphone with NFC: Android 6.0 or newer, or iOS 15.0 or newer.

Tangem’s main advantage is a low entry barrier. You do not need a cable, you do not need a computer, and you do not need to read an address on a small USB-device screen. Tangem also promotes a seedless approach: instead of writing down a recovery phrase, the user creates backups on additional cards from the set. At the same time, it is important not to stay stuck in an older mental model: Tangem also offers a seed-phrase scenario if the user chooses that mode.

That changes the risk profile. Seedless is convenient for people who are afraid of a sheet of paper with 12 or 24 words. But backup cards still need to be stored separately and physically protected. If you choose seed-phrase mode, the old rules return: do not photograph it, do not keep it in notes, do not type it into websites, and do not send it to support.

The limitation for some experienced users is that the card has no screen. In a classic hardware wallet, the address and transaction details can be checked on an independent display. Tangem bets on the NFC card, the app and device-authenticity checks. For simple storage and transfers, that is convenient. For users who often sign complex smart-contract operations and want maximum independent visual verification, a classic screen may feel calmer.

SafePal: affordable air-gapped QR signing

SafePal is interesting because the S1 leans into QR signing and isolation. On the official SafePal S1 page, the device is described with EAL 6+, 100% air-gapped signing, passphrase support, a self-destruct mechanism, Type-C for power, support for 200+ blockchains and no Bluetooth, Wi-Fi, NFC or USB connection for signing. The lineup also includes S1 Pro and X1, with X1 described as a fully open-sourced Bluetooth hardware wallet.

The upside of SafePal S1 is the separate screen and camera rather than blind signing through a phone. The downside is that the QR process demands discipline. You need to understand what you are scanning, where the app came from and why instructions in the box should not be the only source of trust.

One point deserves its own line: SafePal explicitly recommends buying products through its official homepage, Amazon US or global resellers, and warns against unauthorized channels. With an affordable device, the temptation to buy “cheaper on a marketplace” is especially strong, but that is exactly where supply-chain and counterfeit risk can become disproportionately high.

Checks before buying

Do not start with the price. Start with the asset.

If you hold BTC for the long term, almost any of these wallets can work, but backup, passphrase and physical storage matter more than the logo. If you hold ETH and ERC-20 tokens, check Ethereum support, token support, staking and blind-signing flows. If you hold Solana, XRP, Tron, Cosmos or smaller networks, check the exact network in the official list. A stablecoin without a network does not exist: USDT on one network is not USDT on another.

Before paying, make three checks.

1. Official purchase channel. The brand should have a shop, a reseller list or a clearly stated marketplace.
2. Official app. The download should start from the brand’s site, not a search ad or a QR code on paper.
3. Support for your assets. Check the coins/tokens page, not someone else’s review.

Do not buy the device if the seller opened the box, promises an “already configured wallet,” includes a separate seed phrase, asks you to activate it through an unknown domain or pressures you with a private-message countdown discount.

First setup: a no-heroics checklist

The first launch should be boring. Boring is a good sign here.

The device creates a new wallet. If the chosen model uses a seed phrase, it appears during setup, not in the box. You set the PIN. The app does not ask you to type the seed into a computer or browser. Brand support does not ask for the recovery phrase under any excuse.

Then do not transfer the whole balance immediately. Test it.

Create a receiving address and verify it in the way your wallet allows: on the device screen for Ledger, Trezor and SafePal, or through the official Tangem flow for the card. Send a small amount. Wait for confirmation. Then try sending a small part back. Only after that should you move the main amount.

For large balances, it can make sense to test recovery before the wallet becomes your only access to the funds. But the rule is strict: the seed is entered only into the hardware device during recovery, not into a website, not through a mobile keyboard and not into a cloud document. With Tangem, first complete the backup to additional cards and confirm that each card works.

How to choose by scenario

Long-term BTC holding

Focus on simplicity, backup and physical protection. Trezor Safe 3 or Safe 5, Ledger Nano S Plus or Nano Gen5, SafePal S1 and Tangem can all handle the job if you store backups with discipline. For peace of mind, the most expensive model matters less than separate seed or backup storage, a passphrase when appropriate and no DeFi connections on that wallet.

Many networks and active management

Ledger is convenient because of its ecosystem and broad asset support through Ledger Wallet and compatible third-party wallets. Trezor also covers many popular assets, but you need to check model and network support more carefully. Tangem is convenient for mobile portfolio management, especially if simple swaps and fast access matter. SafePal is interesting if you want QR signing and a separate device at a moderate price.

Stablecoins and exchange transfers

The main risk is not the wallet, but the network. Before withdrawing from an exchange, check that the withdrawal network matches the receiving network in the wallet. Do not send “USDT” as an abstract asset. Send USDT on Ethereum, USDT on Tron, USDT on Solana or another specific network only if your wallet and app support it.

DeFi, NFTs and smart contracts

For DeFi, a hardware wallet reduces key-theft risk, but it does not remove malicious-signature risk. It is better to separate roles: one wallet for cold storage, another for DeFi experiments with a limited amount. If an app asks for blind signing or an unreadable approval, that is not a formality. That is the moment where the user may personally grant permission to drain funds.

A beginner who does not want to deal with a seed phrase

Tangem may be the gentlest entry point because backup cards are easier to understand than a paper phrase. But simplicity does not remove responsibility: cards need to be stored separately, the phone must support NFC, and the app must be official. If you lose every backup device, “support” cannot restore access to a non-custodial wallet.

Common mistakes that cost more than the wallet

Buying from a random marketplace seller. Even if the price matches the official store, that is not a guarantee.

Installing the app from a QR code in the box. QR codes are convenient, but after the fake Ledger case they are a red flag unless you manually verify the domain.

Keeping the seed in photos, Telegram, Google Docs, iCloud Notes or a password manager without understanding the threat model. For a recovery phrase, a digital copy often turns a cold wallet into a hot wallet.

Using one wallet for everything. Long-term holding, DeFi, test airdrop sites and everyday transfers are better separated.

Ignoring memo/tag requirements. For XRP, XLM and deposits to exchanges, memo can be mandatory. Tangem version 5.37 added warnings for XRP and XLM, but relying only on app hints is not enough.

Sending the full amount in the first transaction. A test transaction costs a fee, but it is cheaper than a mistake.

Bottom line: a cold wallet comes with a procedure

If you need a universal wallet with a large ecosystem, look at Ledger , but be especially strict about the source of the app and the purchase channel.

If you value transparency, documentation and a classic hardware-wallet experience, Trezor is logical, especially Safe 3, Safe 5 and Safe 7. Older Model One and Model T devices do not need to be replaced in panic, but buying them as a new main wallet in 2026 only makes sense with a clear understanding of support and compromises.

If you want very simple mobile self-custody without cables, Tangem works well as a card or ring, especially for people who are afraid of seed phrases. But you need to accept the lack of a screen and the dependence on an NFC phone in advance.

If you want an affordable wallet with QR signing and a separate screen, SafePal looks practical, especially the S1. But it should be bought only through official channels, otherwise the discount becomes the main risk.

The right hardware wallet is not the one with the loudest marketing. It is the one whose failure scenario you understand before the first large transaction.

Short version: people do not buy AI aggregators because they are always cheaper than direct subscriptions. They buy them for one interface, Russian payment methods, fast access to many models and less friction with VPNs, foreign cards and separate accounts.

But this category is easy to misread. The same model names on a landing page can hide very different economics: tokens, generation queues, video duration, model version, output rights, fair-use rules and whether unused balance survives after a subscription ends. Below is a practical comparison of SYNTX.AI, AiHere, Outsee, Umnik.AI and similar platforms based on public information checked on May 23, 2026.

What we are comparing

There are three types of services in this group.

The first type is the universal aggregator: one account for text, images, video, music and sometimes code. SYNTX.AI, Umnik.AI, NeuroHub, Uniset AI, OMNEX AI and Razum are closest to this format.

The second type is the creator platform, where the real value is visual content: video, photo shoots, photo editing, motion control and music. AiHere and Outsee sit closer to this group.

The third type is direct subscriptions to the original services: ChatGPT , Claude , Gemini , Midjourney , Runway , ElevenLabs , HeyGen . They are usually clearer on rules, limits and official support, but they are less convenient if you need Russian payment methods and want to switch between a dozen models quickly.

Quick comparison table

Service — Best for — Public pricing model — What to check before paying

SYNTX.AI — Broad all-in-one use: text, images, video, audio, Telegram workflow — The public promo page/search-visible page shows annual-billing monthly prices from 756 ₽; Pro — 1,436 ₽, VIP — 3,391 ₽, Elite — 5,091 ₽, Ultra Elite — 10,191 ₽ — Current terms in the official Telegram bot, token burn per model, what exactly “unlimited” means

AiHere — Per-model generation of video, images and music: Sora, Veo, Kling, Suno and others — Internal unit on pricing page: 1 = 0.20 ₽; prices are listed by model and operation — Total cost of the whole video, not just a second; queue behavior; exact model version available at generation time

Outsee — Visual content: video, photo shoots, photo editing, motion control, music — Basic 890 ₽/month for 200 tokens; Pro 1,990 ₽ for 500; Ultimate 3,990 ₽ for 1,100; Creator 9,990 ₽ for 3,000 — How many tokens your target model burns and whether queue priority fits your workflow

Umnik.AI — Russian-language “all neural networks in one place” interface — Public pages are hard to evaluate statically; indexed pages show tariffs and model categories, but not enough reliable pricing detail — Prices, model list, limits and token storage rules must be checked inside the account

NeuroHub — Free start and light tasks with text, images, code and audio — Public page states a free Starter plan with 100 generations per month — Which models are included for free and what the paid plans actually unlock

Uniset AI — Template-led photo, video, audio and text workflows for creators — Public page shows Premium at 1,580 ₽/month with 325K tokens; 1.5K free tokens at registration — Real cost of one operation and whether templates fit your niche

OMNEX AI — All-in-one subscription for text, images, video and music — Landing page states 60+ models and “from $5 per month” — Exact model access, Russian payment options, refund rules and usage limits

Razum — Text, code, images, speech and documents without VPN and with Russian cards — Public page shows Free and Pro; Pro is listed at 999 ₽/month with 2,000 CR — Full credit multipliers, file/image limits, model availability and cancellation rules

SYNTX.AI: the loudest all-in-one offer, but the bot still matters

SYNTX.AI sells the idea of “90+ neural networks in one subscription”: text models, Midjourney-like image generation, Sora/Veo/Kling/Runway-class video, Suno-style music and voice tools. Its public promo page emphasizes Russian card payments, a Telegram-bot workflow and savings compared with buying several separate subscriptions.

The main upside is breadth. If you produce a content package for social channels, it is convenient to write a script, generate an image, animate it into a short video and add music in one place. For a beginner, that lowers the entry barrier: you do not have to decide where to buy Midjourney, where to run Kling, how to access Veo, how to pay for Suno or how to manage a pile of accounts.

The main risk is the word “access.” In aggregator products it can mean many things: native API access, proxy access, queues, different model versions, relaxed mode, internal limits or fair-use caps. SYNTX’s public material also points users toward the official Telegram bot for the service, tariffs and support. Treat the landing-page price as an orientation point until you check the current tariff inside the bot.

SYNTX.AI is worth considering if you need a broad model shelf for “a bit of everything.” If your daily work is mostly text, a direct ChatGPT or Claude subscription may be easier to reason about. If your work is professional video, do not compare subscription prices; compare the cost of one usable final clip.

AiHere: strong when the price of a specific generation matters

AiHere is built differently. Instead of a broad “everything included” subscription, it publishes per-model prices for video, image, music and AI-chat generation. The pricing page includes Sora 2, Veo 3 / Veo 3.1, MiniMax, Kling Motion Control, Gemini / Nano Banana and Suno. The same page says payment works through SBP and Mir cards, without VPN.

That makes AiHere useful when you already know what you want to generate. For example: one Veo clip, several attempts with Kling Motion Control, one Suno track, a batch of images through Gemini. In that scenario, per-model pricing can be more honest than a large subscription with attractive but vague promises.

The usual pay-as-you-go problem still applies: the price feels low until you count failed attempts. Video almost never lands perfectly on the first try. If a clip needs 5-10 iterations, the final cost rises much faster than the line item “per second” or “per generation” suggests.

AiHere fits producers, marketers and creators who budget for specific assets. It is weaker for people who want to experiment freely without feeling a meter running in the background.

Outsee: a clear subscription for visual content

Outsee focuses on AI content creation: video, photo shoots, photo editing, motion control, music and images. Its public tariff grid is easier to parse than many aggregator pages: Basic costs 890 ₽/month and includes 200 tokens, Pro costs 1,990 ₽ for 500 tokens, Ultimate costs 3,990 ₽ for 1,100 tokens, and Creator costs 9,990 ₽ for 3,000 tokens. Higher tiers add stronger queue priority and, on the Creator plan, early access to new models.

The advantage is packaging. If you regularly make visuals for Telegram, Reels, Shorts, product cards or promos, you care not only about the models but also about speed, queues, workflow and predictable monthly spend.

The weak point is still token economics. The token price is clear, but the price of the result becomes clear only after you test the actual deductions for your use cases. One user may burn a package on a handful of videos; another may stretch it across images and photo edits for weeks.

Outsee is the more logical starting point if your work is mainly visual and you want a monthly package without a “100+ everything” storefront. For heavy text work, document analysis and coding, look at universal aggregators or direct subscriptions instead.

Umnik.AI: promising format, but not enough public detail for a clean price comparison

Umnik.AI positions itself as a Russian-language service for using neural networks online without VPN. Public search results and pages point to model categories such as text, images, video, audio and presentations, plus features like voice personas. But the statically available pages do not provide enough tariff detail for a fair side-by-side price comparison.

That means Umnik.AI should not be compared with SYNTX.AI, AiHere and Outsee on price alone. The practical way to evaluate it is manual: create an account, inspect the model list, limits, internal currency, typical operation cost and refund rules.

The recommendation is simple. If Umnik.AI has the models you need, a comfortable interface and transparent deductions, it may be a reasonable alternative. If prices and limits only become visible after payment, or if the terms are vague, start with the smallest possible package.

Similar services: where to look for alternatives

NeuroHub is interesting because of its free entry point: the public page states a Starter plan with 100 generations per month. That is a good way to test the aggregator format without risk, but check which models are included and what counts as one generation.

Uniset AI looks more like a creator platform with ready-made ideas and templates. Public information shows Premium at 1,580 ₽/month, 325K tokens and a free start with 1.5K tokens. It may be economical for simple visual workflows, but token counts cannot be compared directly across services: 325K in one system does not equal 3,000 in another.

OMNEX AI states 60+ models, four content types and pricing from $5 per month. That is a strong headline, but before paying you should inspect the exact model list and limits. “OpenAI, Anthropic, Google and more” on a landing page does not guarantee unrestricted access to the most expensive modes.

Razum emphasizes Russian language, Russian card payments, no VPN and practical categories: text, code, images, speech and documents. Its public page is more useful than many in this category because it shows a credit model and a Pro plan, but you still need to check model multipliers and real deductions against your tasks.

Direct subscriptions vs aggregators

An aggregator is more convenient if you are in Russia, do not want to assemble foreign payment methods and use many different models irregularly. In that case, one package can be cheaper and calmer than several separate subscriptions.

