Nobody stole the keys. That's what makes this one different. Bitget lost $387.5 million on Thursday making it the largest crypto theft of 2026. Attackers got into a backend system in its wallet infrastructure & fed it forged transaction data. Bitget's own authorization process then approved the transfers as routine. Nineteen transfers across five blockchains, including 103 million XRP worth about $157 million. Private keys were never compromised. Cold wallets were never touched. Every security control people usually argue about held & the money left anyway, because the system was persuaded rather than broken. Attribution points at North Korea. CEO Gracy Chen says IP addresses match VPNs used by a DPRK-linked group & the on-chain behaviour matches past operations. Elliptic calls it highly likely & notes the stolen funds touched addresses previously used to launder last year's $1.4 billion Bybit theft. Bitget says its User Protection Fund covers the whole loss. Worth knowing what that fund is: roughly 5,500 bitcoin. So the insurance is denominated in the same asset class it's insuring & its dollar value moves with the market it's meant to protect users from. Withdrawals are paused. Deposits & trading stayed open. There's a bounty paying 5% for freezing attacker funds, 5% for recovery & some blockchain foundations have already frozen addresses. Custody debates usually stop at hot wallet versus cold wallet. This one says the vault was fine. The paperwork wasn't.

Claude just computed a nine-loop scattering amplitude in planar N=4 super-Yang-Mills, past the eight-loop record. What that means. Scattering amplitudes predict how particles behave & physicists compute them in layers of correction called loops. Each loop makes the answer more precise & costs exponentially, sometimes factorially, it’s more work. Most real amplitudes have only been taken to two or three loops. The most precise prediction in particle physics, the electron's anomalous magnetic moment, used five. The challenge came from physicist Matt von Hippel, who publicly asked whether an AI could push past eight loops using only compute an academic could afford. Given one prompt & periodic instructions to continue, Claude ran largely unsupervised for days & did it, using bootstrap methods that Lance Dixon & collaborators developed. Total cost: a few thousand dollars. Dixon, who held the prior record, checked the result independently. Three caveats that matter more than the headline: 1. N=4 super-Yang-Mills is a toy model which is a testing ground for techniques, not a theory describing our universe. No new real-world particle prediction came out of this. 2. Claude applied existing human methods rather than inventing new ones. 3. A human-led group at the Chinese Academy of Sciences, using GPT-6 for parts of it, had concurrently obtained most of the same result & published a dataset on September 17. So the honest version isn't that a machine beats humans. It's that a hard calculation at the edge of a field now costs only a few thousand dollars & takes several days, by more than one route at once.

Structures for the protein complexes of more than 2,800 viruses went into the AlphaFold Database this week, free for anyone, academic or commercial. The scale: roughly 1.7 to 1.8 million high-confidence complex predictions across about 2,812 viral proteomes. Around 30% of the protein interactions in it are new to science, not in the Protein Data Bank at all. Why complexes rather than single proteins matters. Joe Grove at Glasgow, part of the effort, put it simply: many viral proteins don't act alone, they work with partners. Viruses had been a blind spot in a database that already covers nearly every catalogued protein & has over three million users. The engineering: AlphaFold2 running on NVIDIA's BioNeMo Inference Runtime, which brought prediction down to minutes per structure at proteome scale. X-ray crystallography takes years & thousands of dollars per structure. The pipeline itself has been open-sourced. Target selection was guided by the UK Health Security Agency's priority pathogen tool, so it's weighted toward viral families known to infect humans. The necessary caveat & Nature ran it prominently: these are predictions, not experiments. They lack the sugar molecules coating many viral proteins - the ones that help viruses evade immune detection, & they'll need experimental confirmation before anyone builds on them. The timing isn't accidental either. It landed alongside a UN General Assembly session on pandemic preparedness. Still, the point stands. The useful version of preparedness is having the structural work already done before you need it, rather than starting the week a new virus arrives.

