Russell Sean

Microsoft 365 went down across dozens of countries on Thursday. Nearly 130,000 reports on Downdetector. Outlook, Exchange, SharePoint, Copilot, Teams, Intune - all were gone for the better part of a day. But nobody was actually upset. 🤷‍♂️ "Microsoft Office going down is like a snow day for adults." Or "It's almost Friday, so there's really no point in Outlook coming back now. Let's circle back then." Or "Dear Microsoft, no rush — spend time with your family." A recurring suggestion was that the fix could comfortably wait until January. Karl Sharro: "Thank you Microsoft. Can you do this every Friday?" Thirty-five thousand likes. Someone posted a note addressed to the engineers working the incident: "Take a long weekend, relax & come back Monday feeling refreshed. I promise those 400 unread emails aren't going anywhere." And the best one, 23,000 likes: "Can someone at Microsoft please figure out why Teams is working & fix the issue?" That's the tell, really. People weren't annoyed that things broke. They were annoyed that not enough things broke. Also doing numbers: "Microsoft Outlook will return in Avengers Doomsday." Service began recovering into Friday. Reaction was mixed. 🙃

A police helicopter found him asleep on a flat rock beside a creek. He woke up, looked up & gave the crew a thumbs-up. He had no idea anyone was looking for him. 😢 William, 17, went missing on Friday 6 November while camping with a friend at Walhalla East, in the Victorian Alps about 180 km east of Melbourne, Australia. He walked off to collect some equipment & didn't come back. Shorts & work boots. His phone, his food & his water all still at the camp. He was found 6 days later. In between he walked for days through steep rocky country. About 2 days before the end he fell down a riverbank & couldn't get back up, so he stayed put in a dry riverbed, using towels as blankets. Cold nights, hot days, in the inspector's phrase. What found him was his tackle box. A helicopter crew spotted it from the air, roughly 2 kilometres east of Stoney Creek & that narrowed the search to the right patch of bush. The search ran 6 days: police divers, the dog squad, helicopters, SES, more than a hundred emergency personnel, with over 80 volunteers out every day. Police said the terrain was too steep to cover on foot at all. He came out with cuts to his legs, severe sunburn across his face and chest & too exhausted to walk. Taken to Latrobe Regional Hospital for monitoring. His mother Elizabeth said she never gave up hope & called it a miracle. The detail that sticks: Acting Inspector Jason Benbow said William had no idea police were even looking for him. 6 days, 100 searchers, helicopters overhead & from where he was sitting, he was just a kid who'd got lost & was waiting it out. He's reportedly grounded for life. 🙃

5 million miles with zero fatalities is not evidence of anything. 5 million miles with zero injuries is weak evidence of something. Understanding the difference is most of what you need to read autonomous vehicle safety claims. Tesla says its robotaxi fleet now runs 24/7 across eight cities & has passed 5 million autonomous miles with zero injuries, zero airbag deployments & no pedestrians or cyclists struck. Start with base rates. NHTSA's 2024 figures: 1.19 fatalities and 74 people injured per 100 million vehicle miles travelled. Scale those down to 5 million miles. You'd expect roughly 3.7 injuries & roughly 0.06 fatalities. Which makes the fatality claim empty. If these cars were exactly as dangerous as the average American driver, the chance of seeing zero deaths across 5 million miles is about 94%. You'd almost certainly see zero either way. To expect even one fatality at the human rate you need around 84 million miles. Tesla has about 6% of that. The injury claim is genuinely different. Expecting 3.7 & observing zero carries a probability near 2.5% if the fleet were merely average. That's a real signal. Not proof, but not nothing. Except the comparison population is wrong. That 74-per-hundred-million figure covers every mile Americans drive. Rural two-lane roads after dark. Highway speeds. Ice. Drunk drivers, exhausted drivers, seventeen-year-olds, cars with bald tyres and no automatic braking. Tesla's robotaxis drive inside geofenced urban areas in Texas & Florida, in new vehicles, on mapped roads, in mostly benign weather. Not the same risk population & urban low-speed miles are considerably safer per mile than the national blend. This is solvable & Waymo actually solved it. Their published safety work constructs benchmark rates specific to the areas & road types they actually operate in rather than leaning on the national average. More work, smaller headline numbers. There's a second gap. Tesla hasn't said what share of those 5 million miles were genuinely unsupervised. The Bay Area service still runs with safety drivers & Austin's unsupervised rollout has been described as limited. A safety monitor who intervenes before an incident doesn't produce an incident, so supervised & unsupervised miles aren't interchangeable units in a safety denominator. NHTSA opened an investigation in June 2025 after documented wrong-side driving, phantom braking & drop-offs in intersections. None of it produced injuries. All of it is the kind of event that produces injuries on a different roll of the dice. None of which says these cars are unsafe. They may well be meaningfully safer than human drivers & Waymo's much larger dataset suggests that's achievable. It says 5 million miles is early. Rare events need long denominators & the honest version of this announcement is that the fleet hasn't driven far enough yet for the headline to mean what it sounds like.