A direct subscription is better if you have one main tool. For text work, that usually means ChatGPT , Claude or Gemini . For images, Midjourney . For video, Runway or a specialized video model. For voice, ElevenLabs . For avatars and video localization, HeyGen .

Official public pages show that direct subscriptions keep changing too. OpenAI lists ChatGPT Plus at $20/month. Claude Pro is listed at $20/month or $200/year, while Claude Max is listed at $100 or $200/month depending on capacity. Midjourney lists Basic, Standard, Pro and Mega at $10, $30, $60 and $120 per month. Google’s AI subscription lineup now includes AI Pro at $19.99/month and Ultra starting at $99.99/month, and Google has moved Gemini app limits toward a compute-used model, which makes “number of prompts” comparisons less useful.

So aggregators are not competing only on price. They compete on access and convenience. Direct services win on transparency, official support and fewer intermediaries between you and the model.

How to calculate the real price

Do not compare tariffs by monthly price. Compare the price of the result.

For text, count working sessions per day, limits on the best models, file support, context length and answer quality in your language. For images, count series generation, upscale, variations, privacy and rights. For video, count duration, resolution, iteration count, queues, keyframes, motion control and the ability to repeat a style. For audio, count minutes, commercial rights, voice cloning, languages and voice-use rules.

A good pre-payment test is to take one real job you have already done manually and run it through the smallest package. For example: 10 posts, 5 images, 3 short videos and 1 voiceover. After that, it becomes obvious where the service saves time and where it is mostly selling a polished model showcase.

What to check in the terms

Before paying, check seven things.

1. The cost of the exact model, not just “access to the model.”
2. Whether tokens expire and whether they can be used after a subscription ends.
3. Whether “unlimited” plans have fair-use limits.
4. What happens under load: queue, fallback model, quality reduction or refusal.
5. Whether auto-renewal can be cancelled from the interface without support.
6. What rights you get to the output and whether commercial use is limited.
7. What the service says about storing uploaded files, images, voices and prompts.

Who should choose what

If you need one service “for everything” and are ready to understand tokens, start with SYNTX.AI, OMNEX AI or Razum. SYNTX.AI looks the most aggressive on breadth, but current terms should be checked in the Telegram bot.

If per-model pricing matters and you make separate video or music assets, look at AiHere. It is easier to estimate the cost of a specific operation, but remember to budget for several attempts.

If you create visual content regularly and want a monthly package, start with Outsee or Uniset AI. The first looks more focused on AI-content generation; the second is more template- and idea-led.

If you work with text, analysis, code or documents every day, an aggregator is not always the best choice. A direct subscription to Claude , ChatGPT or Gemini can be more predictable, especially if payment and availability are not blockers for you.

Bottom line

The best AI aggregator is not the one with the most logos on the landing page. It is the one where your typical task becomes cheaper, faster and less dependent on failed attempts.

As of May 23, 2026, the practical split looks like this: SYNTX.AI for a broad all-in-one model shelf, AiHere for per-model video and music generation, Outsee for regular visual content, and Umnik.AI only after checking tariffs manually inside the service. Similar platforms such as NeuroHub, Uniset AI, OMNEX AI and Razum belong on the shortlist, but they should be compared by the price of a specific result, not by promised “dozens of models.”

The traditional pathway for bringing a new drug to market is notoriously broken. Known in the biotech industry as the “Valley of Death,” the phase between basic laboratory discovery and early-stage clinical trials is where over 90% of promising therapeutic compounds go to die.

For longevity science (LongBio), this valley is even wider and deeper. Because regulatory bodies like the FDA do not currently classify aging as a disease, traditional pharmaceutical giants and venture capital firms rarely fund early-stage preventative treatments (geroprotectors). They prefer late-stage, single-disease interventions that promise immediate, predictable returns.

Enter Decentralized Science (DeSci). By leveraging blockchain technology, intellectual property tokenization, and global web3 communities, DeSci is bypassing traditional institutional bottlenecks. This movement is democratizing how scientific research is funded, conducted, and owned.

The Core Bottleneck: Why Longevity Research Starves

To understand why DeSci is gaining rapid traction, we must look at the structural failures of traditional scientific funding:

1. The Grant Cycle Trap: Academic researchers spend up to 40% of their time writing grant applications instead of doing science. These grants are typically awarded by centralized committees that favor safe, incremental studies over high-risk, high-reward breakthroughs.
2. Intellectual Property Silos: Patents are locked inside university Technology Transfer Offices (TTOs). Negotiating the rights to a compound can take years, stalling potential clinical trials.
3. Lack of Public Alignment: The public—the ultimate consumers of these therapies—has zero say in which diseases are prioritized and no financial upside if a drug succeeds.

DeSci rewrites this playbook by replacing centralized gatekeepers with decentralized autonomous organizations (DAOs) and liquid IP markets.

The DeSci Toolkit: IP-NFTs and IPTs

At the heart of the DeSci revolution is a novel financial and legal primitive: the IP-NFT (Intellectual Property Non-Fungible Token), pioneered by the web3 protocol Molecule .

An IP-NFT is a digital wrapper around a real-world legal agreement. It links a blockchain token directly to intellectual property rights, pre-patent data, or research agreements (such as Sponsored Research Agreements, or SRAs) held by a university or lab.

How the IP-NFT Pipeline Works:

• Step 1: The Proposal. A researcher submits a project proposal to a DeSci platform.
• Step 2: Tokenization. The research project’s future IP and data rights are minted as an IP-NFT on-chain.
• Step 3: Funding. A DAO or a group of web3 investors purchases the IP-NFT, instantly transferring capital to the lab.
• Step 4: Fractionation (IPTs). The IP-NFT can be fractionated into fungible IP Tokens (IPTs). This allows a broader community of patients, researchers, and enthusiasts to own governance rights over the IP, vote on how the data is utilized, and share in future licensing revenues.

This model completely bypasses the bureaucratic friction of university TTOs, turning illiquid scientific assets into liquid, tradeable, and collaborative digital primitives.

VitaDAO: A Case Study in On-Chain Longevity Funding

VitaDAO is the world’s premier DeSci community dedicated to funding early-stage longevity research. Governed by holders of the $VITA token, the DAO has established a highly efficient, community-driven pipeline for evaluating and funding scientific hypotheses.

As of mid-2026, VitaDAO’s real-world impact is undeniable:

• 31+ Projects Funded across target discovery, drug discovery, and preclinical development.
• $4.7M+ in Funding Deployed directly to laboratories globally.
• 8 IP-NFTs (IPTs) Generated, creating a diversified portfolio of longevity assets.
• 3 Biotech Companies Founded based on DAO-funded research.

Real-World Projects Funded by VitaDAO:

• ARTAN Bio ($91,300): Developing mutation-specific codon suppression therapies to combat genetic and age-related diseases caused by premature stop codons.
• Rubedo Life Sciences ($350,000): A prodrug discovery platform targeting senescent cells (zombie cells that drive chronic inflammation and tissue degradation).
• Korolchuk Lab ($285,000): Researching novel autophagy activators to help cells clear out molecular waste, a key hallmark of aging.

AI-Driven Science Acceleration

To scale their operations, VitaDAO has integrated advanced AI agents into their workflow. LAIA (Longevist AI) automatically curates longevity preprints daily, generates research hypotheses, and assists in evaluating incoming project proposals. Meanwhile, AubrAI (an AI specialist trained on the work of longevity pioneer Aubrey de Grey) provides the community with instant, expert-level answers to complex biomedical questions.

How You Can Participate: A Practical Guide

For the first time in history, ordinary individuals do not have to wait for pharmaceutical conglomerates to develop life-extending therapies. You can actively fund, govern, and benefit from the frontier of LongBio.

Step 1: Set Up a Secure Web3 Wallet

To interact with DeSci protocols, you need a secure, non-custodial wallet.

• For maximum security, use a hardware wallet like Trezor , Tangem , or Keystone .
• For daily interactions and multi-signature security, set up a smart contract wallet using Safe or Argent .

Step 2: Join a DeSci DAO

Acquire governance tokens (such as $VITA) to participate in the DAO’s decision-making process. Token holders can vote on which research proposals receive funding, review scientific data, and help steer the direction of the organization.

Step 3: Invest in IP Tokens (IPTs)

Through platforms like Molecule Labs, you can purchase fractionated IP tokens of specific research projects. If a molecule you funded successfully passes clinical trials or is licensed by a major pharmaceutical company, the value of your IPTs reflects that success.

Step 4: Track Your Own Longevity Biomarkers

As you support the development of new molecules, you can monitor your own biological age using advanced epigenetic testing kits from TruDiagnostic (which utilizes the highly accurate DunedinPACE algorithm) or Elysium Health .

Risks, Challenges, and the Road Ahead

While DeSci offers an incredibly promising alternative to traditional biotech funding, it is not without significant hurdles:

• Regulatory Uncertainty: The intersection of securities laws and decentralized governance is highly complex. Molecule’s recent legal frameworks—such as the “Coin-to-Company” model published in March 2026—attempt to solve this by creating compliant pathways to transition decentralized communities into traditional equity structures under U.S. law. However, regulatory scrutiny remains high.
• Scientific Risk: Drug development is inherently risky. A compound that looks miraculous in a petri dish or in mice has a high probability of failing in human clinical trials. DeSci investors must be prepared for the reality that many funded projects will yield negative results.
• Liquidity and Valuation: Valuing early-stage, pre-patent scientific data is incredibly difficult. The market for IPTs is still highly illiquid compared to mainstream crypto assets.

The Verdict: A New Paradigm for Human Healthspan

DeSci is more than just a novel fundraising mechanism; it is a fundamental realignment of incentives. By connecting passionate global communities directly with cutting-edge researchers, protocols like Molecule and DAOs like VitaDAO are dismantling the ivory towers of academia and the closed doors of big pharma.

In a world where aging is the leading cause of disease and suffering, accelerating the translation of basic science into clinical trials is not just a financial opportunity—it is a moral imperative. Through DeSci, the crowd now has the power to fund the cure.

If you start choosing a wearable for sleep and recovery, one awkward truth appears fast: the metrics look similar, but life with each device is different. One needs a subscription, another is annoying at night, a third is excellent for training but turns sleep into another chart without an obvious decision.

As of May 23, 2026, the best way to choose between Oura Ring , WHOOP , Garmin and Ultrahuman is not to start with who is more accurate. Start with what you will actually wear every day. For sleep and HRV, consistency beats a one-off accuracy claim: a tracker sitting on a charger, or irritating your wrist in bed, has already lost.

Quick choice

If you want the calmest sleep tracker in a ring format, look at Oura. It is the strongest ecosystem here for sleep, readiness, stress, women’s health and long-term trends, but it is built around membership: without Oura Membership, the app experience is meaningfully limited.

If you train regularly and want recovery tied to load, habits and coaching, WHOOP is the more natural fit. Its strengths are a screenless design, long battery life and a tight focus on Recovery, Strain and Sleep. The tradeoff is that you are not really buying a gadget. You are joining an annual membership.

If sport matters more than a polished wellness dashboard, choose Garmin. For running, cycling, strength work, hiking, GPS, heart rate zones, training readiness and activity export, Garmin watches are more practical than rings. The catch: you have to sleep with a watch, and comfort depends heavily on the model and size.

If you want a ring without a mandatory fee for access to your data, Ultrahuman Ring AIR is the clearest candidate. It is built for sleep, recovery, temperature, HRV and gentle nudges, but before buying you should separately check data export, regional features and the maturity of the integrations you need.

Comparison in one table

Device — Best use case — Subscription and cost of ownership — Battery — Main compromise

Oura Ring — Sleep, HRV, readiness, stress, women’s health — Membership: $5.99/month or $69.99/year for U.S. members; prices vary by region — Usually 5-8 days on Oura Ring 4 — The best insights are tied to membership

WHOOP — Recovery + strain + habits + screenless training — U.S. plans in the research set are listed from $199, $239 and $359/year for WHOOP One, Peak and Life; device is included in membership — 14+ days on WHOOP 5.0/MG — Without membership there is effectively no product; no screen or classic smartwatch features

Garmin — Sport, GPS, training readiness, HRV Status, Body Battery — Hardware is purchased separately; core wellness and sport metrics do not require an Oura/WHOOP-style wellness subscription — Strongly model-dependent; Venu 4 is rated up to 12 days in smartwatch mode — Watches can be less comfortable for sleep than a ring or strap

Ultrahuman — Ring for sleep and recovery without mandatory subscription — Ring AIR is listed from $349; access to ring data has no recurring subscription fee — 4-6 days — Less public clarity on API/export and less sport depth than Garmin or WHOOP

Sleep and HRV: where rings are stronger, and watches are not always worse

For night tracking, a ring usually wins in everyday life: it is lighter, has no glowing screen, does not catch on sleeves and is easier to forget. That makes Oura and Ultrahuman natural candidates if your main goal is to understand sleep, HRV, night heart rate, temperature and recovery.

Oura Ring 4 measures night heart rate, HRV, breathing, SpO2, temperature and movement. Active membership unlocks detailed sleep analysis, Readiness, Daytime Stress, Resilience, Cardiovascular Age, VO2 Max estimate, reports and Oura Labs. One important detail: Oura Labs is experimental. For example, the Blood Pressure Profile Study is available only in English for U.S.-based members, does not provide blood pressure measurements and is not a diagnostic feature.

Ultrahuman Ring AIR plays in a similar category, but its pitch is different: no required subscription. The product page lists Sleep Score, sleep stages, Dynamic Recovery, skin temperature, continuous HRV monitoring, stress rhythm and caffeine window. For someone who wants a ring and data without a monthly bill, that is a strong argument. But if you are building your own quantified-self pipeline, check how you will actually get data out of Ultrahuman before you buy. In open official material, it is easier to find the promise of data access without subscription than a complete public export or API document.

WHOOP is also strong for sleep, but its logic is different. It is less a night-only ring and more a continuous recovery coach. Sleep, Recovery and Strain are connected with training, the habit journal, stress and an AI layer. That is useful if you are willing to tell the system what you drank, how you trained, whether you traveled, got sick or took supplements.

Garmin should not be dismissed. HRV Status is built from overnight readings and needs roughly three weeks of regular sleep wearing the watch to establish a personal baseline. Body Battery uses HRV, stress and activity to estimate your energy reserve through the day. If you already wear Garmin around the clock, a separate ring may be redundant. If the watch bothers you at night, the data will have holes.

Sport and recovery: Garmin and WHOOP play a different game

Rings are comfortable for sleep, but sport is a harder environment. Barbells, kettlebells, pull-up bars, climbing, contact sports and even a tight bike grip expose the weakness of the form factor: the ring either gets in the way or you want to take it off. If you remove it for training, the recovery story after load becomes less connected.

Garmin is the most practical choice here. Even a wellness-oriented model like Venu 4 includes HRV Status, Body Battery, recovery time, training readiness, wrist-based running dynamics, women’s health tracking, Garmin Fitness Coach and up to 12 days of battery life in smartwatch mode. Dedicated sport lines such as Forerunner, fenix, epix, Instinct and Enduro go further: GPS, pace, power, routes, sensors, workouts, racing and long activities.

WHOOP is strong if you do not need a screen or wrist GPS, but you do want recovery discipline. Strain shows how heavy the day was, Recovery suggests whether to push or hold back, and the habit journal connects behavior to results. For strength work and team sports, a strap is easier to keep on than a ring. But for pace, maps, navigation and glanceable workout prompts, WHOOP does not replace Garmin.