Anthropic says Claude found a previously undescribed enzyme system in bacteriophage DNA. One research brief, a database of 1.9 billion protein clusters, roughly 950 agent sessions running 21 hours & consuming about 210 million tokens. The system is called ART: array-associated reverse transcriptase & has three parts: the enzyme, a partner gene of unknown function, & an array of 3 to 21 evenly spaced DNA repeats. The repeat layout is what drew the CRISPR comparison. In CRISPR, an array like that stores RNA guides, which is what makes the system programmable. Now the caveats, which matter more than the headline. The enzyme itself wasn't actually new, earlier studies had identified it. What's new is the recognition of the array & the partner protein alongside it. Nobody knows what ART does. Lab work confirmed the array is expressed as several short RNAs & that's where the biology currently stops. No demonstrated function, no gene-editing capability, preprint not peer-reviewed. Feng Zhang, one of CRISPR's pioneers, reviewed it & called the finding intriguing & worth further investigation, which is a careful scientist's phrasing, not an endorsement. Other researchers have been less impressed. 🤷‍♂️

Nineteen humanoid robots waltzed with human partners in Shanghai this week. The interesting problem isn't the dancing. It's the partner. 🤖 Choreography alone is solvable, a robot executing a fixed sequence on a flat stage is a well-understood control problem. But adding a human changes the category. A partner applies unpredictable forces through the arms and torso, arrives fractionally early or late & shifts weight in ways no script anticipates. To the robot, that's a continuous stream of disturbances it has to reject while staying upright, in sync with music & in position on a crowded stage. Balance under unmodelled contact is the same problem that stops humanoids working in warehouses and homes. Objects shift, floors slip, people bump into things. A stage with a human partner is a rare public test of exactly that. The hardware: Unitree H2, 180 cm, roughly 70 kg, 31 degrees of freedom, leg joints rated to 360 newton-metres, onboard compute scaling to 2,070 TOPS on Nvidia's Jetson AGX Thor. The eyes are cameras. Unitree says the routines ran autonomously in real time. That means no teleoperation & no remote correction, not improvisation. The choreography was learned in advance; what's live is the balance & synchronization. Nineteen machines holding that simultaneously, in front of an audience, with no single failure visible, is the real measurement worth noting here. 📐

What kind of useless gadget is this.

Meta announced Muse Charm: a keychain-sized device with a 2-inch touchscreen, a camera, at least three microphones & a fingerprint sensor to wake it. Ships December. No price. The architecture is the interesting part. Meta reportedly gave it built-in 5G, which means it doesn't need a paired phone & on something this size there's no room for meaningful onboard compute. So the device captures, the network carries & the model runs in Meta's cloud. Sensor at one end, intelligence at the other. That's the same admission as the VR Glasses at the same event, where the processor sits in a separate puck. Both say the limit on wearables isn't chips. It's thermal load & battery mass on the body. Zuckerberg's own framing: if you're not wearing glasses, this is the fastest way to show Muse what's going on around you. Which is a camera, on a keyring, in whatever room you're in. The specs that actually matter haven't been published like battery life, whether capture is continuous or on demand & what signals to people nearby that it's recording. Meta says the layout isn't even finalized. Humane's Pin ran hot & died. Rabbit's R1 underdelivered. This one offloads both the heat & the thinking, which fixes the hardware problem but also creates a different one.

Meta just announced VR Glasses today at $1,299.99 & the engineering story is a single design decision: take everything heavy off the face. 100 grams on the head which is five times lighter than Quest 3. The processor, battery & storage sit in a roughly 300-gram puck in your pocket, connected by an optical tether. Magnesium alloy frame, dual micro-OLED panels, Qualcomm's new Snapdragon Reality Elite chip. Why this matters more than the spec sheet suggests: weight on the face is the constraint that has capped VR session length since the beginning. Head-borne mass loads the cervical spine at a long lever arm, and the discomfort compounds with time. Halve the weight and you don't get a slightly nicer headset, you change the viable use case from twenty minutes to a feature film. The trade is optical. Field of view drops to 70 degrees horizontal, against roughly 110 on a Quest 3. Pancake lenses with micro-OLED buy you thinness & pixel density - 37 pixels per degree, 5K - at the cost of how much of your vision the image fills. Lighter face, narrower window. Everything else follows from that: eye-tracked foveated rendering to make the resolution affordable, eye tracking doing double duty as UI input and biometric unlock, hand tracking with no controllers included. Reality Labs has burned around $60 billion since 2019. This is the first device where the physics of wearing it stopped being the main objection.