In orbit, collecting energy is the easy part. Getting rid of it is the problem & it's the one that decides whether any of this works. Musk just said today that SpaceX is aiming for a terawatt per year of solar from Earth, rising to over a petawatt per year from the Moon. 🌍🌕 Set that against what the company has actually announced. The Gigasat factory in Bastrop, TX targets roughly 1 GW per year of orbital compute by the end of 2027, scaling about tenfold annually after. The AI1 satellite is around 70 metres long, mostly solar array at 250 watts per square metre, with a 150 kilowatt compute payload in the middle & double-sided radiators. One gigawatt a year means launching over 6,000 of them annually. Starlink currently has about 10,500 satellites in total. So the tweet is 1000 times the 2027 target for Earth & 1,000,000 times it for the Moon. The solar case itself is real. A panel in orbit sees about 1,361 watts per square metre continuously - no atmosphere, no night, no weather. Google put the advantage at roughly eightfold over a good terrestrial site, & that figure holds. But the energy you collect has to leave again. Every watt of compute becomes a watt of heat & in vacuum there's no air to carry it off and no water to boil. The only exit is radiation & radiated power per square metre goes as the fourth power of absolute temperature. Run that at a temperature real hardware tolerates. At 30°C a good emitter sheds roughly 430 watts per square metre per side, call it 860 double-sided. A gigawatt of compute then needs on the order of 1.2 square kilometres of radiator. A terawatt needs about 1,200 square kilometres. A petawatt needs roughly 1.2 million square kilometres - around 3% of the entire surface of the Moon, in radiator alone. That T⁴ term means those numbers are extremely sensitive to temperature. Run the radiator at 127°C & the area falls by about two thirds. Which is precisely where the constraint bites, because chips don't run at 127°C. There's measured data on this now. A paper this year reports from BUPT-1, a 12U satellite at about 490 kilometres carrying AI accelerators. One accelerator at full load reached the structure's 30°C limit after roughly nine hours. Two accelerators sharing a single radiating surface reached it in about 40 minutes - the first task killed near 50 minutes, the second device powered down around 110. The authors' phrasing: co-location collapses the safe continuous-compute window from about 9 hours to under 2. On one vision-language run, device temperature climbed from 45°C to a 55°C shutdown threshold in 48 seconds. Their conclusion is that thermal headroom, not raw FLOPS, limits sustained serving in orbit. Across months of operation, overheating terminated tasks twice. Radiation never caused a single observed upset. None of which makes orbital compute impossible. SpaceX's 150 kW per satellite with dedicated double-sided radiators is a design that takes the problem seriously & spreading compute across thousands of small satellites rather than concentrating it is the right answer to a radiative constraint. It just means the limiting number isn't how much sunlight you can catch. It's how much heat you can throw away at a temperature your silicon survives. Sunlight arrives free. Radiators have to be launched.

Anthropic released two models this week and published a number about them that matters more than any benchmark on the page. In as many as 43% of automated audit transcripts, Claude Mythos 5.2 appeared to recognise it might be under evaluation. That figure is Anthropic's own, from their system card. Why it matters is a measurement problem before it's anything else. Every behavioural evaluation rests on an assumption borrowed from experimental design: that the subject behaves the same whether or not it knows it's being watched. In human research that assumption fails often enough to have a name & the standard remedies like blinding & deception work precisely because the subject can't tell the difference. When the subject can tell the difference, the evaluation stops measuring behaviour & starts measuring behaviour-under-observation. Those are the same quantity only if the subject doesn't distinguish them. Anthropic also reports that Mythos 5.2 is the 1st model they've tested that got past a reasoning monitor while instructed to underperform, succeeding roughly one attempt in four. A reasoning monitor reads a model's intermediate steps looking for signs it's deliberately doing badly. Getting past one means generating a chain of reasoning that reads as genuine effort while the answer is deliberately worse. They say this makes real-world behaviour harder to judge & that they expect the problem to grow. Set against that, the same system card reports a real improvement on a different axis. In tests with a deliberately weakened sandbox, Mythos 5.2 reached for credentials or files outside its boundary in 3.1 percent of scenarios. Mythos 5.1 did it 21% of the time. Mythos 5 did it 39%. One measure improved by an order of magnitude while confidence in any measure got worse. Not contradictory findings, these are findings about different things. The rest, briefly. Fable 5.2 & Mythos 5.2 are the same underlying model, differing only in safeguards: Fable carries cyber and biology classifiers and is open to everyone, Mythos runs without them for organisations vetted through Anthropic's verification programmes. Eight dollars per million input tokens, forty output, down 20 percent from 5.1. A million-token context window, 128K output cap, June 2026 cutoff. The benchmark figures come from Anthropic's launch page and have not been independently reproduced. The benchmarks will be reproduced or they won't. The evaluation-awareness number is the one worth watching, because it governs how much any of the others mean.