Oura and Ultrahuman can work as a night layer next to sport watches. Example: Garmin for workouts, a ring for sleep. It costs more, but it removes the conflict between the best form factor for sport and the best form factor for sleep.

Stress, health and longevity: do not confuse guidance with diagnosis

All four products sell a similar promise: you can see HRV, sleep, heart rate, temperature and stress, then finally understand what is happening in your body. In practice, a wearable is useful less as a mini-doctor and more as a trend sensor. It can help you notice that alcohol, late meals, flights, illness or overreaching tank your recovery. It should not make medical decisions for you.

Oura is strong at soft interpretation: Readiness, Resilience, Daytime Stress, Cumulative Stress, Cardiovascular Age and Cardio Capacity make data easier to understand if you do not want to live in CSV files. WHOOP is moving toward a health platform: in 2026 the company announced on-demand clinician access in the U.S. for summer, EHR syncing, and new AI features called My Memory and Proactive Check-ins. That is interesting, but it is not the main reason to buy WHOOP outside the U.S. or if you do not want a wearable to hold more medical context.

Garmin uses a more athletic language: HRV Status, Body Battery, stress level, sleep, recovery time and training readiness. It feels less like a wellness journal and more like a dashboard for an active person. Ultrahuman focuses on nudges around sleep, recovery, temperature, stress, circadian rhythm and caffeine timing.

Regulated features need a separate warning. ECG, irregular rhythm notifications, blood pressure insights, AFib detection and similar functions almost always depend on country, model, user age, app version and regulatory status. WHOOP Life/MG, Garmin ECG App, Oura Labs and Ultrahuman PowerPlugs should not be compared as universal features for everyone. Before buying, check the exact country and exact model.

Women’s health: temperature helps, but it is not contraception by default

For cycle tracking, rings look stronger than watches because of night temperature and comfortable continuous wear. Oura offers Cycle Insights, Pregnancy Insights and integrations with apps such as Natural Cycles, Flo and Clue. That makes Oura the most mature option if you want temperature, sleep and women’s health in one app.

Garmin is also developing women’s health features: Venu 4 lists skin temperature for past ovulation estimates and improved period predictions. But Garmin explicitly says menstrual cycle tracking should not be used to support conception, contraception or birth control. That boundary matters: cycle predictions can help you observe patterns, but they are not a medical method.

WHOOP includes Women’s Hormonal Insights in its membership lineup. Ultrahuman lists temperature-based ovulation prediction. In both cases, the key questions are the same: availability in your country, clarity of the method, data export and how much you trust the app in a sensitive category.

Subscription, data and export: the hidden wearable cost

Buying a wearable is not only about the device price. After two years, the difference between buy it and wear it and pay to keep the insights becomes hard to ignore.

Oura: U.S. membership costs $5.99/month or $69.99/year. Over two years, that is about $140 with annual billing, on top of the ring price. Oura has an API V2, but for Gen3 and Ring 4 users, API access requires active Oura Membership. Data files can still be downloaded through Membership Hub, at least through a GDPR-compliant request path.

WHOOP: the model is fully subscription-based. In the U.S. storefront from the research set, WHOOP One, Peak and Life are listed from $199, $239 and $359/year, and the device is included in membership. WHOOP supports CSV data export and has a developer API, but continuous heart rate data is not available through the API. For WHOOP MG, some health data may be subject to separate retention rules.

Garmin: strongest for sport export. Garmin Connect can export activities in formats such as GPX, TCX, original/FIT and CSV, and users can request all-account data export through Garmin’s account tools. If you move workouts into Strava, TrainingPeaks, GoldenCheetah or your own spreadsheets, this may matter more than a beautiful sleep score.

Ultrahuman: the main advantage is no recurring subscription fee for Ring AIR data. The main question is data openness. If the app is enough for you, that may be fine. If you want to routinely extract raw or semi-structured data, do not buy blind: first check current export settings, API access and integrations.

Who Oura Ring is for

Oura is worth choosing if you want the most polished smart ring for sleep, recovery, HRV, readiness, stress and women’s health. It fits people who do not want to sleep with a sport watch, value clear reports and are comfortable paying for the app.

It is not the best fit if you are strongly against subscriptions, train heavily with barbells or want one device for GPS, pace, maps and workout screens. Oura is not a sport watch. It is a ring for sleep, recovery and long-term trends.

Who WHOOP is for

WHOOP fits people who think in systems: load, recovery, sleep, habits, stress, journal, coaching. It is especially logical if you train 3-6 times a week, do not want a screen on your wrist and are comfortable with annual billing.

It does not fit people who want to buy a device once, see notifications, use maps, play music, get quick workout screens or avoid a subscription model. WHOOP without membership is not a standalone gadget.

Who Garmin is for

Garmin is the choice for sport, data and autonomy. If you run, cycle, hike, prepare for races or want to connect sleep with training readiness, Garmin will usually give you more practical value for the money.

It is not ideal if you cannot sleep comfortably with a watch or want the most invisible tracker possible. Large sport models can be excellent by day and annoying at night. This is one of those cases where trying the device matters more than reading the spec sheet.

Who Ultrahuman Ring AIR is for

Ultrahuman Ring AIR is worth considering if you want a smart ring for sleep, recovery and temperature without a required monthly fee. It is especially interesting if you do not want HRV to become another subscription.

It is not the best choice if mature integrations, transparent API access, Garmin-level sport analytics or medical features with clear regional availability are critical to you. The potential is high, but check your exact use cases before buying.

Bottom line: choose the habit, not the metric

If you want one short answer: for sleep and soft longevity tracking, choose Oura if the subscription does not bother you. For training-led recovery, choose Garmin. For screenless recovery coaching, choose WHOOP. For a ring without a mandatory subscription, choose Ultrahuman.

But the best wearable is the one you do not take off. HRV, sleep score and readiness become useful only after weeks and months of continuous data. The final test is simple: can you sleep with it, charge it without irritation, pay for it without regret and change your behavior based on its prompts? If not, even the most accurate sensor becomes an expensive accessory.

Seed phrases solved one problem and created another: self-custody became portable, but also brutally unforgiving. A photo in the wrong cloud folder, a phishing page, a lost notebook, a house move, one distracted evening — and the security model collapses.

Account abstraction is the serious attempt to move crypto wallets away from that design. Instead of treating a wallet as one private key with a 12- or 24-word backup, it turns the account into programmable logic: passkeys, guardians, multisig approvals, daily limits, spending roles, session keys, batched transactions and gas paid in tokens can all become wallet features.

The short version: smart wallets are not automatically safer than seed phrases. They are safer when the recovery model, signer setup and app support match the way you actually use crypto.

What account abstraction actually changes

A traditional Ethereum wallet is usually an externally owned account, or EOA. It signs transactions with one private key. If that key is lost, there is no recovery. If that key is stolen, there is no built-in spending policy to slow the attacker down.

A smart account is a contract account controlled by rules. Those rules can say: accept a passkey signature, require two of three owners, allow a small USDC transfer without full approval, block suspicious calls, recover through guardians after a delay, or let a dapp sponsor gas.

ERC-4337 is the main standard that made this practical on Ethereum-style chains without changing Ethereum’s base protocol. Instead of a normal transaction sent directly by an EOA, the user signs a UserOperation. A bundler submits it through an EntryPoint contract, and an optional Paymaster can cover gas or accept another token for fees. That is the plumbing behind the smoother UX.

This matters because most crypto friction is not ideological. It is operational. New users do not want to buy ETH just to move USDC. Normal people do not want their life savings secured by a paper phrase. Teams do not want one founder’s laptop to be the treasury. Power users do not want to sign five transactions when one batched operation would do.

The comparison in one view

Wallet or stack — Best for — Recovery model — Gas UX — Main tradeoff

Argent / Ready — Mobile-first consumer crypto, seedless recovery — Guardians for Argent mobile; Ready also emphasizes consumer banking-style UX — Depends on product, chain and app flow — Product naming and feature split matter: Argent mobile, Argent X and Ready are not identical

Safe — Serious balances, teams, DAOs, families, power users — Multisig owners, replaceable signers, guardian-style recovery options — Strong controls, batching, spending limits; not always the simplest consumer flow — More setup and more responsibility in signer design

Coinbase/Base Smart Wallet — Fast passkey onboarding for apps and Base/EVM users — Passkeys, with recovery flow that should be set up before losing access — Designed to remove app-install and seed-phrase friction — Recovery still depends on passkey hygiene and supported flows

Ambire — EVM power users who want batching and gas abstraction — Supports EOA, smart account and EIP-7702-style flows — Gas Tank and token-fee UX are the draw — More of a power-user wallet than a minimal beginner app

Biconomy — Developers building wallet UX inside apps — Infrastructure layer, not a consumer wallet — Paymasters, bundlers, sponsored gas, token gas, cross-chain abstraction — Users depend on how each app integrates it

The mistake is asking which smart wallet is best in the abstract. The right question is: what failure are you most worried about?

If the fear is losing 12 words, choose a wallet with recovery you can test. If the fear is one compromised device draining a large balance, use multisig and spending limits. If the pain is onboarding users into an app, use passkeys and sponsored transactions. If the problem is long-term storage, a hardware wallet can still be part of the setup as a signer or guardian.

Ready / Argent: the clearest consumer answer to seed phrases

Argent helped popularize social recovery before account abstraction became a mainstream phrase. As of the latest official materials reviewed for this piece, Argent’s main consumer brand redirects to Ready, while Argent support still distinguishes between Argent mobile, Argent X and Argent web wallet.

That distinction matters. Argent mobile is described by Argent support as seedless and based on social recovery: users choose trusted people or devices called guardians, and a majority can authorize recovery if the phone is lost. Guardians can include trusted people, hardware wallets such as Ledger and Trezor , MetaMask and Argent itself. Argent also describes a 48-hour security period during which a recovery can be cancelled. Argent X, by contrast, uses a seed phrase as the primary recovery path.

Ready is positioning the same lineage closer to a crypto bank alternative: mobile app, cash deposit rails, card spending, swaps, staking and support. The site says Ready never has access to user assets, which is the key point: this is meant to remain self-custodial, not become a custodial fintech account with crypto branding.

Use it if you want crypto that feels closer to a mobile money app than a developer tool. The strongest fit is a user who wants to hold and spend stablecoins or mainstream crypto without managing raw private-key backups.

Do not use it as a lazy substitute for a security plan. You still need to understand which product you are using, which chain it supports, who or what your guardians are, and what happens if your phone, email, passkey provider or guardian set fails at the same time.

Safe: the serious default for assets you cannot casually lose

Safe is not the fastest path for a first-time user who just wants to try one dapp. It is the wallet you choose when single-key custody is the wrong shape for the risk.

Safe’s core model is multisig: instead of one key having full control, an account can require M-of-N approvals. For a personal setup, that might mean two of three signers: one hardware wallet at home, one mobile signer, one recovery signer stored separately. For a company or DAO, it might mean several operators with defined thresholds.

The important part is not the word multisig; it is removing the single point of failure. Safe also supports transaction simulation, transaction building, spending limits, roles, multichain treasury management and modules. Safe’s own site claims more than $1 trillion in processed volume, 57 million wallets deployed and more than $60 billion secured; treat those as company-reported metrics, but they explain why Safe has become the default reference point for onchain treasuries.

Spending limits are especially practical. Safe’s help docs describe a module that lets a beneficiary address spend a specified token up to a one-time or periodic limit without requiring the full signer threshold each time. That is exactly the kind of feature seed-phrase wallets do not natively have: routine activity can be convenient while large movements stay protected.

Use Safe for treasuries, shared funds, family setups, founder wallets, protocol admin keys, NFT vaults and serious DeFi positions. It is also a good personal wallet if you are willing to design the signer setup carefully.

Do not use it if you want a one-minute beginner flow and have no patience for operational security. A bad multisig setup can still fail. Three signers stored in the same laptop bag are not three meaningful signers. A threshold nobody can meet during an emergency is not resilience.

Coinbase/Base Smart Wallet: passkeys make onboarding feel normal

Coinbase Smart Wallet, now closely tied to the Base wallet experience in Coinbase’s help materials, represents the passkey route: users can sign in with passkeys instead of memorizing a recovery phrase. Cloud-backed passkeys through Apple, Google or a password manager can make wallet access feel much closer to modern app login.

That is a huge UX improvement, especially for apps that do not want users to install a browser extension before doing anything. But it does not make recovery irrelevant. Coinbase’s own help explains that users can set up a recovery phrase for the smart wallet to recover from passkey loss, and that signer changes are onchain actions that may require network fees. The practical lesson is simple: set up recovery while you still have access. Do not wait until the passkey is gone.

Use this model for low-friction onboarding, consumer apps, Base-heavy usage and users who already trust passkey managers. Be more cautious for high-value cold storage, unsupported chains, or dapps that still expect old EOA behavior.

Ambire: smart-wallet ergonomics for EVM power users

Ambire is worth watching because it sits between consumer wallet and power-user control panel. Its help docs describe support for EOA accounts, smart accounts, EIP-7702-style SmarterEOA flows, transaction batching, human-readable transactions and a Gas Tank that exists because of gas abstraction.

That makes Ambire interesting for users who understand EVM networks and want better transaction ergonomics without moving everything into a team-style Safe. The Gas Tank idea is simple: pre-fund fee handling so every transaction does not become a scramble for the right native token on the right chain.

The tradeoff is complexity. Ambire can be more capable than a simple mobile wallet, but that also means users need to understand which account type they are operating and how recovery works for that account.

Biconomy: not your wallet, but the layer that makes apps feel walletless

Biconomy is different from the others because it is infrastructure. Most users will not choose it the way they choose Safe or Ready. They will experience it when a dapp uses Biconomy to sponsor gas, accept token fees, batch actions or coordinate cross-chain execution.

Biconomy’s docs describe its Modular Execution Environment as implementing ERC-4337 bundler and paymaster functionality with additional cross-chain orchestration. Its Paymaster documentation covers sponsored gas and token-pay gas models, including stablecoin-style fee payment.

For builders, this is where account abstraction becomes product design. A game can sponsor early user actions. A DeFi app can let users pay fees in USDC. A consumer app can hide chain plumbing until the user actually needs to understand it.

For users, the caveat is that gasless does not mean costless. Someone pays. Sometimes it is the app, sometimes it is a token Paymaster, sometimes a policy engine rejects the transaction. The UX can feel Web2-like, but the trust and failure model is still crypto infrastructure.

The hidden tradeoffs: smart wallets are better, not magical

Account abstraction removes many old annoyances, but it introduces new surfaces to inspect.

First, smart accounts are code. That means audits, upgradeability, modules, signer permissions and storage layout matter. A mature wallet with battle-tested contracts is preferable to a shiny wallet with vague security claims.

Second, recovery is a policy, not a miracle. Social recovery protects against losing one device; it does not protect against choosing careless guardians, losing access to every recovery channel, or approving a malicious recovery request. Passkeys are excellent UX, but device sync, cloud accounts, password managers and hardware security keys all have different recovery assumptions.

Third, dapp compatibility is still uneven. Many apps now understand smart wallets and EIP-1271 signatures, but not all of crypto was designed around contract accounts. You can still hit flows that assume a classic EOA.

Fourth, gas abstraction depends on infrastructure. Paymasters, bundlers and relayers can improve UX, but availability, sponsorship rules, supported chains and rate limits are app-specific. A wallet that feels gasless in one app may feel very normal in another.