Skild AI says it trained a robot to play football by letting it play against itself for 140 years inside a simulation. ⚽️ The clip is labeled autonomous, 1x. No pilot, no speed-up. Two things are doing the work here. First, time compression which is thousands of robot instances running in parallel physics sims, so what would be a century & a half of practice takes days of wall-clock training. Skild builds these on NVIDIA's Isaac Lab & Omniverse & describes gaining millennia of experience in days. Second is self-play. Nobody scripted a dribble or a shot. The system plays opponents that are copies of itself, & because the opponent improves whenever it does, the difficulty curve is generated automatically - never so easy the policy gets lazy, never so hard it learns nothing. It's the same mechanism that produced superhuman Go. The hard part is the last step. A policy trained in simulated physics usually falls apart in a real body, where friction, motor lag & a slightly wrong mass estimate break everything. Getting it to transfer is the actual result & Skild has shown related generalization before with its models recovering from a stuck wheel in seconds, or walking with a broken leg after a few attempts. DeepMind did knee-height soccer robots this way in 2024. This is the full-size version.

Predator is a steel humanoid that got knocked to the floor repeatedly this week & stood back up unassisted every time. Fall recovery is harder than walking, & it's the part worth watching. Upright locomotion is a balance problem with a known starting state. A fall erases that. Predator lands in an arbitrary orientation, has to reconstruct its position from its own sensors, then sequence dozens of joints through a push-up while its center of mass moves the entire time. One error & it's back down. It did this under adversarial conditions, against Frankie LaPenna trying to prevent it. To be clear about what's being measured: REK's founder told NBC News these robots aren't autonomous. They are "100 percent" human pilots in VR headsets. So Predator is testing teleoperation latency, impact tolerance & balance recovery, not machine decision-making. Which makes it a more honest benchmark than most. Lab demos happen on flat floors under controlled conditions. You can't stage a knockdown from someone actively fighting you. Getting up is the gate on humanoids working anywhere near people. Predator cleared it in public, repeatedly, on video. 🦾

Pioneer Labs announced sPL.001 today which is a microbe engineered to grow on Martian dirt & turn it into building material. The organism is Cupriavidus necator, a bacterium that naturally stores carbon as PHB, a bioplastic. The recipe is genuinely elegant: 40 grams of Martian regolith per liter of water supplies nitrogen & phosphorus, while the carbon comes from acetate made from atmospheric CO2 using a small electrocatalysis device. Dirt, water & air in, printable plastic out. The reason regolith is hard: it's heavily salted & contains 0.5 to 4% perchlorate, a bleach-like compound that kills most lab strains. Getting a chassis organism to grow in that at all is the achievement here. The idea isn't new, which is a point in its favor. A 2015 study estimated that C. necator producing PHB on Mars could cut the shipped mass needed to 3D-print a six-person habitat by 85%. Pioneer Labs is trying to actually build the organism. Now the caveats. This is a company announcement with no paper and no peer review yet. The claim that one rocket's equipment yields a small city's worth of material in four years is a model, not a demonstration. Despite the terraforming language, the microbe lives in sealed heated bioreactors - their own roadmap describes indoor biomanufacturing as a stepping stone toward an outdoor organism, which is a much harder problem. Still, the framing is right. You cannot ship a city. You can ship a bacterium and let it eat the planet.

Belleville Police have confirmed an open criminal investigation into the death of an 83-year-old woman under Canada's MAiD program. Brigitte Stegemann died by lethal injection on July 10 at The Pearl care home in Cannifton, Ontario. She had stage 4 stomach cancer, diagnosed in February, with a prognosis of six months to two years. Her granddaughter had cared for her for twelve years & held power of attorney for personal care for six. She says her grandmother had refused MAiD repeatedly on religious grounds & that the assessment was arranged while she was away on a ten-day trip. The family's complaint to Ontario's Chief Coroner, filed in August, makes three specific claims: that she lacked capacity to consent, that the power of attorney was bypassed & that there was no final express consent on the morning. On the capacity assessment, the family says she was asked how many siblings she had - she had 13 - answered none. Asked how many were living, she said none. 2 are alive. The care home & the practitioners involved have not publicly addressed the allegations. They maintain that she legally consented. Nothing has been proven & the investigation is ongoing. Here's what makes this bigger than one family. Every argument for assisted dying rests on capacity & consent being verifiable. This case will test whether Canada can actually demonstrate that after the fact or whether the paperwork is the only record that survives.