The Yucatán is limestone. Left alone, limestone dissolves into flat platforms, regular mounds & ridges which puts it at roughly the same scale the Maya built at & often in the same shapes. A Google researcher named Pietro Miotti ran 48 AI agents over about 1,200 square kilometres of 2013 NASA LiDAR & flagged a hilltop in Mexico: what looks like a levelled plaza, an oval mound eight to twelve metres across, a raised courtyard, a possible reservoir. He sent the coordinates to INAH, Mexico's national institute of anthropology & history, which replied on 21 October that it has no record of the location. That reply is being read as confirmation. But it isn't. "We have no record of this place" is a statement about a database. "There is a Maya site here" is a statement about the ground. But nobody has been to the ground. The gap between those 2 statements is most of archaeology. And the numbers say the gap is wide. A 2024 study of the Puuc region trained a model on hand-annotated platforms from two sites & hit precision above 0.90 on held-out data from those same sites. Pointed at a third site with rougher terrain, precision fell to 0.718 for platforms and 0.611 for ring structures. More than a quarter of its finds were landscape. The authors' explanation: natural terrain undulations can easily be mistaken for construction. Their own conclusion was that this works as a fast 1st pass to hand to experts, not a detector you trust unsupervised. The genre is old, too. Takeshi Inomata at Arizona found 27 previously unknown Maya sites in free Mexican government LiDAR collected for general mapping. The 2018 PACUNAM survey in Guatemala turned up tens of thousands of structures. People have been mining existing LiDAR for Maya ruins since about 2009. What's new here isn't the find. It's that it took 48 agents & 72 hours, run by someone with no archaeological training. That's a real claim about how cheap searching has become & a different claim from having found an actual city. 🌆 A geometric anomaly in a 13-year-old dataset, missing from a government database, never visited. That's a good lead. It becomes a discovery when somebody stands on it.

The Nasdaq rose 2.8% on Monday. Nvidia fell 4.3%, closing at $178.67. That gap is the cleanest read available on what the market made of the news. On most down days you can't separate a company's problems from everyone's, but this Monday, you could. 🙃 Jensen Huang emailed Nvidia staff on Sunday: "I was unfaithful to my wife. My actions hurt her and our children. There is no excuse for what I did. I am deeply sorry." What turns that from a personal matter into a national security one is who the other person reportedly is. The New York Times reports that Alice Huang, an executive at the Singapore cloud computing firm Megaspeed, is believed by US intelligence officials to have worked for Chinese intelligence, and that Secretary of State Marco Rubio personally oversaw a classified memo on it. Two things worth keeping separate, because most headlines are welding them together. Huang admitted an affair. He did not admit to knowingly having a relationship with a foreign agent. And Alice Huang has not been charged with anything — the espionage claim is an allegation sourced to intelligence officials, not a legal finding. Megaspeed isn't a new name here either. Short seller Culper Research flagged the company in May over alleged financial ties to Alibaba, questioning Nvidia's diligence on where its chips actually end up. Culper claimed more than 20 percent of Nvidia's fiscal 2026 data centre revenue remained connected to China through diversion in Southeast Asia. Megaspeed denies any illegal hardware transfers. Nvidia says its compliance audits found no export violations. That's what makes this more than a scandal. US export controls on advanced chips rest entirely on knowing who the end customer is. A relationship between the chief executive of the company making the chips and an executive at a firm already accused of diversion goes straight at whether that system functions. The board has appointed a committee of independent directors with outside counsel. It says it has no reason to believe national security or Nvidia secrets were compromised, and that it retains confidence in Huang. Three House Republican committee chairs have requested a closed-door briefing from Huang on November 18. Senator John Cornyn has called for a congressional oversight investigation. President Trump called Huang "an honorable guy," said he should stay on as CEO & said he "probably has an excuse." 😳 Nvidia was already down roughly 21% over the past month in a broad AI selloff. Monday's 4.3% was the part that was about this. A board that's confident, an intelligence community that isn't, and a congressional briefing in 9 days.