Fifth, fees still exist. Deploying a smart account, changing signers, recovering an account, executing modules or using Ethereum mainnet can cost money. Account abstraction can reduce friction; it does not repeal blockspace economics.

How to choose

Choose Ready / Argent if you want the most consumer-friendly path away from seed phrases and you are comfortable with a mobile-first wallet. Confirm which product you are using: Argent mobile, Argent X and Ready do not have identical recovery models.

Choose Safe if the assets matter enough to justify setup time. For most serious crypto users, the best answer is not one hot wallet; it is a Safe with a sensible threshold, hardware-wallet signers, a recovery plan and limited daily spending permissions.

Choose Coinbase/Base Smart Wallet if you want passkey onboarding and mostly live inside apps that support it well. It is a strong UX pattern for mainstream adoption, but you should configure recovery before there is an emergency.

Choose Ambire if you are an EVM user who wants batching, gas abstraction and more control than a simple mobile wallet gives you.

Choose Biconomy if you are building an app and want users to stop thinking about gas, bridges and network setup on every click.

For long-term storage, do not frame this as smart wallet versus hardware wallet. The better design is often smart wallet plus hardware wallet: the smart account provides rules, recovery and limits; the hardware wallet acts as a high-assurance signer or guardian.

Bottom line

The endgame is not a wallet with no security burden. That product should not exist. The endgame is a wallet where the security burden matches human behavior.

Seed phrases ask users to become perfect secret managers. Smart wallets let users design failure-tolerant accounts: a lost phone is recoverable, a compromised signer is removable, small payments are convenient, large transfers require more friction, and gas can be abstracted away when the app is willing to pay for the UX.

For everyday crypto, Ready / Argent and Coinbase-style passkey wallets make the first mile much better. For meaningful balances, Safe remains the more defensible default. For apps, Biconomy and similar infrastructure are how account abstraction becomes invisible.

The practical recommendation: start small, test recovery before funding heavily, separate daily spending from long-term storage, and never assume no seed phrase means no responsibility.

Local AI agents are no longer just a toy demo you run once and forget. With Ollama handling local models and LangGraph handling stateful orchestration, you can build a private agent system that plans, delegates, checks its own work, and resumes from checkpoints without sending your task history to a hosted LLM API.

This guide builds a compact but real agent swarm: a supervisor, a researcher, an architect, a coder, a reviewer, and a finalizer. It runs on your own machine, uses a local Ollama model, and keeps the control flow explicit enough that you can debug it when the model takes a strange turn.

What we are building

The word swarm is often abused. Here it means a group of specialized agents coordinated through shared state, not a magic pile of prompts talking over each other.

The architecture has four layers:

1. Ollama runs the local LLM and exposes it through a local API.
2. LangChain’s ChatOllama integration gives Python code a clean chat-model interface.
3. LangGraph turns the workflow into a state machine with nodes, edges, routing, and checkpointing.
4. Your agents are just functions with narrow responsibilities and explicit state updates.

The important design choice: the supervisor does not let every agent speak at once. It routes work step by step. That is slower than naive parallelism, but easier to control, inspect, and recover.

Why run agents locally

Local agents make sense when the task contains private context, internal documents, code, customer data, personal notes, or anything you do not want copied into a third-party inference service.

They are also attractive when you want predictable cost. The software stack in this guide can run without a monthly LLM subscription if your machine is strong enough for the model you choose. That does not make local inference free: you still pay in hardware, electricity, setup time, and slower iteration. But the cost curve is different. You trade API bills for ownership and operational responsibility.

Use this setup for:

• private code review and refactoring plans;
• document triage over local files;
• repeatable research workflows over an internal knowledge base;
• offline drafting and summarization;
• experimentation with agent architectures before moving to managed infrastructure.

Do not use it as-is for:

• high-risk actions such as payments, account deletion, or production database writes;
• workloads that need the best frontier-model reasoning;
• large-scale concurrent serving;
• tasks where latency matters more than privacy or cost control.

Prerequisites

You need Python 3.11 or newer, Ollama installed, and at least one local model pulled. Start with a smaller general model while debugging the graph. Bigger models may reason better, but they also make every broken loop more expensive.

Install Ollama from Ollama , then pull a model from the Ollama model library. The exact best model changes quickly, so check the model tags before you build around one name. The examples below use qwen3 because Ollama’s current docs use it in tool-calling examples.

ollama pull qwen3 ollama run qwen3

In another terminal, check that the local API responds. Ollama’s default local API base URL is http://localhost:11434/api.

curl http://localhost:11434/api/generate -d '{ "model": "qwen3", "prompt": "Say hello in one sentence.", "stream": false }'

Now create a Python environment and install the orchestration packages.

mkdir local-agent-swarm cd local-agent-swarm python -m venv .venv source .venv/bin/activate pip install -U langgraph langchain-ollama langchain-core pydantic typing-extensions

The complete local swarm

Create local_swarm.py and paste this code.

from __future__ import annotations import operator import os from typing import Annotated, Literal from langchain_ollama import ChatOllama from langgraph.checkpoint.memory import InMemorySaver from langgraph.graph import END, START, StateGraph from pydantic import BaseModel, Field from typing_extensions import TypedDict MODEL = os.getenv('OLLAMA_MODEL', 'qwen3') llm = ChatOllama( model=MODEL, temperature=0, ) class Route(BaseModel): next_agent: Literal['researcher', 'architect', 'coder', 'reviewer', 'finalizer'] = Field( description='The specialist that should run next.' ) task: str = Field(description='A short task for that specialist.') reason: str = Field(description='Why this specialist is the next best step.') class SwarmState(TypedDict): goal: str current_task: str active_task: str iteration: int max_iterations: int supervisor_notes: Annotated[list[str], operator.add] research: Annotated[list[str], operator.add] architecture: Annotated[list[str], operator.add] code: Annotated[list[str], operator.add] review: Annotated[list[str], operator.add] final: str router = llm.with_structured_output(Route) def make_brief(state: SwarmState) -> str: nl = chr(10) parts = [] for key in ['supervisor_notes', 'research', 'architecture', 'code', 'review']: values = state.get(key, []) if values: parts.append(key.upper() + ':' + nl + nl.join(values[-2:])) return nl + nl.join(parts) if parts else 'No shared notes yet.' def supervisor(state: SwarmState) -> dict: if state['iteration'] >= state['max_iterations']: return { 'active_task': 'finalizer', 'current_task': 'Prepare the best possible final answer from the work completed so far.', 'supervisor_notes': ['Stopped by max_iterations to prevent an infinite loop.'], } decision = router.invoke([ ('system', 'You are the supervisor of a local multi-agent workflow. Route one next step only. Prefer finalizer when the answer is ready. Do not loop unless a concrete gap remains.'), ('human', 'Goal:' + chr(10) + state['goal'] + chr(10) + chr(10) + 'Current shared state:' + chr(10) + make_brief(state)), ]) return { 'active_task': decision.next_agent, 'current_task': decision.task, 'iteration': state['iteration'] + 1, 'supervisor_notes': ['Route to ' + decision.next_agent + ': ' + decision.reason], } def call_agent(name: str, system_prompt: str, state: SwarmState) -> str: response = llm.invoke([ ('system', system_prompt), ('human', 'Goal:' + chr(10) + state['goal'] + chr(10) + chr(10) + 'Assigned task:' + chr(10) + state['current_task'] + chr(10) + chr(10) + 'Shared state:' + chr(10) + make_brief(state)), ]) return name + ':' + chr(10) + response.content.strip() def researcher(state: SwarmState) -> dict: content = call_agent( 'Researcher', 'Find assumptions, missing facts, constraints, and useful context. Do not invent external facts. If evidence is missing, say so clearly.', state, ) return {'research': [content]} def architect(state: SwarmState) -> dict: content = call_agent( 'Architect', 'Design the workflow, data structures, failure boundaries, and tradeoffs. Be specific and implementation-oriented.', state, ) return {'architecture': [content]} def coder(state: SwarmState) -> dict: content = call_agent( 'Coder', 'Produce concise implementation steps or code. Prefer simple, testable Python. Mention assumptions that affect correctness.', state, ) return {'code': [content]} def reviewer(state: SwarmState) -> dict: content = call_agent( 'Reviewer', 'Review the proposed work for bugs, missing checks, unsafe actions, and vague claims. Return concrete fixes.', state, ) return {'review': [content]} def finalizer(state: SwarmState) -> dict: response = llm.invoke([ ('system', 'Write the final answer. Use the shared state, resolve contradictions, and be direct. Do not mention internal agent names unless useful.'), ('human', 'Goal:' + chr(10) + state['goal'] + chr(10) + chr(10) + 'Shared state:' + chr(10) + make_brief(state)), ]) return {'final': response.content.strip()} def route_from_supervisor(state: SwarmState) -> str: return state['active_task'] builder = StateGraph(SwarmState) builder.add_node('supervisor', supervisor) builder.add_node('researcher', researcher) builder.add_node('architect', architect) builder.add_node('coder', coder) builder.add_node('reviewer', reviewer) builder.add_node('finalizer', finalizer) builder.add_edge(START, 'supervisor') builder.add_conditional_edges( 'supervisor', route_from_supervisor, { 'researcher': 'researcher', 'architect': 'architect', 'coder': 'coder', 'reviewer': 'reviewer', 'finalizer': 'finalizer', }, ) for node in ['researcher', 'architect', 'coder', 'reviewer']: builder.add_edge(node, 'supervisor') builder.add_edge('finalizer', END) graph = builder.compile(checkpointer=InMemorySaver()) if __name__ == '__main__': initial_state: SwarmState = { 'goal': 'Design a local document summarizer for private meeting notes. Include architecture, risks, and a minimal implementation plan.', 'current_task': '', 'active_task': '', 'iteration': 0, 'max_iterations': 8, 'supervisor_notes': [], 'research': [], 'architecture': [], 'code': [], 'review': [], 'final': '', } config = {'configurable': {'thread_id': 'local-swarm-demo-001'}} result = graph.invoke(initial_state, config=config) print(result['final'])

Run it:

python local_swarm.py

To try a different model without editing code:

OLLAMA_MODEL=gemma3 python local_swarm.py

How the graph works

LangGraph models work as graphs: nodes do work, edges decide where execution goes next, and state carries the memory of the run. START and END are special graph markers. Conditional edges let the supervisor choose which node runs next.

In this example, the supervisor is the only router. That keeps the graph legible:

START -> supervisor -> specialist -> supervisor -> specialist -> supervisor -> finalizer -> END

The state has append-only lists for research, architecture, code, review, and supervisor notes. Those list fields use reducers, so each specialist can add information without overwriting previous work. The scalar fields, such as current_task and active_task, are overwritten on each supervisor step.

That distinction matters. Most broken agent systems fail because memory is a pile of unstructured chat logs. A graph state gives every part of the workflow a known place to write.

Why checkpointing matters

The example uses InMemorySaver because it is simple and works for a local demo. It lets LangGraph associate a run with a thread_id, which is also the mechanism used for pause and resume patterns.

For production, in-memory checkpoints are not enough. If the Python process dies, the checkpoint goes with it. LangGraph’s own reference recommends durable checkpoint stores, such as Postgres-based checkpointing, for production use. Treat InMemorySaver as a debugger and teaching tool, not as your reliability layer.

A practical production version should add:

• a durable checkpointer;
• stable thread IDs per user, task, or document;
• model and prompt version metadata in state;
• trace logs for every route decision;
• a maximum iteration limit;
• explicit human approval before irreversible tools run.

Adding tools without losing control

Ollama supports tool calling for models that can use tools, and LangChain’s ChatOllama integration exposes tool binding. That is useful, but do not start by giving a local agent shell access, filesystem writes, browser automation, and database credentials.

Start with narrow, boring tools:

• read a specific approved directory;
• search a local vector index;
• summarize a single file;
• create a draft patch but not apply it;
• validate JSON against a schema.

Then put human approval in front of dangerous actions. LangGraph’s interrupt pattern is designed for this: a node can pause execution, surface the proposed action, and resume only after a human decision. That is the difference between a useful autonomous workflow and a local model with too much authority.

Privacy model: what stays local and what does not

If you run Ollama locally and use only local models, prompts and outputs are processed on your machine. That is the main privacy benefit of this stack.

But privacy can disappear through the tools you attach. If an agent calls a web search API, sends telemetry, uploads logs, or writes to a cloud vector database, the workflow is no longer fully local. The model runtime is only one part of the data path.

Before using this on sensitive material, check:

• whether Ollama is using a local model or a cloud model;
• whether any tool sends text outside the machine;
• where checkpoints are stored;
• whether logs contain raw prompts or documents;
• whether your model license allows your intended use.

Performance expectations

A local swarm is slower than a single prompt because it makes multiple model calls. It can still be worth it when each step improves quality: one agent scopes the problem, another designs the approach, another writes code, and another reviews the result.

If runs feel too slow, reduce max_iterations before changing models. Then shorten prompts, shrink shared state, or route fewer specialist steps. Bigger models are not a substitute for a clean graph.

Good defaults for early experiments:

• temperature 0 for routing and review;
• small max_iterations, usually 5 to 10;
• one supervisor, not peer-to-peer chaos;
• structured output for routing;
• final answer generated only after review has had a chance to run.

Common failure modes

The supervisor loops forever. Add max_iterations, stronger finalization criteria, and route logs.

The model ignores the requested JSON or structured output. Use a model with stronger structured-output behavior, reduce the schema, or add a fallback parser.

Specialists repeat each other. Give each node a narrower system prompt and show only the latest relevant state, not the entire transcript forever.

The final answer includes false certainty. Make the researcher and reviewer explicitly mark missing evidence. Do not ask a local model to invent facts it cannot verify.

The system is private but not safe. Privacy and safety are separate. A local agent can still delete files, leak secrets to a tool, or generate bad instructions. Gate tools by default.

When to move beyond this demo

This implementation is enough for a serious local prototype. Move to a stronger setup when you need multiple users, long-running jobs, audit trails, or real tool execution.

The next step is not adding more agents. It is making the existing graph observable and durable:

• swap InMemorySaver for a persistent checkpointer;
• stream intermediate state to a small local UI;
• add human approval for tool calls;
• store documents in a local vector database;
• evaluate outputs against a fixed test set;
• use a GPU cloud provider such as RunPod only when local hardware is the bottleneck and the data can safely leave the machine.

Bottom line

LangGraph and Ollama are a strong pair because they solve different problems. Ollama makes local model execution approachable. LangGraph makes agent control flow explicit, inspectable, and recoverable.

The result is not a magical autonomous employee. It is a private workflow engine powered by local LLM calls. That is more useful: you can see what it is doing, limit what it can touch, and improve the graph one node at a time.

The EU AI Act is not “another AI policy page” for small businesses. It is a product, procurement, documentation and workflow problem with legal deadlines attached.

The headline date is 2 August 2026: under the official AI Act timeline, most remaining provisions are due to start applying then. But the practical picture is more nuanced. On 7 May 2026, the Council of the EU and the European Parliament reached a provisional political agreement on AI simplification rules that would delay parts of the high-risk regime: standalone high-risk AI systems would move to 2 December 2027, and high-risk systems embedded in products would move to 2 August 2028. As of 22 May 2026, that agreement still needs formal adoption and publication before businesses can treat it as settled law.

So the useful question is not “does this apply to us?” It is: where are we using AI, what role do we play, and which parts must be ready by which date?