Mercedes-Benz just signed on to put Wayve's AI Driver into production cars within two years. That makes three major automakers: Nissan, Stellantis & now Mercedes. The interesting part is how Wayve drives, because it breaks with how self-driving has been built for fifteen years. The traditional approach is a stack of separate modules: one detects objects, one predicts what they'll do, one plans a path & they all rely on detailed pre-built maps of every street. It works, but every new city needs its own mapping & engineering. Wayve uses a single end-to-end neural network. Camera & sensor data go in, driving decisions come out, with no high-definition maps. It learns driving the way a language model learns language- from enormous amounts of examples. Wayve says it drove "zero-shot" in more than 500 cities across Europe, North America & Japan within a year. Zero-shot means no city-specific training before arriving. If that holds up at scale, it solves the biggest cost problem in autonomy. What consumers get first is more modest. From 2027, hands-off driving on supervised roads - the car steers, navigates and handles traffic, but you are still responsible & watching. Three automakers betting on the same approach is the real signal here. They're not buying a feature. They are betting the module stack was the wrong architecture.

Amazon says it will put more than 20,000 pairs of AI smart glasses on delivery drivers by the end of 2027. The glasses switch on when the van parks. They show which package to grab, then walking directions to the door, hazards along the way & a proof-of-delivery photo - all in the driver's line of sight instead of on a phone. Amazon says 500 drivers have already used them on 275,000 deliveries & estimates they save up to 30 minutes per shift. The science question here is “attention”. A driver looking down at a phone isn't looking at the stairs, the ice, or the dog. Putting information at eye level removes that glance. But aviation researchers found the catch decades ago with pilot heads-up displays: attentional tunneling. When information floats in front of you, your eyes stay on it & your awareness of everything else narrows. The display is in your line of sight, but so is the world & the brain doesn't always split them well. So the real test isn't whether the glasses work. It's whether they make drivers safer or just faster. Amazon reports serious vehicle crashes are down more than 23% from its broader safety upgrades, but that's the company's own figure across multiple changes, not a study of the glasses. 30 minutes a shift, multiplied across millions of daily deliveries, is the business case. Whether the driver's eyes end up on the doorstep or the display is the one worth measuring. 📐

Sleepagotchi wants to pay you to sleep better. But it’s worth separating the real part from the token. The real part is that about 78,700 people use its AI sleep coach every day, most opening it within ten minutes of waking. In behavioral medicine, a habit loop like that is rare. And the underlying need is real - roughly a third of American adults don't even get 7 hours. The platform is tied together by a crypto token & you earn rewards for hitting sleep goals. That's where the medicine gets complicated. Sleep researchers already have a name for people who sleep worse because they're anxious about their tracker's score: orthosomnia. Attach money to that score & you raise the stakes on exactly the number that's causing the anxiety. Plus, you also create a reason to game the tracker instead of actually sleeping. And the history of earn-by-doing tokens isn't kind. StepN paid people to walk & its token collapsed once new users stopped arriving to fund the old ones. There's a data question too. The CEO has said openly that the real prize is the biometric data the app generates. Worth asking who ends up owning yours. A sleep coach that works would be genuinely valuable, but it shouldn’t need a token to prove it. What it needs is a trial & a proper architecture like QuanMed AI.

The human hand has about 27 degrees of freedom. Around 21 of those are in the fingers & thumb - four in each finger, five in the thumb with the rest in the wrist. Unitree's new Dex5-S hand has 22. Real-hand size for $6,500. On paper that's roughly finger-for-finger parity with you. In practice it isn't & the gap is the interesting part. Your hand is packed with sensors - touch, pressure, temperature, stretch & the motion is only half of what makes it a hand. Robotic hands have been matching joint counts for years while lagging badly on feeling what they're holding. What Unitree got right is closer to biology than the joint count: every joint is backdrivable. Push a finger & it yields rather than locking. Human hands work the same way where tendons & muscles are compliant, which is why you can catch a falling glass without shattering it. Rigid robot fingers can't. And the price is the other story. Dexterous hands used to cost more than most of the robots they were attached to. At $6,500 they become something labs can buy in quantity, break & iterate on. That matters beyond humanoids. Prosthetics research runs on exactly this kind of hardware & cheaper, more human-like hands are how better ones get built. Twenty-two joints is the easy part. Touch is still the hard one. 🦾