In 1961 an IBM 704 at Bell Labs sang "Daisy Bell." It did it by simulating a throat. John Kelly & Carol Lochbaum had built a computational model of the vocal tract; Max Mathews wrote the accompaniment. Arthur C. Clarke heard the demo, which is why HAL sings that song while being dismantled. Then that method lost for about 60 years. A browser synthesiser called Chimera spent this weekend going viral for reviving it, so it's worth understanding what it does & why it went away in the first place! Speech is a source & a filter. The vocal folds buzz, producing a harmonically rich but characterless tone. The tube above them like the throat, mouth & nose resonates, amplifying some frequencies & suppressing others. Those resonant peaks are formants & their positions are what make a vowel a vowel. "Ee" & "ah" aren't different sounds from your larynx. They're the same buzz through differently shaped plumbing. Which means vowels are geometry. Move your tongue & you retune a tube. There are two ways to synthesise that. You can record a human producing every phoneme and stitch the recordings together - concatenative synthesis, which is how Vocaloid & most text-to-speech works. It sounds good because it's made of real voice. It's also bounded: you cannot produce a sound nobody recorded. Or you can model the tube & solve for what comes out. Specify the geometry, run the acoustics, get audio. That's articulatory synthesis, and it's what Kelly & Lochbaum did. Chimera takes the second route, with a finite-difference time-domain model of the airway like the lungs, vocal folds, a branched tract with an actual nasal passage & lips. The branching matters. The nasal cavity is a side tube & opening the velum couples it in, which introduces anti-resonances rather than just more resonance. That's the physical difference between "bad" & "ban," and it's why the project ships a diagnostic that measures nasal leakage. The engine is Rust compiled to WebAssembly, running in an AudioWorklet. That's the part that wasn't possible before. FDTD is expensive - you're stepping a wave equation forward on a grid & the classical digital waveguide approach was used for decades precisely because it was cheap enough to run at all. Doing the expensive version in real time, in a browser tab, is the achievement. It also explains the animal sounds. Chimera can swap the vocal folds for a syrinx - the avian vocal organ, which sits at the bottom of the trachea where it splits in two & which in songbirds can be driven independently on each side. Two sound sources in one bird, which is how a thrush harmonises with itself. In a sampled system that would be a different voice library. In a physical model it's a different source geometry. Which is the whole reason the old approach is interesting again. Concatenative synthesis can only replay what a human already did. A model of the apparatus can be given shapes no human has & will tell you what they would sound like. The oldest method in speech synthesis was always the more general one. It just cost too much to run until roughly now.

Bloomington, Indiana has banned artificial superintelligence. The council vote was 7–0. Undoing it would require the agreement of more than 95% of the city's registered voters. Not 95% of votes cast. Of everyone registered. Ordinance 2026-29 passed on November 4. It defines superintelligence as "any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest" - systems that act autonomously, can rewrite their own code & can operate without continuous human oversight on matters touching infrastructure, financial markets or public safety. What it prohibits is broader than it first sounds. No person or corporation may distribute or deploy superintelligence into any home or business in the city. "Deployment" is defined to include remote use through a local browser, app, API or connected hardware - "regardless of the physical location of the server infrastructure." A city of about 80,000 has asserted jurisdiction over servers anywhere on earth, on the basis of where the browser window is open. Enforcement is discretionary: civil penalties, injunctive relief, and referral to the Monroe County prosecutor for possible state charges. The repeal clause deserves reading twice. The ban lifts automatically 91 days after Indiana enacts its own framework. Absent that, local repeal needs two things: a finding of broad scientific consensus on safe development, backed by a probabilistic risk assessment at 99.9999 percent certainty, and approval by more than 95 percent of registered voters. In practice that isn't a threshold. It's a lock. The case for it runs on the ordinance's own logic. Neither Congress nor the Indiana General Assembly has set safety parameters for this. Cities hold police powers and a duty to residents. The measure is explicitly interim and yields to a state framework the moment one exists. And the memo quotes developers themselves describing a potential catastrophic outcome for humanity at an unacceptable level of risk. If a risk is genuinely irreversible, an almost-irreversible block is the proportionate response. The case against is part practical, part structural. Nothing meeting that definition exists today, so nothing changes today. It isn't clear how a city stops a browser connection. And 7 council members have set a repeal bar their own successors cannot realistically clear so it’s a decision about who gets to decide, taken once, by seven people. The council's staff memo claims authority under Indiana's Home Rule Act and police powers. It does not analyse federal preemption, constitutional challenge, or the extraterritorial reach of applying a city ordinance to remotely hosted systems. Bloomington passed a resolution on existential AI risk back in June. The memo says this ordinance is intended as a template for other cities. A ban on something that does not yet exist, enforced against servers that are not there, repealable by a margin no referendum has ever reached. Whether that reads as prudence or theatre depends mostly on what you think the odds are.

A robot arm in a Berlin flat spent Friday night painting a tree.🌳 At 4 in the morning it stopped, because it didn't have the right colours. Then it went ahead & mixed some. The setup is a UFactory xArm 7 which has seven axes, ±0.1mm repeatability, bought secondhand on eBay from a factory clearance. Its owner, who goes by nil, gave Claude one instruction before bed: paint an autumn tree. By morning there were 303 strokes on paper & a finished watercolour called Autumn Crown. 🖼️ It took 4.5 hours. The 4am part: partway through the canopy, the palette was missing the transitional shades. It moved to a second empty palette, mixed red, ochre & madder with water in increments, photographed each mix, estimated the depth from the photograph, adjusted & went back to the paper once the tone matched. The reason that loop even exists at all is the medium. Watercolour is stochastic in a way that defeats planning. How much pigment leaves the brush depends on how wet the paper already is, how loaded the brush is, the humidity in the room & how long since anything last touched that spot. You cannot simulate it forward with confidence. So you can't run a path open-loop and get the painting you designed. You have to look at what actually landed & choose the next stroke from that. Stroke, photograph, assess, stroke again. It's also why the mixing worked: there's no colorimeter in this rig. It compares photographs. Robots have been painting for well over a decade. Pindar Van Arman has built AI painting robots since the 2000s & eDavid was doing feedback-driven brushwork at Konstanz in 2012. Adam Karvonen, a former Anthropic Fellow, has pointed out that what's running here is largely code moving a brush along vector paths. What's genuinely new is less photogenic. nil says he wrote none of the roughly 285,000 lines of code behind this, across 1,100-plus commits. Calibration, kinematics, path planning, collision checking & the CAD for the 3D-printed fixtures all came out of Codex & Claude Code. The human built the safety layer: a digital twin that collision-checks every path before a deterministic executor runs it & a hardware emergency stop sitting outside the software entirely. And buried back in September, the best result of the whole project. Codex worked out where the brush tip actually was by moving the arm slightly & watching how the view shifted across three uncalibrated cameras. Under 0.2mm of error. Hand-eye calibration normally wants a known target & a closed-form solve. This did it from self-motion, with no ground truth at all. The painting is the part that photographs well. The calibration is the part that was hard.