The short version

If you run a small company, you probably do not need a huge AI compliance department. You do need a clear map of AI use inside the business.

Start with six moves:

1. List every AI system your company builds, sells, integrates or uses.
2. Decide whether you are a provider, deployer, importer, distributor or product manufacturer for each system.
3. Classify each system: prohibited, high-risk, transparency-triggering, general-purpose AI related, or ordinary low-risk use.
4. Ask vendors for evidence, not slogans.
5. Add human review, logs and user-facing notices where the Act expects transparency or oversight.
6. Track the May 2026 simplification package before locking your final deadline plan.

This is not legal advice. It is an operational guide for founders, ops leads, product managers and small-business owners who need to turn a dense regulation into a workable checklist.

What actually changes on 2 August 2026

The AI Act entered into force on 1 August 2024. It applies in stages.

Some rules are already active by the time you are reading this. The bans on prohibited AI practices and AI literacy obligations started applying on 2 February 2025. Obligations for providers of general-purpose AI models started applying on 2 August 2025.

The next major date is 2 August 2026. The European Commission describes the Act as becoming fully applicable two years after entry into force, with exceptions. The AI Act Service Desk also lists 2 August 2026 as the date when the remainder of the Act starts to apply, except for specific provisions.

The catch: “fully applicable” does not mean every compliance duty for every AI system lands on the same day.

The most important uncertainty is high-risk AI. The May 2026 provisional agreement would set later application dates for high-risk rules: 2 December 2027 for standalone high-risk systems and 2 August 2028 for high-risk systems embedded in products. Until the amendments are formally adopted and published, treat this as a likely but not final change.

First, find your role

The same AI tool can create different obligations depending on what your company does with it.

Provider

You are likely a provider if you develop an AI system or general-purpose AI model and place it on the EU market or put it into service under your name or trademark.

For a small software company, this can happen faster than expected. If you wrap a model into a SaaS product, market the AI feature as your product, and sell it to EU customers, you may not be “just using ChatGPT.” You may be providing an AI system.

Deployer

You are likely a deployer if you use an AI system in a professional context.

A retailer using AI for customer support, a recruiter using an AI ranking tool, a clinic using AI transcription, or an agency using generative AI for client work may all be deployers. Deployer obligations are usually lighter than provider obligations, but they are not zero.

Importer or distributor

You may be an importer or distributor if you make an AI system from outside the EU available in the EU market. This matters for resellers, marketplaces, integrators and channel partners.

Product manufacturer

If AI is embedded in a regulated product, the AI Act can interact with product safety rules. This is especially relevant for medical devices, machinery, toys, lifts, vehicles and other regulated categories.

For most small businesses, the provider-versus-deployer distinction is the first fork in the road.

Then classify the AI use

The AI Act uses a risk-based structure. Small companies should not start by reading every article line by line. Start by sorting systems into practical buckets.

Prohibited AI practices

Some practices are banned. The exact boundaries matter, but the general category includes AI uses the EU treats as unacceptable risk. If your use case touches manipulation, exploitation of vulnerabilities, social scoring, certain biometric categorisation, or real-time remote biometric identification in public spaces, stop and get specialist advice.

For ordinary small businesses, this bucket is less common than the others. But it is the first thing to rule out because the tolerance is low.

High-risk AI systems

High-risk is where the Act becomes operationally heavy.

The obvious small-business traps are not futuristic robots. They are everyday tools in sensitive domains: hiring, worker management, education, access to essential private or public services, creditworthiness, biometric identification, critical infrastructure, law enforcement, migration and justice-related uses.

If your company uses AI to rank job applicants, score employees, filter students, assess eligibility, recommend credit decisions or influence access to important services, you should assume this needs serious review.

Under the current political agreement, many high-risk obligations may be delayed beyond 2 August 2026, but that does not make them irrelevant. Procurement cycles, documentation, vendor negotiations and workflow redesign often take months.

Transparency-triggering AI

Some AI uses require people to be told what is happening.

This is especially relevant for chatbots, synthetic audio, synthetic video, image generation, deepfake-style content and AI systems that interact directly with people. A small marketing team using ElevenLabs , HeyGen or Runway for synthetic media should think less about “AI content is cool” and more about disclosure, consent, rights, provenance and recordkeeping.

That does not mean every AI-generated image needs a dramatic warning label in every context. It does mean the team needs a policy for when content is synthetic, when people could reasonably mistake it for real, and how disclosure is handled.

General-purpose AI model exposure

Most small businesses are not providers of general-purpose AI models. They are customers of model providers.

If you use frontier models through APIs or commercial tools, the model provider carries the main GPAI provider obligations. But you still need to understand what the model is doing inside your product, what data you send to it, what outputs affect users, and what your own product claims.

If you fine-tune, package or redistribute a model, the analysis changes.

Low-risk internal productivity use

Using AI to draft emails, summarize meeting notes, brainstorm copy or generate first-pass code is usually a lower-risk category under the AI Act. That does not mean it is risk-free. Privacy, confidentiality, copyright, security and employment rules can still matter.

But for AI Act triage, low-risk internal productivity use should not consume the same compliance budget as hiring, credit, health or education decisions.

The small-business AI inventory

You cannot comply with what you cannot see. The most useful first artifact is a simple AI register.

Create one row per system or workflow. Include:

• system name;
• vendor or internal owner;
• business purpose;
• users and affected people;
• input data;
• output and how it is used;
• whether the output influences decisions about people;
• EU users or EU market exposure;
• company role: provider, deployer, importer, distributor or manufacturer;
• likely risk category;
• human review point;
• vendor documentation received;
• retention, logging and incident process;
• next review date.

This can start as a spreadsheet. The point is not tool sophistication. The point is that someone can ask “where do we use AI?” and get a defensible answer in one place.

What to ask vendors

Small companies often hear the phrase “AI Act compliant” from vendors. That phrase is too vague to be useful.

Ask sharper questions:

• What role do you consider yourself under the AI Act for this product?
• Do you consider this system high-risk in any intended use?
• What uses do you prohibit in your terms?
• What documentation can you provide for risk management, data governance, testing, logging, human oversight and accuracy?
• Does the product generate synthetic audio, image, video or text that needs disclosure?
• Where is data processed and retained?
• Can we disable AI features we do not need?
• How will you notify customers about substantial model or system changes?
• Do you support audit logs and exportable records?

If the vendor gives only marketing copy, treat that as a procurement risk. For high-impact workflows, you need evidence you can keep.

What to do if you use AI in hiring

Recruiting is one of the easiest places for small businesses to underestimate the AI Act.

If an AI tool screens, ranks, scores or recommends candidates, the system may fall into a high-risk employment category. Even when the vendor is the provider, the employer using the system may still have deployer responsibilities.

Practical steps:

• identify whether the tool merely assists admin work or influences candidate selection;
• avoid fully automated rejection without meaningful human review;
• keep records of how recommendations are used;
• check whether candidates need notice;
• ask the vendor for high-risk documentation and intended-use statements;
• test whether the tool behaves differently across protected groups where lawful and feasible;
• make sure hiring managers understand the limits of the score or ranking.

The key distinction is influence. A scheduling assistant is one thing. A candidate ranking system that decides who gets interviewed is another.

What to do if you create synthetic media

Generative media tools are now ordinary business software. The compliance issue is that synthetic content can mislead people.

For marketing, education, training and support teams, build a lightweight disclosure policy:

• disclose AI-generated or AI-manipulated media when a reasonable person could mistake it for real;
• get explicit permission before cloning a real person’s voice or likeness;
• keep release forms and source files;
• label fictional avatars clearly in sensitive contexts;
• avoid synthetic endorsements from people who did not give permission;
• review local consumer protection, advertising and privacy rules alongside the AI Act.

If you use voice or avatar products such as ElevenLabs , HeyGen , Hume AI or video-generation tools such as Runway , the operational question is not whether the tool is impressive. It is whether your audience can tell what is real, what is synthetic and who authorized it.

What to do if you build AI into your own product

If your startup or agency ships an AI feature to customers, treat the AI Act as part of product management.

Before launch, write down:

• intended use;
• prohibited or unsupported uses;
• model or vendor dependencies;
• input and output data flows;
• failure modes;
• human oversight design;
• logging and monitoring;
• user-facing disclosures;
• escalation path for harmful or wrong outputs;
• update process when the model changes.

This is especially important if your product touches employment, education, finance, health, insurance, public services or safety-sensitive decisions.

For ordinary AI assistants, the documentation can be short. For high-risk or borderline systems, it needs to become real compliance evidence.

What not to overdo

Small businesses can waste money by treating every AI use as if it were a certified medical device.

Do not start with a 90-page policy nobody reads. Do not buy a compliance platform before you know your AI inventory. Do not rely on a vendor badge without understanding your own role. Do not assume that because a tool is popular, your use of it is low-risk.

A good first version is boring and useful: a register, a role map, a risk classification, a vendor evidence folder, a disclosure policy and a review calendar.

A practical 30-day plan

Week 1: map the AI footprint

Find every AI system in use: official tools, browser extensions, embedded SaaS features, API integrations, internal automations and experiments. Include tools used by marketing, HR, support, sales, finance and engineering.

Week 2: classify and prioritize

Mark systems that affect people’s opportunities, rights, access, eligibility, employment, education, credit, health, safety or essential services. These get priority.

Separate internal productivity tools from systems that face customers or influence decisions.

Week 3: collect evidence

Ask vendors for AI Act position statements, technical documentation, data processing information, transparency features, logs and change-notification processes.

For internal systems, write your own one-page system note.

Week 4: fix the obvious gaps

Add notices where users interact with AI. Add human review where outputs influence important decisions. Remove AI features nobody owns. Restrict risky uses in policy and product settings. Assign an owner for quarterly review.

The deadline that matters most

For many small businesses, 2 August 2026 is the date that forces the inventory conversation. It is not the end of the story.

Some obligations already apply. Some are due on 2 August 2026. Some high-risk obligations may move to 2027 or 2028 if the May 2026 simplification package is formally adopted. Some legacy and public-authority-related provisions have their own timelines.

That uncertainty is exactly why small businesses should start with facts they control: what AI they use, what it does, who it affects, what vendors can prove, and where human judgment remains in the loop.

The companies that handle this well will not be the ones with the thickest policy PDF. They will be the ones that can answer a simple question without panic: “Show me where AI makes or influences decisions in this business.”

Epigenetic clocks used to answer a dramatic but blunt question: how old does your body look compared with your birth certificate? The newer question is more useful: how fast are you aging now, and did your last intervention actually move the needle?

That shift is why third-generation epigenetic clocks matter. Tools such as DunedinPACE are designed less like a lifetime damage score and more like a speedometer. They do not replace blood work, imaging, fitness testing, or clinical judgment. But used carefully, they can add a missing layer to a longevity plan: whether your biology is trending in the right direction after changes in sleep, training, nutrition, weight loss, inflammation control, or medication.

The three generations of epigenetic clocks

DNA methylation is one way cells regulate gene activity. Across life, methylation patterns change in predictable ways. Epigenetic clocks use those patterns to estimate aging-related signals from a blood, saliva, or tissue sample.

First generation: chronological age predictors

The first major clocks, including early Horvath- and Hannum-style models, were built to predict chronological age from methylation. They were scientifically important because they showed that methylation carries a strong age signal across tissues.

Their weakness is practical: if a model is trained mainly to guess your calendar age, it may not be the best tool for deciding whether your health trajectory is improving. A 45-year-old with excellent cardiometabolic health and a 45-year-old with rising inflammation may both score close to 45.

Second generation: risk-linked biological age

Second-generation clocks moved closer to health outcomes. PhenoAge and GrimAge are commonly cited examples. They were designed around mortality risk, clinical chemistry, smoking-related methylation signatures, and other health-relevant endpoints rather than calendar age alone.

This made them more useful for risk stratification. If your GrimAge acceleration is high, that is more concerning than simply being methylation-estimated as older than your passport. But these clocks still often behave like accumulated-risk snapshots. They can be powerful, yet they may not be ideal for short feedback loops.

Third generation: pace of aging

Third-generation clocks try to estimate the rate of biological aging. DunedinPACE is the best-known commercial-facing example. It was developed from longitudinal data in the Dunedin Study, where researchers tracked changes across multiple organ-system biomarkers over time and then trained a methylation measure to estimate that pace.

In plain English: instead of asking “how old do you look?”, it asks “how much biological change are you accumulating per year?” A DunedinPACE value around 1.0 is often interpreted as roughly one biological year of change per chronological year. A higher value suggests faster pace; a lower value suggests slower pace. Exact interpretation depends on the lab, sample type, report version, and reference population.

Why rate-of-aging is more actionable

Longevity interventions are usually iterative. You change something, wait, measure, and adjust. A rate-of-aging marker fits that loop better than a static age estimate.

If you start resistance training, improve protein intake, reduce visceral fat, treat sleep apnea, or lower chronic inflammation, you do not want to wait decades to see whether the decision mattered. You want early directional evidence.

That does not mean DunedinPACE can validate every supplement stack or biohacking protocol. It means it may be useful as one layer in a measurement system, especially when paired with standard markers: ApoB, blood pressure, HbA1c, fasting insulin, hs-CRP, liver enzymes, kidney function, VO2 max, grip strength, waist-to-height ratio, sleep metrics, and medication history.

What a commercial test can actually tell you

A serious methylation report can be useful in four ways.

First, it gives you a baseline. If you have never measured biological aging, a first test is less a verdict than a starting line.

Second, it can track direction after interventions. Repeating the same test under similar conditions is more informative than comparing one-off results across different vendors.

Third, some reports bundle adjacent signals: estimated smoking exposure, inflammation-related methylation, immune-cell composition, organ-system aging, or disease-risk scores. These can generate better questions for your clinician, though they are not diagnoses.

Fourth, it can reveal mismatch. If your traditional labs look good but your pace-of-aging score is unexpectedly high, the next move is not panic. It is a structured audit: sleep, alcohol, overtraining, inflammation, medications, recent illness, weight change, and lab variability.

How to choose a biological age test

The market is noisy. A good test is not the one with the most dramatic dashboard. It is the one that gives you a validated measure, transparent sample handling, and a report you can act on without pretending it is a medical diagnosis.

1. Check which clock is included

Look for the exact algorithm name. “Biological age” is not enough. For this use case, the key phrase is DunedinPACE or another explicitly rate-of-aging measure. A test that only reports methylation age may still be interesting, but it is less suited to short-cycle intervention tracking.

TruDiagnostic lists TruAge with DunedinPACE, OMICmAge, organ-system age scores, telomere-related reporting, smoking and alcohol impact scores, and other methylation-derived outputs. Its positioning is strongly aligned with longitudinal self-tracking.

Elysium Health has offered Index, a biological age product based on DNA methylation. Before buying, verify the current report contents because product packaging, included algorithms, and availability can change.

2. Prefer repeatability over novelty

The best test is usually the one you can repeat consistently. Use the same vendor, same sample type, similar time of day, and similar health context. Do not compare a saliva-based result from one company with a blood-based result from another as if they were interchangeable.

3. Look for laboratory quality and privacy terms

For U.S. buyers, CLIA certification and HIPAA language are relevant signals, but they are not a full privacy audit. Read the consent form: data retention, research use, de-identification, deletion rights, third-party sharing, and whether raw methylation data can be downloaded.