Stanford just built a fake biotech company staffed entirely by AI agents & put a chief scientific officer in charge of it. 37,075 agents, organized into divisions like target identification, trial design & the usual org chart. The CSO assigns the work. Each agent got exactly one late-stage clinical trial to analyze, out of more than 55,000 trials across a huge range of conditions. The finding that came out of that pile: drugs targeting proteins that are active in specific cell types were nearly 50% more likely to reach market than other drugs. A structural predictor of trial success, pulled from data that was already public and already sitting there. Then they pointed it at a real question: is CD276 a good target for lung cancer? The system confirmed the target from existing data & designed an approach: an antibody that recognizes CD276, tethered to a cancer drug. Outside reviewers called it promising. Here's the part to keep honest about. Nothing was tested. No experiments, no trials, no validation in a lab. A very sophisticated literature-and-data analysis produced a plausible plan & plausible plans fail in pharma all the time. The lead author's own framing is that this is about how far agent teams can go, not that they've arrived. Still, 37,000 agents reading 55,000 trials in parallel is a thing no research group could do any other way.

Meta launched Muse, its AI agent, on September 8. Ten days later Zuckerberg opened it to outside developers. The pitch is simple: you bring the API, Muse brings the agent. You tell it what you want and it goes and does it - reads your Notion docs, pulls your meeting notes, books the table, pays for it through Stripe at more than a million businesses. It hit number one on the App Store. Now read Meta's own help center. There are two kinds of connectors. The ones in the directory, which Meta reviews. And custom ones Muse builds on the fly for a single user from any service's public API - which Meta does not review. So the agent that can spend your money can also wire itself into services nobody at Meta has checked. Meta's own setup guidance says it plainly: Muse may take unexpected actions. Monitor it carefully. To be fair, there are real guardrails. Requests run in isolated virtual machines, Stripe never hands Muse your full card number & it's supposed to ask before anything consequential. But "supposed to ask" is doing a lot of work in a product whose whole selling point is that you don't have to.

This week, a human boxer fought a six-foot humanoid in a cage in San Francisco. 🥊 The machine is an EngineAI T800 1.73 m, 75 kg, the same model used in China's URKL combat league. First one imported to the US, brought in specifically for sport rather than research or industry. It is not fully autonomous. REK's fighters wear VR headsets & arm-mounted controllers and pilot the robot in real time. So the bout is a human boxer against a human operator with a metal proxy. The engineering problem being solved is teleoperation latency & balance under impact, not machine decision-making. But this is exactly why it's interesting. Getting a 75 kg bipedal machine to absorb strikes & stay upright is a genuinely hard control problem & combat is an unusually honest test of it, because you can't stage a fall. Two things that make this more than spectacle. REK says it's training models on fight footage, so the piloted data becomes training data for autonomous movement later. This same hardware appearing in Chinese league matches is now landing in the US through entertainment rather than industry, which says something about where early commercial demand for humanoids actually is. Robot theater is a fair description. But also a benchmark nobody can fake.

Every living thing on Earth like bacteria, trees, you writes its instructions in the same four-letter code. A, T, C, G. Four letters, every organism, no exceptions, for billions of years. Scientists at UC San Diego just added 4 made-up letters & the enzyme that reads DNA handled them anyway. 🧬 Here's why that's strange. DNA isn't useful on its own. An enzyme has to crawl along it to copy it out & that enzyme has spent billions of years seeing only those four letters. Show it something new & it should choke. It didn't. It read the synthetic letters as accurately as the real ones. They took atomic-level pictures to find out why & the answer is that the enzyme isn't only checking for the specific chemistry of the natural pairs. A synthetic pair that fits, trips the same recognition machinery. Right shape, it copies. Never mind that nothing in nature has ever looked like this. The second paper is the stranger one. They used synthetic letters built without the chemical bonds every textbook says hold DNA pairs to their partners. The enzyme worked on those too. So life's alphabet has four letters not because four is the limit, but because that's what it started with. Turns out the machinery was always capable of more. This was purified enzyme in a test tube, not inside a living cell. That's the next problem & it's a hard one. What it's for, eventually? DNA with extra letters can be designed to do things natural DNA can't. Earlier work built synthetic DNA that latches onto liver cancer cells. This is the step that makes running that kind of thing inside a living organism plausible.