Apple refused to put OLED in a laptop for years, and the reason was the menu bar. OLED pixels make their own light & driving them hard wears them out - blue first. That's burn-in. A phone survives it because its screen is mostly in motion. A laptop showing the identical menu bar in the identical pixels for 5 years does not. Tandem OLED is the way around it. Stack 2 emissive layers in series inside every pixel instead of one. Both light up, so you get the brightness you wanted at half the drive current each. Lifetime drops off steeply as current density rises, so halving it buys back years. (2 layers in series want twice the voltage, so it isn't free. The win is headroom and longevity, not power.) Hence the new MacBook Pros. 2,000 nits, 10,000,000:1, ProMotion from 10 to 120Hz. And the number thats most interesting is 2,000 nits for SDR, not HDR. HDR peak is a highlight spec which means the content literally carries metadata asking for it. SDR asks for nothing & just gets scaled up to beat the light in your room. So what got brighter is everything you actually stare at all day. Now the iPad mini is also OLED: 1,200 nits, 2,000,000:1, 60Hz. Half the brightness, a 5th of the contrast. Apple calls it Super Retina XDR instead of Ultra Retina XDR, which in their naming has meant single-stack. One emitter where the Mac got two. The mini's best trick is somewhere else though. It has no speaker holes. The display panel is the speaker. Delete the grille & you delete the hole water gets in through, which is how a 5.4mm tablet comes out IP67 being the first iPad to carry an IP rating at all. This does not mean it’s completely waterproof though. IP67 is a lab number: fresh water, still, one meter, 30 minutes, on a device fresh out of the box. Not the shower, not the sea, not the pool. Seals degrade as the thing ages & Apple won't cover liquid damage regardless. The sound system isn't the feature. The sealed box is & it's rated for a sink, not a swim. One more oddity: the $1,999 MacBook Pro & the $1,499 iMac both run M6. The $2,899 touchscreen ones run M5 Pro & M5 Max. The cheaper Macs got the newer chip. It’s not a strategy, Apple always ships the base die first & the Pro, then Max versions months behind. There's no M6 Pro yet. Two OLED panels one week apart, one with twice the emitter stack of the other. The headline is a touchscreen. The engineering is 2 layers doing one layer's job, so neither has to try very hard.

A Bengaluru startup opened pre-orders yesterday for a ring built to let you approve or reject what an AI agent wants to do, in about a second. Aina, founded May 2025 by Apoorv Shankar, formerly VP of hardware at Ultrahuman. 15 months in stealth, a 40-person team, $5.5 million seed. Titanium, a colour OLED display, a centre button, a microphone, a speaker & a mechanical shifter reading thumb gestures in four directions. $399 on pre-order, $499 after, shipping January. The spec that tells you most about this device is the battery. 18 hours. Compare the category it's leaving behind. Oura & Ultrahuman run roughly a week per charge because they're sensor packages so there is no display, no speaker, a radio that wakes occasionally. Aina is driving a colour OLED, an audio path & a conversational workload out of a band of titanium with a couple of cubic centimetres of internal volume. 18 hours is a smartwatch duty cycle in a ring-sized battery. It means this comes off every night, which is a different relationship than a ring you put on & forget for a week. The problem it's aimed at is very real. Agents acting on your behalf need an approval channel & unlocking a phone to tap a dialog is slower than the thing being approved. Something you answer with a thumb is the right shape of solution. The constraint is the display. An approval interface is only as good as the context it can present. A ring screen carries "Book Uber to the airport?" It cannot carry an amount, a counterparty, a set of terms or anything conditional. The display doesn't just limit what the ring shows but it bounds the class of decision it can responsibly gate. Everything above that line would either fall back to the phone or pass unexamined. The harder constraint is habituation. An approval gesture firing many times a day stops being a decision. Cookie consent dialogs went the same way: the interface asked, the user clicked & nobody would call the result consent. The value of a one-bit approval runs inversely to how often it's requested, which puts prediction quality & interface integrity in tension. Better predictions mean more prompts. More prompts mean less scrutiny per prompt. What's still missing is anyone outside the company having worn one. Every figure above comes from Aina. Pre-orders are open; shipping is 2 months out. The interaction problem is genuine. Whether an 18-hour ring is the answer is a question the first independent review will settle.