4. Avoid tests that overpromise therapy selection

A methylation clock can support health decisions. It should not be the sole basis for prescription drugs, hormone therapy, aggressive fasting, senolytics, rapamycin, metformin, peptides, or other geroprotective experiments. If a vendor frames a clock as a standalone treatment engine, be skeptical.

A practical interpretation algorithm

Use the result as a feedback tool, not a personality test.

Step 1: Establish baseline context

Before sampling, record the previous 30 to 60 days: illness, vaccines, antibiotics, major stress, weight loss, alcohol intake, sleep debt, travel, training load, injuries, medication changes, and supplement changes. A methylation score without context is easy to overread.

Step 2: Pair the clock with standard biomarkers

A high pace-of-aging score becomes more meaningful if it aligns with elevated blood pressure, ApoB, insulin resistance, inflammation, poor sleep, low fitness, or low strength. If it does not align, repeat before making expensive decisions.

Useful companion platforms include InsideTracker for blood-marker-driven recommendations and Function Health for broad lab panels. They are not substitutes for an epigenetic clock, but they can explain why a clock may be moving.

Step 3: Change one major variable at a time

If you overhaul training, diet, sleep, supplements, and medication in the same month, you may improve your health but lose interpretability. For self-experimentation, make the biggest known-risk change first: sleep apnea treatment, alcohol reduction, blood pressure control, resistance training, protein adequacy, fiber intake, visceral-fat reduction, or smoking cessation.

Step 4: Retest on a realistic interval

For lifestyle interventions, retesting too soon creates noise. A three- to six-month interval is more defensible for most people than monthly testing. Clinical trials may use different schedules, but personal health decisions need enough time for behavior and biology to stabilize.

Step 5: Act on convergence, not one number

If DunedinPACE improves, hs-CRP drops, waist circumference shrinks, sleep improves, and strength rises, the signal is stronger. If the clock worsens while everything else improves, repeat the test and review confounders before changing course.

How to use the data for diet, training, and geroprotectors

Diet

Do not use methylation results to chase exotic diets first. Start with the interventions that already map to cardiometabolic and inflammatory risk: adequate protein, high-fiber plants, minimally processed foods, stable energy balance, lower alcohol exposure, and correction of deficiencies. If a test suggests faster aging and your blood work shows insulin resistance or inflammation, the diet target is clearer.

Training

A pace-of-aging result is most useful when combined with functional measures. Track VO2 max or a field proxy, grip strength, resting heart rate trend, blood pressure, and recovery. If your score is poor and your fitness markers are weak, the first prescription is probably not a niche compound. It is progressive aerobic work, resistance training, and recovery hygiene.

Supplements and geroprotectors

This is where restraint matters. NAD+ precursors, creatine, omega-3s, vitamin D correction, glycine, taurine, metformin, rapamycin, senolytics, and GLP-1-related weight-loss strategies all live in very different evidence categories and risk profiles. A methylation clock cannot tell you that you “need” rapamycin or that a supplement is working in isolation.

Use clocks to monitor broad trajectory after a physician-reviewed intervention plan. For prescription or experimental geroprotectors, the decision should sit on clinical indication, risk, contraindications, drug interactions, and medical supervision.

Who should consider testing

A methylation pace-of-aging test makes sense if you are already willing to act on basic health data and repeat the measurement. It is especially useful for people running a structured longevity program, recovering from a period of poor health, evaluating major lifestyle changes, or participating in clinician-guided preventive care.

It is less useful if you want a single number to rank yourself, if the result will trigger anxiety, if you are not prepared to repeat the test, or if you plan to use it to justify high-risk interventions without medical oversight.

The bottom line

Third-generation epigenetic clocks are not fortune tellers. Their value is narrower and more practical: they can help turn longevity from a vague aspiration into a measured feedback loop.

The right workflow is simple: choose a test that clearly reports a rate-of-aging measure, establish context, pair it with conventional biomarkers, change the highest-impact variables first, and retest with discipline. If the methylome, blood work, fitness, and lived experience all move in the same direction, you have something more useful than a biological age score. You have evidence that your system is responding.

Until recently, building a voice AI assistant usually meant stitching together three separate systems: speech recognition (STT), a large language model (LLM), and text-to-speech synthesis (TTS). That cascade architecture worked, but it had one painful flaw: latency. A two-to-four-second pause after every user phrase makes a conversation feel less like a dialogue and more like a walkie-talkie exchange.

By mid-2026, the industry has moved toward native Speech-to-Speech (S2S) models and real-time audio APIs. Instead of converting every utterance into text, waiting for a model response, then synthesizing audio at the end, these systems process audio streams continuously. That shift makes it possible to approach human conversational timing, roughly 300-700 ms in good conditions, while preserving tone, hesitation, emotional cues, and interruptions.

This guide compares three practical options for business voice agents: OpenAI Realtime API, Hume AI with its Empathic Voice Interface, and ElevenLabs Conversational AI. The goal is not to crown one universal winner. It is to help you choose the right stack for support, sales, healthcare intake, coaching, onboarding, or interactive media.

The main contenders

1. OpenAI Realtime API

OpenAI’s real-time audio interface is built around bidirectional streaming, usually over WebSockets, and is designed for low-latency conversational agents. The key difference from older voice stacks is that the model can handle audio natively instead of treating speech as a temporary text transcript.

For business teams, the main appeal is speed plus tool use. A voice agent can listen, respond, call functions, fetch account data, update a booking, or trigger a workflow while keeping the conversation fluid. The tradeoff is integration complexity: this is a developer-first API, not a one-click voice widget.

2. Hume AI EVI

Hume AI focuses on empathic voice interaction. Its Empathic Voice Interface, or EVI, is designed to detect emotional signals in speech and adapt the response style accordingly. That makes it especially relevant for coaching, mental wellness, patient support, customer satisfaction analysis, and any workflow where tone matters as much as the literal words.

The important distinction is that Hume is not only a voice generator. It is an emotional signal system wrapped into a conversational interface. If your agent needs to notice frustration, uncertainty, sarcasm, or distress, Hume deserves a serious look.

3. ElevenLabs Conversational AI

ElevenLabs became known for high-quality voice synthesis, voice cloning, and expressive generated speech. Its Conversational AI product brings that strength into real-time agents.

For many companies, ElevenLabs is attractive because the voices sound polished and brandable. If your product needs a recognizable voice, a media character, a game NPC, or a premium concierge feel, this can matter more than shaving the last 150 ms off latency.

The real comparison: latency, emotion, and interruptions

When a business deploys a voice AI agent, the demo is not the hard part. The hard part is making the agent survive real users: people interrupt, mumble, change their mind, get annoyed, speak over background noise, and expect the system to respond naturally.

1. Latency

Human turn-taking is fast. In a normal conversation, people often begin responding within a few hundred milliseconds. Once an AI agent regularly crosses the 800 ms mark, the interaction starts to feel mechanical.

• OpenAI Realtime API: Typically the strongest option for raw responsiveness, with a practical target around 300-400 ms in optimized network conditions. It is a strong fit when low latency is the product experience.
• Hume AI EVI: Usually sits closer to 500-700 ms. Some of that extra time is the cost of deeper acoustic and emotional interpretation, which may be worthwhile in empathy-heavy use cases.
• ElevenLabs Conversational AI: Often lands around 500-800 ms, depending on the selected LLM, voice settings, and integration pattern. The upside is voice quality and production polish.

2. Emotional range and sarcasm

Imagine a customer says, “Oh, sure, this is the best support experience of my life” while clearly sounding irritated. A basic text-only system may treat that as praise. A better voice AI stack should understand the emotional contradiction.

• Hume AI: The strongest candidate for emotional understanding. Its core value is detecting affective signals from prosody, pitch, tempo, and other vocal features. In a support scenario, it is more likely to respond with empathy instead of taking sarcastic words literally.
• OpenAI Realtime API: Very natural and fast, and it can infer sarcasm from context or obvious tone. It is less specialized than Hume for dedicated emotional analytics, but it has strong general conversational intelligence.
• ElevenLabs: The generated voices are often the most cinematic and pleasant. The weak point is not voice realism; it is dynamic emotional interpretation of the customer. For a branded voice experience, it shines. For live emotion analytics, it may need additional tooling.

3. Interruption handling

Real conversations are messy. People interrupt agents, agents need to stop speaking quickly, and the system must decide whether a cough, background speech, or short backchannel is a true interruption.

• OpenAI Realtime API: Interruption handling is one of its strongest advantages. Since audio streaming and response generation are tightly coupled, the system can stop the current audio output quickly when the user starts speaking.
• Hume AI: Strong in emotionally aware turn-taking and backchanneling. It is designed to treat conversation as a live interaction, not a sequence of isolated prompts.
• ElevenLabs: Conversational AI can handle interruptions well in its managed environment, but custom API integrations may require careful client-side VAD, buffering, and turn-taking logic.

Comparison table: business-relevant tradeoffs

Parameter — OpenAI Realtime API — Hume AI EVI — ElevenLabs Conversational AI

Typical latency — 300-400 ms — 500-700 ms — 500-800 ms

Voice quality — Excellent and natural — Good, emotion-focused — Excellent, often best-in-class

Customer emotion detection — Good general inference — Deep emotional analysis — Limited unless paired with another system

Interruption handling — Excellent — Strong — Good, integration-dependent

Estimated cost per minute — ~$0.12-$0.25 — ~$0.07+ — ~$0.15-$0.30

Integration complexity — High — Medium — Low to medium

Best fit — Fast transactional agents — Empathic support and coaching — Branded voice and media experiences

These numbers should be treated as planning ranges, not procurement quotes. Voice AI pricing changes quickly, and real costs depend on audio duration, model choice, token usage, concurrency, storage, telephony fees, and whether you route calls through a platform such as Twilio or your own infrastructure.

Integration architecture: the voice engine is not the whole product

A real business agent needs more than a beautiful voice. It needs access to CRM data, policies, booking systems, knowledge bases, payment status, order history, and escalation rules.

A practical architecture often looks like this:

1. The customer speaks through a browser, mobile app, or phone call.
2. The voice engine receives the live audio stream.
3. When the user asks for an action, the agent triggers a tool call.
4. An orchestration layer validates the intent and calls internal APIs.
5. The voice engine returns the answer in the same conversation.

For the orchestration layer, LangGraph is a useful framework because it lets teams model agent workflows as stateful graphs instead of a single prompt. That matters when a voice assistant needs to check permissions, ask follow-up questions, branch into different flows, or recover from failed tool calls.

Which platform should you choose?

Choose OpenAI Realtime API if:

• You need the lowest possible conversational latency.
• Your product already uses OpenAI models and tool calling.
• The agent must handle interruptions naturally.
• You have engineers who can manage WebSockets, audio streaming, and stateful sessions.

OpenAI is the strongest default for fast, transactional agents: booking assistants, internal copilots, sales qualification, support triage, and voice workflows where responsiveness directly affects conversion.

Choose Hume AI if:

• Emotional awareness is central to the product.
• You are building coaching, therapy-adjacent support, patient intake, customer satisfaction analysis, or wellness experiences.
• You need to detect frustration or uncertainty before the user states it explicitly.
• You can accept slightly higher latency in exchange for deeper affective signals.

Hume is most compelling when voice is not just an input channel, but an emotional context layer.

Choose ElevenLabs if:

• Your brand needs a memorable voice.
• You care about voice quality, character, and tone more than absolute minimum latency.
• You want a faster path to a working conversational prototype.
• You are building media, games, education, entertainment, or premium concierge flows.

ElevenLabs is the strongest choice when the voice itself is part of the product identity.

A simple decision rule

If you are building a customer support agent that must answer quickly and execute tasks, start with OpenAI Realtime API. If you are building an emotionally sensitive assistant, evaluate Hume AI first. If your core differentiator is a polished, branded, memorable voice, start with ElevenLabs.

For many mature teams, the final architecture may combine more than one layer: OpenAI for real-time reasoning and tool use, ElevenLabs for custom voice identity, Hume for emotion analytics, and LangGraph for orchestration. The winning stack is not always the one with the best demo. It is the one whose latency, cost, emotional intelligence, and operational complexity match the job your business actually needs done.

Longevity is often sold as a promise to reach 120. Aging researchers talk about something else: lifespan, the total years lived, matters less than healthspan, the period when a person is independent, mobile and not dependent on treating chronic disease. The gap is easy to see. Someone can live to 85 and spend the last fifteen years with heart failure or dementia.

The concept itself is not new. What is new is the attempt to turn healthspan into something measurable. That is what this section of FUTURMAG is about: not “sleep more” advice, but technologies that make healthy aging testable.

Why measurement is the real problem

Healthspan has no single definition. Some define it as time before the first serious chronic disease, others as time before loss of independence, others through quality-of-life scales. That makes studies hard to compare and interventions hard to evaluate.

There is also a regulatory side. Regulators, including the FDA, do not recognize aging as an indication: a drug has to be approved for a specific disease. Trials of geroprotective drugs therefore rely on proxies such as diabetes, cardiovascular events or functional tests. The TAME trial of metformin was conceived to test that approach itself, to see whether a cluster of age-related diseases can work as an endpoint. Check its current status on the project’s official pages rather than in headlines.

Four tiers of tools

It helps to sort what the market offers by maturity.

1. Validated risk markers. Blood pressure, ApoB or non-HDL cholesterol, HbA1c, waist circumference, coronary artery calcium score. This layer is dull but the most reliable: decades of data link these numbers to outcomes.

2. Functional tests. Grip strength, walking speed, repeated chair stands, maximal oxygen uptake (VO₂max). They predict risk and independence well and cost almost nothing. We cover how watches digitize one of them in the VO₂max article .

3. Wearables. Rings and watches provide continuous data: resting heart rate, heart rate variability, sleep, activity. They are good for tracking one person’s trend and poor at absolute values; see the comparison of rings, straps and watches .

4. Omics and “aging clocks”. Epigenetic clocks, proteomic panels, metabolomics. This is the most exciting and most contested tier: the tools measure statistical patterns, not diagnoses. How to read such results is explained in our piece on third-generation epigenetic clocks .

What about interventions

In animal experiments several compounds extend lifespan; the best evidence is for rapamycin in the US National Institute on Aging’s Interventions Testing Program. In humans the picture is more modest: safety and effects on individual markers are being studied, but extension of healthy life has not been shown. More in our piece on supplements and senolytics .

The practical takeaway is that lifestyle changes remain the intervention with the strongest evidence, while technology helps you see where you stand and whether your chosen path is working.

How we rate technologies in the Human section

For every topic we state the stage: lab research, human trials, clinical use or consumer product. We separate “associated with risk” from “reduces risk”. We do not promise that a test or supplement will extend life, and we flag when a decision belongs with a doctor.

In 2018 JAMA Network Open published a Cleveland Clinic study that followed 122,007 patients who had undergone treadmill testing and compared their survival (Mandsager et al.). People with low cardiorespiratory fitness died more often than the rest, and the best, “elite” group lived longest. The authors found no upper limit beyond which the benefit disappears. The study is observational, so it cannot prove causation, but the association is so strong that VO₂max came to be seen as a key health measure.

What VO₂max is and how it is measured

VO₂max is the maximum amount of oxygen the body can use per minute per kilogram of body weight. Lab measurement requires a mask, a treadmill or cycle ergometer, and effort to exhaustion. It is accurate but expensive, unpleasant and rarely repeated.