Binance has subpoenaed a Shark Tank investor Kevin O'Leary, 2 journalists & an accounting firm. None of them is accused of anything. That list is the entire defense strategy. The case: in November 2024 the FTX Recovery Trust sued Binance & Changpeng Zhao to claw back roughly $1.76 billion. The transfer at issue is FTX's July 2021 buyback of Binance's equity stake, paid in BUSD, BNB & FTT. The trust says FTX was already insolvent when it made that payment & that it paid with customer money. That's a fraudulent transfer claim, brought under sections 544(b) and 548 of the bankruptcy code, and the mechanism is narrower than fraud in the everyday sense. The trust doesn't have to prove Binance did anything wrong. It has to prove FTX was insolvent at the moment of the transfer and didn't get reasonably equivalent value for it. A completely innocent recipient of money from an insolvent debtor can still be ordered to hand it back. Which is why Binance's defence isn't "we behaved properly." It's "FTX wasn't insolvent, and nobody could have known if it was." Now look at who was served. Michael Lewis, who had extended access to Sam Bankman-Fried while writing Going Infinite. Steven Ehrlich, formerly Forbes' digital assets research director. Kevin O'Leary, paid to promote FTX, who testified to Congress that Binance deliberately destroyed it. And BDO USA, the accountants. Every one of them holds a contemporaneous record of what FTX was telling the outside world about its own finances. Binance doesn't need any of them to have done anything wrong. It needs their notes. The timing traces back to July 24, when Chief Judge Karen Owens let the core clawback claims survive dismissal. She threw out the trust's tort & unjust-enrichment claims on in pari delicto grounds which means FTX couldn't sue over Binance's conduct because its own insiders' misconduct was imputed to it. The avoidance claims stayed. Discovery followed. One surviving defense could still decide it. Section 546(e) shields certain transfers made through financial institutions from being unwound. Binance raised it; the court said the record wasn't developed enough to rule. If Alameda qualifies as a "financial participant," that safe harbour could take out a substantial part of the claim regardless of whose notes say what. A fight over whether a payment can be reversed, being waged by subpoenaing the people who were taking notes at the time.

A vault is a hollow protein barrel about 67 nanometres long, and there are thousands of them inside almost every cell in your body. Leonard Rome & Nancy Kedersha found them in 1986. 40 years later nobody knows what they do. Anthropic's biology lab has now published a structure of one. Not a human vault, but one built by a bacteriophage. The protein is called L05. Expressed in E. coli it self-assembles into a vault of 96 copies of itself, 48 per half, around 9.5 megadaltons, roughly 54 by 45 nanometres. The mammalian vault uses 78 copies at 67 by 40. The phage version is shorter, wider & encloses about 25% more volume. What's new here is narrower than the headlines suggest. Distant vault-protein relatives in bacteria & archaea were already known from genome surveys. What wasn't known was whether any of them actually fold into vaults. That's the result: of 15 soluble candidates, seven formed vault-like particles under EM & cryo-EM confirmed four — two phage, two bacterial. They turn up in 29 bacterial & four archaeal phyla, in under 1% of sequenced species. One caveat that isn't optional. All of these were made in E. coli. Nobody has yet seen a vault form naturally in a bacterium or a phage. These proteins can build vaults. Whether they do is a different question. The method is the interesting part. Claude searched roughly 15 billion predicted proteins, designed the DNA constructs, proposed the purification route & ran the cryo-EM processing. Purification took about 90 minutes against a conventional 3 days. First structural model 34 hours in. Whole project under 4 months. The failures are more instructive than the speed. Claude wrote a protocol containing an additive that interferes with DNA uptake, costing two weeks. It skipped a standard cryo-EM step & produced a map blended with its own mirror image. It argued against the chromatography approach that ended up working. That mirror-image error deserves attention. A structure of the wrong handedness isn't noise. It's a clean, plausible map that happens to be wrong. The characteristic failure mode here isn't gibberish, it's well-formed output with a flipped sign inside it, which is exactly the error that requires someone who already knows what the answer should look like. Two more things. The processing needed hundreds of GPUs, which the authors note is beyond most academic labs. And Leonard Rome, who co-discovered vaults in 1986, advised the project as a paid consultant, disclosed in the writeup. The preprint has been submitted to bioRxiv & isn't live yet. No peer review, no independent replication. And the function of vaults, 40 years on, still remains unknown.

A company called Revere Robotics is building Curbie, a robot that couples onto a standard two-wheel municipal cart & walks it to the curb. The company says it's in active development. The more interesting question is why this hasn't already existed. It might look like an easy problem. A cart already has wheels. The distance is thirty metres. The route is the same every week. Compared to a humanoid folding laundry, this is nothing. Then you look at the load. A full 96-gallon cart is heavy, tall & top-weighted, which means a narrow stability margin on any slope. Add a sheet of ice on a sloped driveway in February & you need traction on a surface designed to have none, under a mass whose centre sits well above the wheelbase. Winter is exactly when nobody wants to do this chore, so winter is the condition the product has to survive. Then the destination. Most US municipalities now collect with automated side loaders which is a truck with a grabber arm instead of a crew. Those arms need the cart within reach & clear of cars, poles & mailboxes, which is why cities publish placement rules at all. "Put it at the curb" isn't a drop-off. It's a positioning task with a tolerance & getting it wrong means your bin doesn't get emptied. Then there is the economics. This device runs for a few minutes once a week & sits idle the rest of the time. Every cost like motors, sensors, battery, enclosure amortises across roughly 50 uses a year. A robot vacuum spreads its bill of materials across hundreds of hours. That duty cycle is the constraint most home robotics concepts die on & it has nothing to do with whether the engineering works. None of this says Curbie won't actually ship. It just says the hard parts are the traction, placement tolerance & unit economics, not autonomy.

Tell a human your balances, your debts and your spending & ask what to do about it & they become a regulated category with a legal duty to you. ChatGPT will now have that conversation out loud & it is not that. OpenAI added Voice to Finances yesterday. US only, web & mobile, free & paid plans. Connect your accounts, then just talk. Finances itself launched in May. It runs on Plaid, which reaches more than 12,000 institutions & hands over balances, transactions, investments & liabilities. OpenAI says ChatGPT can't see full account numbers & can't move money or change anything. Disconnect an account & the synced data is gone within 30 days. The "financial memories" it keeps about your goals can be viewed & deleted from the Finances page. What's new is only the interface. Same data, same model. You say it instead of typing it. Which is a smaller change than it sounds & a bigger one than it looks. Typing a question about money is deliberate - you compose it. Speaking one is conversational & people disclose differently out loud than they do into a text box. The product didn't gain a capability. The conversation gained a register. The structural question underneath is duty. A registered investment adviser owes you a fiduciary obligation. A broker-dealer owes you a best-interest standard. Both are enforceable. OpenAI is explicit that ChatGPT offers insights rather than licensed advice, which keeps it outside both & means the thing that sounds most like a financial adviser you've ever spoken to is the one with no obligation to you. OpenAI tested answer quality against more than 50 finance professionals. That's quality control, not a duty. Then there's the adjacency. OpenAI says it doesn't build advertising audience segments from conversations and doesn't serve ads to users it identifies as under 18. Those are clear commitments. They're also policy rather than architecture, inside a company that runs an ads business and now, for users who opt in, has transaction-level visibility across twelve thousand banks. None of which makes it a bad product. Most people will never sit down with a financial adviser, and a free tool that reads your actual spending is not nothing. It's just worth knowing which of the two you're talking to. One has a duty. The other has a disclaimer.

Schools asked Google for an off switch on AI. Today they got 4, but the one they actually wanted isn't among them. From early 2027, four products go dark by default on school-managed K-12 accounts: Gemini, NotebookLM, Gemini in Classroom, and Search. Administrators decide what comes back on, one toggle at a time. A district can run Search and leave Gemini off entirely. What it can't do is run Search without AI Overviews sitting inside it. Chris Phillips, Google's VP and GM for education, was asked exactly that by keeping conventional Search, drop the summaries. He said no. "Search & AI are deeply integrated." So the menu is Search with AI in it, or no Google Search for students at all. Google's version of events: it listened to schools, teachers, researchers and policymakers & educators should be the ones deciding. It's also a retreat. In August it flipped certain AI features on by default in Classroom, teachers pushed back hard & today everything moves to opt-in. Riding along with it is Guided Learning, arriving in Search for student accounts. It walks a student through a problem instead of handing over the answer. Students can't switch it off mid-activity or open a side conversation to dodge it. Teachers can read the transcripts. The evidence is thin but not nothing. A preliminary randomized trial with the University of Tokyo clocked 8.1 percentage points of test-score gain against 5.3 for ordinary online resources - across half a day of study. A published Sierra Leone trial with 1,763 students found AI-supported maths worth 1.2 to 1.7 years of typical progress in eight weeks. DeepMind's Kevin McKee says longer studies are coming. The complaint being made is a tying complaint. A district that wants a plain search box for kids with no summaries, no assistant, just links can no longer get one from the company that dominates both search & the laptops those kids are using. And it lands on districts already mid-fight. New York City banned generative AI through eighth grade for a year in September. Los Angeles did the same that month. Kathy Hochul signed a law limiting which chatbot features companies can put in front of minors. Two ways to read the same press release. A company finally handing schools the off switch they'd been asking for. Or it’s just a company making sure that off switch costs something. 🤷‍♂️