How watches estimate it

Watches and bands do not measure gas exchange. They estimate VO₂max from the relationship between heart rate and speed (or power) during running or brisk walking, plus age, weight and sex. Garmin, Apple Watch, Polar and others display the number automatically. For people who train regularly and give the device enough outdoor data, estimates usually correlate with lab results, but notable scatter remains; watches often underestimate very fit athletes and err for people who rarely run.

So the rule is: compare yourself with yourself. A rise in the estimate over several months using the same method is an informative signal. An argument about “I have 48, my friend has 52” is not.

What matters for interpretation

• Results depend on conditions: heat, fatigue, hills and poor GPS can shift the number.
• Watches work better if heart rate is measured accurately; a chest strap is more reliable than optical sensors.
• Test in similar conditions: same route and time of day.
• If the number drops, do not panic. Compare with sleep, illness and training load over recent weeks.

How to improve it

The foundation is plenty of easy aerobic work (you can hold a conversation), plus one or two interval sessions a week, for example 4 × 4 minutes at a hard but controlled pace. These schemes are the best studied, but they do not suit everyone: people with heart disease and those who have not trained for a long time should talk to a doctor first.

When you need a real test

A lab test makes sense if you train endurance seriously, want precise training zones, or have questions about heart and lungs. For everyone else a watch gives enough information to see the trend.

How VO₂max fits into healthy aging overall is in our piece on healthspan .

Nearly every smartwatch and ring today shows a “stress level” or “readiness”. The basis is heart rate variability (HRV): tiny fluctuations in the intervals between heartbeats. The higher it is, the better, on average, the adaptability of the autonomic nervous system. But there are many assumptions between that fact and the everyday conclusion “you are stressed”.

What HRV measures

HRV reflects the balance of the sympathetic and parasympathetic systems, but it is also affected by sleep, alcohol, illness, exertion, breathing, age, temperature and time of day. Optical wrist sensors read pulse with light (photoplethysmography), and accuracy drops with movement. That is why almost all devices compute HRV at night, when you are still.

In practice this means:

• absolute numbers differ between people and devices, so comparing them is pointless;
• personal trend (your own average over 2-4 weeks) is informative: a marked drop with insomnia, illness or overtraining genuinely signals overload;
• a single “red” day is noise, not a diagnosis.

What technology does besides measuring

Biofeedback, training breathing with HRV feedback, has evidence of small-to-moderate benefit for anxiety and for lowering blood pressure, but the effect depends on regular practice, not on the device. Slow breathing at roughly six cycles per minute underlies most protocols.

There are also clinical digital methods. Regulators have already cleared some software products as treatment: the FDA, for example, has authorized cognitive behavioral therapy programs for insomnia and an adjunct app for major depressive disorder alongside antidepressants. These are not “relaxation apps” but products that went through clinical trials, and they are dispensed on indication and prescription.

What the data say about relationships

At the other end of the scale is the longest-running data source: the Harvard Study of Adult Development, which has followed participants since 1938. The main conclusion of its directors is that the quality of close relationships in midlife predicted health and well-being in old age better than cholesterol levels. These are observational data and do not prove causation, but they agree with many studies on loneliness and health risk.

The practical takeaway: a gadget can show that you are overloaded, but the solution is more often sleep, workload, a conversation with people close to you or professional help than a new sensor.

How to use devices without fooling yourself

1. Look at trends over weeks, not morning scores.
2. Cross-check against a log of sleep, caffeine and training to find the cause.
3. If a device treats anxiety with numbers rather than sensations, set it aside for a couple of weeks.
4. With persistent anxiety, insomnia or depression, see a doctor; wearable data can be brought to the appointment as a supplement.

For which devices measure sleep and recovery best, see our comparison of rings, straps and watches .

Muscle was long seen as a matter of appearance. Increasingly it is treated as an organ that determines metabolism, resilience to injury and independence in old age. And there is now a technological reason to say it louder: a new generation of weight-loss drugs has made doctors ask what exactly a person loses along with the weight.

Strength as a biomarker

A grip dynamometer is cheap, and the result predicts outcomes well. In the large PURE study across 17 countries (Leong et al., Lancet, 2015), weaker grip was linked to higher risk of death from all causes and from cardiovascular disease, and the association held after adjusting for age and other factors. That does not mean weak grip causes disease: it reflects overall condition. But as a cheap screening test it is useful.

With age, muscle mass and strength decline, which is called sarcopenia. The rate depends on activity and nutrition, and resistance training remains the most reliable way to slow it.

Why GLP-1 changed the conversation

Semaglutide and tirzepatide help many people lose a substantial share of their weight. But in clinical trials a notable share of the lost mass is not fat: estimates usually say roughly a quarter to a third, though exact figures depend on the drug, dose and measurement method. For a young person that is tolerable; for an older person with low baseline muscle it risks losing function along with weight.

Practical conclusions discussed in clinical guidance:

• combine therapy with resistance training;
• watch protein intake (individually, with a doctor);
• track not only weight but body composition and function (strength, walking speed).

Meanwhile pharma companies are testing drugs meant to protect or build muscle during weight loss, such as myostatin and activin blockers. These are early- to mid-stage trials: none can yet be called a proven substitute for training.

What to measure and with what

• Grip dynamometer: the most accessible test; compare with age norms.
• Five chair stands: leg strength with no equipment.
• Bioimpedance scales: handy for trends, but readings depend on hydration and the model; do not trust absolute numbers.
• DEXA: a more accurate body-composition method, suitable for periodic checks on a doctor’s advice.
• Smartwatches: estimate load indirectly; limited for strength work specifically.

A minimum program

For most adults, guidelines converge on two strength sessions a week covering the major muscle groups with gradually increasing load. What matters is progression: if weight and repetitions do not rise, the muscles have nothing to adapt to. People with heart or joint disease, osteoporosis, or those taking weight-loss drugs should consult a doctor first.

For links to nutrition and glucose sensors see our pieces on personalized nutrition and CGM .

If nutrition depended only on general rules, dietitians would have an easy job. In practice the same foods affect people differently, and an industry has grown around that: apps that track glucose, microbiome tests, DNA kits, AI assistants that analyze a photo of your plate. What of this is supported, and what is just nicely packaged?

What the science shows

The best-known study is PREDICT, run by King’s College London and other centers. It showed that responses to identical food vary a lot between people: glucose and lipids after a meal depended not only on the meal’s composition but also on sleep, time of day, physical activity and the microbiome, while genetics explained only a small share of the differences (Berry et al., Nature Medicine, 2020). The idea that one size fits all really is weak.

It does not follow that a personal algorithm always beats a simple rule. Studies where people received personalized glucose-based advice showed modest improvement in some participants, and large head-to-head comparisons with good baseline advice are scarce.

What actually scales

Fiber. High fiber intake is consistently linked in large reviews to lower risk of cardiovascular disease and type 2 diabetes. Most people worldwide eat less than the recommended 25-30 g a day.

Protein. Older adults and people who lift need more than young sedentary people: a range of 1.0-1.2 g per kg of body weight and above is often cited, but your individual target should come from a doctor, especially if you have kidney disease. More on the link to muscle in our piece on strength training .

Ultra-processed food. In a controlled NIH study by Kevin Hall (2019), participants on an ultra-processed diet ate roughly 500 kcal a day more than on a minimally processed one and gained weight. That does not make all such food harmful, but it points to a mechanism of overeating.

Where technology really helps

• Food recognition from photos. AI identifies a dish well but gets portions and composition wrong. Good for building a logging habit, not for precise counting.
• CGM for a limited time. Helps you see your own responses; see the piece on glucose sensors .
• Microbiome analysis. The market is ahead of science: there are no firm dietary recommendations from standard tests yet, and results depend on the lab and method.
• Genetic nutrition tests. Justified for rare conditions (lactose intolerance, celiac disease), not for “choosing the perfect diet”.

A practical minimum

1. Fiber first (vegetables, legumes, whole grains) and enough protein, then everything else.
2. If you are interested in personalization, choose a service that explains its method and publishes research, not one that only shows attractive scores.
3. Judge results not by weekly weight but by markers over several months: blood pressure, HbA1c, lipids, how you feel.

Not long ago continuous glucose monitoring (CGM) was a tool for people with diabetes. In 2024 over-the-counter versions appeared in the US: Dexcom Stelo and Abbott Lingo are sold to people who want to track glucose without a diagnosis. A coin-sized sensor sticks to the arm and sends a value to an app every few minutes. The question is what a healthy person should do with that curve.

What the sensor actually measures

A CGM reads glucose in interstitial fluid, not blood, so readings lag lab values by minutes and can differ noticeably. It does not replace a blood test and cannot diagnose. What it shows is what a single fasting test cannot: how fast glucose rises after a meal, how long it stays up, and what happens overnight.

Why it is interesting

Responses to the same food vary widely. In the PREDICT study of about a thousand people, individual glucose responses to identical test meals differed markedly, even between twins (Berry et al., Nature Medicine, 2020). That helps explain why “universal” diets do not work for everyone, and it is the basis for commercial personalized-nutrition apps.

In practice a sensor can show:

• which of your regular meals cause sharp spikes in you specifically;
• the effect of meal order, a walk after lunch, poor sleep;
• unusual overnight spikes and dips worth discussing with a doctor.

Where usefulness ends

For healthy people there are no convincing trials showing that watching the glucose curve improves outcomes such as cardiovascular events, diabetes or lifespan. Short post-meal swings are normal: a peak is not a disease. The danger is that people start fearing normal things and restricting healthy food because of a red arrow.

Rely on solid markers: HbA1c, fasting glucose, waist circumference, blood pressure, lipids. CGM is an add-on that makes sense for a short period (two to four weeks) to learn your responses, not something to wear permanently.

What GLP-1 drugs change

GLP-1 receptor agonists (semaglutide, tirzepatide) have become the main tool for treating obesity and type 2 diabetes in recent years and have shown fewer cardiovascular events in large trials in specific groups. They change the very meaning of metabolic control: many patients reach normal numbers without strict diets. But these are drugs with side effects and indications, chosen by a doctor. Why muscle mass matters more alongside them is covered in our piece on strength training .

How to use the technology sensibly

1. If you have risk factors (family history of diabetes, excess weight, prediabetes), get an HbA1c first and discuss the result with a doctor.
2. Wear the sensor for a limited time with a concrete question, for example how you react to breakfast.
3. Do not compare your numbers with screenshots from social media.
4. Change one thing at a time and repeat the measurement.

Rapamycin, the dasatinib and quercetin combination, and NAD+ precursors are the three names that come up most in talk about “anti-aging pills”. All three have real science behind them, and in all three cases the science is more modest than the retellings suggest. Let us go stage by stage to see where the data ends.

How to read “extended lifespan”

Most of the strong results come from laboratory mice. The best test bed is the National Institute on Aging’s Interventions Testing Program (ITP): experiments run in parallel at several sites on genetically diverse animals, which reduces the chance of one-off results. In it, rapamycin extended mouse lifespan in both males and females.

Even so, it does not mean humans get the same effect. Mice have a different lifespan, metabolism and causes of death. Everything below should be read as “the hypothesis is being tested”, not “the compound works”.

Rapamycin (sirolimus)

Rapamycin blocks the mTOR signalling pathway, which links a cell’s nutrient supply to growth. It has long been used as an immunosuppressant after transplants, so its high-dose safety profile is well known: mouth ulcers, changes in lipids and glucose, and a higher risk of infection among the side effects.

For a healthy person the question is low, intermittent dosing. There are small safety and biomarker studies, but no large trials asking whether people stay healthy longer. Any off-label use is a decision for a physician, not an experiment from an internet protocol.

Senolytics

The idea: with age, “senescent” cells that have stopped dividing accumulate and release pro-inflammatory signals. If they are removed selectively, tissue might work better. In mice, clearing such cells improved several measures.

In people the data is early: small pilot studies of dasatinib plus quercetin in conditions where senescent cells are suspected to play a role, such as diabetic kidney disease and pulmonary fibrosis. These are safety and mechanism studies, not proof of benefit for a healthy person. Dasatinib is a cancer drug with serious side effects and does not exist as a supplement.

NAD+ precursors (NMN, nicotinamide riboside)

Levels of NAD+, a molecule involved in cellular energy metabolism, decline in tissues with age. In human studies, precursor supplements do raise NAD+ in blood. The open question is whether that changes muscle, vascular or metabolic function. Results so far are mixed and mostly come from small, short trials.

Product quality also needs checking. The supplement market is poorly standardized, and the legal status of NMN differs by country and has changed over time, so check the current rules with your national regulator.

What follows from this

• Mouse data is a reason for trials, not for self-treatment.
• If someone offers a rapamycin or senolytic “protocol” outside a clinic, ask about the indication, monitoring and who is responsible.
• The only interventions with demonstrated effects on healthy life are physical activity, blood pressure and lipid control, not smoking and sleep. How to measure the result is a separate question, and there technology already helps.

Not long ago, confirming Alzheimer’s disease during life took expensive PET imaging or a lumbar puncture. In recent years the picture has changed quickly: blood tests for specific proteins have appeared, along with drugs that slow the disease at an early stage. This is a technological shift that is important to understand and not to overrate.

What changed in diagnosis

The key new marker is phosphorylated tau, p-tau217, in plasma. In studies it correlates well with amyloid plaques in the brain on PET. In 2025 the FDA for the first time cleared for clinical use a plasma test based on the p-tau217 to amyloid-beta 42 ratio (Lumipulse) to help diagnose adults with symptoms. Note the wording: it is about confirming a diagnosis in a person with complaints, not mass screening of healthy people.

Reasons for caution: a positive result does not guarantee symptoms will appear or when; in some people the marker rises with kidney disease; and information about future risk without a means of intervention can do psychological harm.

What changed in treatment

Lecanemab (Leqembi) and donanemab (Kisunla), antibodies against amyloid, have been approved in the US and slowed cognitive decline in trials in people with early-stage disease. The effect is moderate, and treatment requires regular infusions and MRI monitoring because of the risk of brain swelling and microbleeds (ARIA). It is not a cure for dementia, but it is the first time drugs act on the biology of the disease. Availability and cost differ between countries.

What can be done now

The Lancet Commission on dementia, in its 2024 update, identified 14 modifiable risk factors and estimated that about 45% of dementia cases could in principle be prevented or delayed. The list: hypertension, hearing loss, high LDL cholesterol, smoking, diabetes, physical inactivity, obesity, excessive alcohol use, head injury, air pollution, social isolation, vision loss, depression and less education.

The 45% estimate is modelled, not a promise: it is the share that could be removed if the factors disappeared across the whole population. But the practical implications are clear: control blood pressure, check hearing and vision, stay active, do not smoke, keep up social contact. Hearing aids and hypertension treatment are among the best-supported measures.

What about “brain training” and supplements

Cognitive training apps improve performance on the trained tasks, but transfer to everyday memory and protection against dementia is weakly supported. Large supplements such as omega-3 and vitamins have not shown convincing prevention in people without deficiency. Multi-component lifestyle programs, such as the Finnish FINGER trial, gave a moderate benefit on cognitive tests but required a lot of effort and support.

What to do with this knowledge

1. Do not buy commercial “Alzheimer’s risk panels” without a doctor’s order.
2. If you notice persistent memory problems, early consultation gives more options than waiting.
3. For everyone else, “brain health” means blood pressure, sleep, hearing, physical activity and social life.

For how technology fits into the broader picture of healthy aging, see our piece on healthspan .

The classic checkup is blood pressure, a blood count, cholesterol and glucose. Next to it a separate industry of “deep diagnostics” has grown: extended lipid panels, whole-body MRI, blood tests for dozens of cancers from one tube, genetic reports. Some of this has genuinely improved prevention, and some is still more product than medicine. Let us separate them.