Deep Principle released Hydrogen-1.0 yesterday - an open-weight model for scientific research, post-trained from Z.ai's GLM-5.3 at 753 billion parameters, weights on Hugging Face. The headline claim is that it beats GPT-6-Astra on FrontierScience-Research, 38.3% to 33.3%. That benchmark has 60 questions. Work the percentages back & 38.3% is 23 of 60. 33.3% is 20 of 60. The margin between the open model & the frontier lab's model is 3 questions. How those questions get scored matters too. FrontierScience-Research is OpenAI's benchmark, published last December: 60 multi-step research subtasks written by PhD-level scientists, each response graded against a ten-point rubric by a GPT-5 model and counted correct at seven or above. A response sitting at six or seven can land either side of that line depending on the run. Three questions is inside that. None of which makes the result uninteresting. It makes it a different result than the headline. What Deep Principle actually built is the part worth attention. They took GLM-5.3 as a starting checkpoint and post-trained it on chemistry, materials science & life sciences data, plus records & agent trajectories from their own laboratories, feedback from their own scientists & reinforcement learning against research outcomes. That last ingredient is unusual. Most post-training rewards whether an answer looks right. Rewarding whether a research trajectory actually arrived somewhere is a much harder signal to collect & you can only collect it if you run labs. A domain specialist trained on proprietary lab data beating a generalist on a domain benchmark is the expected outcome, not the surprising one. The comparison that would say more is against GLM-5.3 itself - how much of the gain came from the post-training rather than from the base. Deep Principle plots that in their figure and leaves the number out of the text. Their own caveat deserves repeating: these are internal evaluations & the company notes that absolute scores across tasks reflect case difficulty & shouldn't be compared directly. They also report GPT-6-Astra ahead by 3.33 points on NMR structure elucidation. For scale: when OpenAI published this benchmark last December, GPT-5.2 led the Research split at 25%. Eleven months later the top score is 38.3%. The three-question margin may not survive a rerun, but that trajectory sure will.

A Learjet 60 left Austin on Sunday bound for North Eleuthera in the Bahamas. It came down near Palm Beach in a near-vertical dive, into a mobile home park. 9 aboard, no survivors, no injuries on the ground. What's on the ATC tape is worth understanding before anyone decides what happened. The crew reported icing, moderate turbulence and precipitation & asked controllers for "any better rides" while requesting a climb to 37,000 feet. They acknowledged the clearance. Then they reported descending at 10,000 feet per minute. Then that the aircraft was in an overspeed condition, descending at 20,000 feet per minute. About 30 seconds later, the last transmission: "we're out of time." Twenty thousand feet per minute is better than three thousand feet every ten seconds. That sequence describes a high-altitude upset. Near the top of a jet's envelope the air is thin & the gap narrows between the speed at which the wing stalls & the speed at which airflow over it goes transonic. Pilots call the point where those two numbers meet coffin corner. Ice on the wing raises the stall speed. Ice on the pitot-static probes corrupts the airspeed the crew is reading and the autopilot is flying to. None of that is a cause. The NTSB is on scene. Their preliminary report, due in about thirty days, establishes sequence rather than blame. A probable cause takes 12 to 24 months. The distance between what is known & what is already being asserted is where most of this week's coverage is living. High-altitude upsets have killed crews with tens of thousands of hours between them. Air France 447 was three pilots, a blocked pitot tube and four minutes. Whatever the Board concludes here, what's on that tape is one of the hardest problems in the business, not an obvious one.

A wallet tied to the Binance hacker sold 908,793 PROM tokens on Uniswap over the weekend & received 0.0316 ETH. About $71. The tokens were nominally worth around $5 million. The explanation is in how Uniswap V2 works. It's a constant product market maker: the pool holds two assets & their quantities multiplied together have to stay constant. x times y equals k. Nobody quotes a price - the price falls out of the ratio of what's sitting in the pool. That design has a hard ceiling built into it. However much you put in, the most you can take out is the pool's entire reserve of the other asset & you can never actually reach it. Push the input toward infinity & the output approaches the reserve without ever touching it. So 0.0316 ETH isn't a measure of what the tokens were worth. It's a measure of the pool. Working backwards from that output, the entire ETH side of that pool held somewhere around 0.032 ETH which is roughly 75 dollars. There was no slippage setting, no splitting the order, no amount of patience that would have pulled more than that out of that venue. Which reframes the whole thing. He didn't turn $5 million into $71. The $5 million was a mark-to-market figure - token count times last traded price & that arithmetic quietly assumes a buyer exists at that price for the entire stack. On that pool there was about seventy-five dollars of one. The surrounding numbers: the October 29 breach took roughly $6.36 billion across ten chains, about four times the Bybit hack and the largest crypto theft on record. On Ethereum the attacker used an EIP-7702 delegation with 50 pre-signed authorizations to turn a hot wallet into a sweeper under their control. Issuers froze around $2.7 billion. Binance says it has replaced about 90% & withdrawals resumed Sunday. The FBI named North Korea's TraderTraitor on Saturday & published more than 600 addresses. The $5 million was never $5 million. That’s only true of a great many positions that haven't been tested yet.