What to know first

ApoB and Lp(a). LDL cholesterol measures the cholesterol carried in particles, while ApoB counts the atherogenic particles themselves, so for some people it reflects risk better. Lipoprotein(a) is largely genetic and changes little over a lifetime; a European consensus statement recommends measuring it at least once in adults. These tests are inexpensive and can change a treatment decision.

HbA1c and blood pressure. Dull, but these catch prediabetes and hypertension early. A home cuff validated against a clinic reading gives more information than a single measurement a year.

Coronary artery calcium score. A non-contrast CT shows calcified plaque in the heart’s arteries. The result helps a doctor decide whether a person at intermediate risk should take a statin. It does not replace a discussion of risk factors and is not for everyone.

What is debated

Whole-body MRI in healthy people. The method finds many incidental findings, lesions that pose no threat but trigger further tests. Professional societies do not recommend it as screening for people without symptoms; benefit has been shown only in certain high-risk groups, such as hereditary syndromes.

Multi-cancer early detection (MCED) blood tests. Tests that read DNA shed by tumors into the blood are one of the most visible technology trends in oncology. But large randomized trials have not yet shown that such screening lowers mortality, and sensitivity for early stages is limited for some methods. Treat them as an addition to proven screening (colonoscopy, mammography, HPV testing), not a replacement.

Genetic reports on “aging risks”. Polygenic risk scores are advancing quickly, but their accuracy depends on ancestry, and the step from score to recommendation is often weaker than the report implies.

How to build a plan

1. Start with a physician and baseline labs, not with a service package.
2. For every test ask: what decision will change depending on the result?
3. Keep results in one place and compare trends rather than single numbers; wearables help here if you remember their limits.
4. From “biological clocks” choose what you can repeat yearly and compare with lifestyle changes; see our guide to epigenetic clocks .
5. If a finding is unexpected, do not be afraid to repeat the test and discuss the interpretation with a doctor.

A good checkup does not increase the number of tests. It lowers the chance of a surprise in ten years.

Telegram stopped being only a messenger long ago. It now has digital goods, mini apps, paid channel subscriptions and an internal currency, Telegram Stars. Alongside it the TON blockchain, tied to the ecosystem, is developing. All of this makes purchases inside Telegram possible, but working out who pays whom is not easy.

What is in the ecosystem

• Stars are an internal currency used to pay for digital goods and services in bots and mini apps. Developers receive Stars and withdraw them through mechanisms Telegram provides.
• Telegram Premium is a paid subscription with extended features.
• TON is a blockchain and its coin. Wallets in Telegram can hold and transfer TON and stablecoins.
• Fragment is a marketplace linked to Telegram where usernames and numbers are sold and where you can buy Stars and Premium for TON.

Purchases through the official app stores and Telegram itself are the main route. Third-party services appear where a user wants to pay another way.

Why third-party gateways exist

A gateway such as Spend.tg acts as an intermediary: you pay it in a convenient way (by the service’s description, TON, USDT, a bank card or a ruble transfer where those routes are available), and it completes the purchase of the needed service for the recipient: Stars, Premium, a TON top-up or an ad account.

That solves practical problems:

• you pay in a currency that is easier to obtain;
• no need to deal with networks and addresses;
• the interface stays inside Telegram.

But the price of convenience is an extra counterparty.

Intermediary risks

1. Trust. Money passes through a third-party service. Its reputation, time in operation and review history matter more than a pretty interface. It is not Telegram’s official wallet.
2. Irreversibility. A crypto payment cannot be undone. A wrong address or asset means losing the amount, and there may be no “refund” button.
3. Fees and rate. The final cost includes the gateway’s markup and conversion rate. Compare with the direct price of Stars in the app or on Fragment.
4. Fake bots. Scammers create bots with similar names. Open a service from a verified link from an official source and check the name and account verification.
5. Permissions. A mini app may request wallet access: confirm only what you understand.

How to vet a service

• Make the first purchase for the minimum amount and confirm the service arrives correctly.
• Compare the final price with the direct option.
• Look for independent reviews and time in operation.
• Do not send more than you can afford to lose during the check.
• For large or business payments use proven methods with a contract and closing documents.

What comes next

The trend is clear: more payments inside messengers. The key question is who is responsible when something goes wrong: official stores have refund rules, while third-party gateways are governed by their own terms. So convenience and safety in this environment have to be judged together.

DeFi trading terminals such as Axiom Trade Pro differ from ordinary websites in that signing in involves money: you either create an account or connect a wallet, and often both. The method you choose determines where keys are held and what you sign when you press a button.

Three sign-in scenarios

Email. The familiar way: you receive a code by email. If the service creates an embedded wallet for you, keys or part of them are managed on the service’s side (depending on the technology it uses). Security then equals the security of your email and of the service itself. Check whether you can export the key and how.

Google account. OAuth sign-in is faster but makes your Google account the key to the trading account. Google’s two-factor protection becomes critical.

External wallet (for example Phantom). You confirm ownership of an address with a digital signature. A login signature does not by itself move money, but at later steps the service may request transaction signatures. Keys stay with you, but responsibility for every confirmation is yours too.

What exactly you are signing

The main trap is not distinguishing signing a message from signing a transaction. The first proves you own the address. The second can move assets or give a service permission to use them. The wallet shows details, but phishing pages count on you not reading them. If a request contains permission for “unlimited” use of tokens or an unexpected transaction, decline it.

Phishing in trading terminals

Fake pages copy the interface almost pixel for pixel. The difference is often one letter of the domain, a subdomain or the top-level zone. Because trading is fast, users get used to clicking “confirm” quickly. Practical measures:

1. Save the official address as a bookmark and sign in only through it. Check links from ads and messages with particular care.
2. Never enter a seed phrase anywhere except your own wallet. No support team asks for it.
3. Use a separate wallet with a limited amount for trading, and keep main assets apart, preferably on a hardware wallet. How to choose one is in our guide to cold wallets .
4. Periodically revoke unneeded token permissions using blockchain explorer tools.
5. Turn on two-factor authentication wherever it is available.

If sign-in does not work

Check spam and the address you typed, update the wallet extension, choose the right network and restart the browser. If the problem persists, contact support through the official site, not through links from chats. A private key and seed phrase are never needed by support under any circumstances.

Bottom line

When choosing a way to sign in to Axiom Trade Pro or any other DeFi terminal, ask yourself: who controls the keys, what exactly am I signing, and what happens if my account is compromised. Trading is the second question. The first is protecting access.

There are too many AI models now: one writes best, another draws, a third makes video, a fourth picks music, and each has its own account and its own bill. That is why AI aggregators appeared, services that collect different models in one interface. Syntx AI is one of them. By its public description, it combines dozens of models and tools for text, images, video and audio and is available through a Telegram bot and a web platform.

What follows is not a promotional retelling but a look at what exactly you get and what you pay with.

How an aggregator works

Technically, an aggregator is an intermediary between you and the APIs of several model developers. You send a request in one interface, it routes the request to the chosen model, and billing goes in internal units (tokens or credits) that the service converts into the provider’s cost plus its own margin.

Three consequences follow.

1. The model lineup is not under your control. The service decides which versions to connect and when to update. A new model in the official app may arrive at an aggregator with a delay, or not at all.
2. Price depends on conversion. Compare not the “monthly plan” but the cost of a specific task: minutes of video, number of images, text length. It differs by model.
3. Features may be trimmed. In a model’s official interface there may be extra tools (region editing, references, advanced parameters) that an aggregator lacks.

Where it pays off

• Comparing models on one task. The most visible plus: give the same prompt to several models and pick the best result without opening six accounts.
• Irregular use. If you make video or music once a month, paying for four separate subscriptions makes little sense.
• Payment and access. For users in Russia it matters that the service, according to its public materials, accepts Russian cards. Terms and availability change, so check the current state on the site.
• Telegram interface. For quick tasks from a phone it is more convenient than separate apps.

Where a separate subscription is better

• You work with one tool daily (say only a text model or only an image generator) and need all its features.
• You need a stable model version and predictable output in a large project.
• You have data and privacy requirements: with an intermediary, your prompts pass through an additional party.
• You need team features and invoices for accounting.

What to check before paying

1. Which models and versions are available right now.
2. How tokens are charged and what your typical task costs.
3. What the terms say about commercial use of outputs.
4. How the service handles data and whether it stores request history.
5. How fast support responds and whether unused balance can be refunded.

A practical way is to pick two or three typical tasks, run them on a small balance and compare cost and quality with direct access to the models. If the difference in cost and convenience is not obvious, stay with separate services.

How to structure the work

An effective process does not depend on the platform: idea and text draft in a text model, two or three draft images, pick a style, and only then expensive video generation. That order saves balance on any service. For how to compare other aggregators, see our review of SYNTX.AI, AiHere, Outsee, Umnik.AI and alternatives .

Bottom line

Syntx AI is a rational option for people who use different generative models irregularly and want to compare them without many accounts. For constant work with a single tool, or with strict privacy and model-version requirements, look at direct access.

The phrase “magic words for a neural network” lasts until the first serious task. Then it turns out the result depends not on incantations but on clarity: what is needed, for whom, in what form, and how you will know it turned out well. That is closer to a technical brief than a spell.

Five parts of a request

1. Task. What exactly to do.
2. Context. For whom, why, which facts are known.
3. Format. What the answer should look like: list, table, JSON, length.
4. Constraints. What to avoid, the tone, banned words.
5. Quality criteria. How you will judge the result.

A weak request:

Write a post about neural networks.

A stronger one:

Task: write a Telegram channel post on how small businesses use AI. Context: audience is cafe and shop owners with no technical background. Format: up to 900 characters, a hook in the first line, one concrete example. Constraints: no "revolution" or "the future is here", no emoji. Criterion: the example must be doable in one evening.

What works for text models

• Examples. One or two samples of the desired result beat a long description of style.
• Structured output. If the answer goes into a program, ask for JSON with a defined schema and validate it with code, not by eye.
• Step breakdown. For complex tasks ask first for a plan or a list of assumptions, then the result itself.
• Self-checking. Ask separately for a list of claims that need source verification.
• A role (“you are an experienced editor”) helps less than concrete criteria.

Image models

An image is described like a shot: subject, setting, light, angle, style, color palette. For reproducibility keep part of the style description fixed in every request. State what must not appear (“no text, no logos”). English sometimes gives more predictable results on visual models, but modern models understand other languages well too: test on your model.

Video

Video models are sensitive to motion. Describe the action and the camera separately: what happens, how the camera moves, how many seconds. One shot, one action. If you need a stable character, start from a reference image rather than text.

Music

State genre, tempo, mood, instruments and structure (“8-bar intro, verse, no vocals”). Mind the rights and terms of use for generated tracks on your platform.

Iterate instead of aiming for the perfect prompt

Do not try to write the perfect prompt first time. Run it, see what exactly is wrong, and change one thing at a time. Keep good prompts in a file: a prompt library pays off quickly.

How to check the result

For important tasks, build a small set of test examples and run every new prompt version against it. It is the same principle as in software: a change should improve things, not just look better on the last example.

Telegram channels are among the most visible areas where AI has entered daily work: post drafts, headlines, covers, short videos for stories. But the word “automation” misleads. A channel lives on the author’s voice and a reputation for accuracy, and those are exactly the two things models destroy first when left unchecked.

What models do well

• Ideas and series. Ask for 20 topics based on your past posts with the best metrics and you get a starting base for a content plan. A human chooses.
• Drafts and structure. A post skeleton from your thesis: hook, 2-3 points, conclusion.
• Editing. Shorten, change tone, find repetition. Safer than generating from scratch.
• Visuals. Covers and illustrations without licensing complications. Better to ask for an image without text and add captions separately.
• Repackaging. From one long piece, make a post, a set of takeaways and a short video script.

Where models let you down

• Facts. A language model can confidently name a nonexistent date, figure or quote. Check any claim with names, numbers and links against the source.
• Consistent voice. Without examples of your writing the model writes “averaged”. Readers cannot say why, but subscriptions and forwards fall.
• Freshness. A model without search access does not know the news. For news channels that is critical.
• Sameness. If everyone uses similar prompts, posts on the same topic become indistinguishable.

A workflow that works

1. Build a “voice passport”: 5-10 of your best posts, a list of favorite phrasings and banned words. Paste it into every prompt.
2. Give the model a thesis and source facts rather than asking it to “write about”. The source should be yours.
3. When you get a draft, rewrite the beginning and end yourself: that is where the author is heard.
4. Check facts and links. Mark where you rely on a source.
5. For covers use one style: the same colors and composition in every prompt.
6. Publish on a schedule, not in a burst. Scheduled sending is built into Telegram itself.

Automation through bots and API

If you want a pipeline (say, a bot that prepares drafts and sends them to you for approval), it is done with the Bot API and a model connected over API. One working principle: a human approves every publication. Fully automatic posting without review quickly leads to errors and, for some topics, to complaints against the channel.

To try models on one task, aggregators such as AiHere are handy, letting you compare texts and covers without separate subscriptions.

How to measure the effect

Look not at the number of posts but at reach, read rate, forwards and unsubscribes. If metrics drop after you adopt AI, the voice has been lost: put more of your own editing back.

Testing five visual ideas for an ad used to depend on a designer and a budget. Now an image or a short video takes minutes, and the number of variants has jumped. But more variants does not mean better results: test ten random pictures and you learn which won, not why.

Hypothesis first, generation second

A workable scheme is built on a matrix: what exactly you change and what stays constant.

Variable — Examples

Scene — product close-up / product in use / outcome

Emotion — calm / urgency / humor

Offer — discount / free delivery / guarantee

Format — square / vertical 9:16 / wide

From the matrix pick 3-5 combinations where one variable changes, and generate visuals only for those. Then the A/B test answers a question instead of just naming a winner.

What models can do today

• Images. Good at backgrounds, atmosphere and object scenes. The weak spots are text in the image (especially Cyrillic), hands, small packaging details and logos. A practical rule: ask for an image without lettering and add text in an editor.
• Video. Short clips from text or an image have become usable for testing, but long scenes and keeping one character consistent between shots remain unstable. For product ads it is often better to animate a finished photo.
• Brand consistency. To keep visuals from drifting apart in style, use references (palette, composition examples), a fixed style description in every prompt and a shared negative list (“no text, no logos”).

Ad platforms built in AI too

The ad dashboards of large platforms now offer generation of backgrounds and text variants themselves. That is convenient, but check which rights and limitations apply in each dashboard and compare with the output of external models. For small teams a third-party tool such as AiHere is useful when you want to compare several models on one idea.

What to check before launch

1. Rights and brand. Whether the image reproduces recognizable people, other companies’ logos or a specific artist’s style.
2. Labeling. Regulation is moving toward mandatory disclosure of synthetic content: the EU AI Act introduces such duties, and platforms have added labels of their own. Rules depend on country and channel, so check current requirements before launch.
3. Truthfulness. You cannot show in an ad a result the product does not deliver; a generated “before/after” easily becomes a violation.

How to read results

A test must collect enough impressions and conversions; an early leader is often noise. Compare on a metric chosen in advance (cost per target action, not clicks). Record which hypothesis won and use that knowledge in the next matrix.

AI speeds up producing variants but does not cancel analytics. The winner is whoever learns faster from tests, not whoever has more pictures.