A claim by OpenAI to have solved one of mathematics’ most intractable problems has led to a dispute with academics who say they were first. CC-licensed photo by Dunk đ on Flickr.
You can sign up to receive each dayâs Start Up post by email. Youâll need to click a confirmation link, so no spam.
A selection of 9 links for you. Does not compute. I’m @charlesarthur on Twitter. On Threads: charles_arthur. On Mastodon: https://newsie.social/@charlesarthur. On Bluesky: @charlesarthur.bsky.social. Observations and links welcome.
OpenAI just claimed a huge math discovery. Some academics are crying foul âą WIRED
Will Knight:
»
OpenAI said on Tuesday that it has found an AI-generated solution to one of the biggest problems in mathematicsâa 200-year-old equation that describes the natural behavior of fluids like water and air.
The announcement appears to demonstrate the stunning power for AI to advance mathematics. But it has been marred by claims from another mathematician, Tristan Buckmaster, who says that OpenAI rushed ahead to solve the problem after learning of the work of himself and a colleague, the Anthropic researcher Levent Alpöge. Buckmaster claims OpenAI also tried to influence who got credit for the work.
The proof concerns the Navier-Stokes equation, one of the unsolved problems in the Clay Millennium prizes, which are each worth $1m.
Sebastien Bubeck, a mathematician and AI researcher at OpenAI, said in a press briefing that the company began training a new AI model with advanced mathematical capabilities on August 28.
After reading rumours that Anthropic was making progress toward solving Navier-Stokes, Bubeck said the company decided to dedicate more resources to tackling the problem. The company had more than 1,000 agents tackle the problem over more than 50 hours; eventually, as many as 10,000 agents were hard at work before the company discovered that it had come up with a solution.
âI thought there must be a mistake somewhere,â Bubeck said. âAnd on Sunday morning we had the final solution, Lean-formalized and everything.â (Lean is a programming language that can be used to formalize mathematical proofs.)
âŠOn Monday, Alpöge and Buckmaster, a mathematician at NYU, posted documents claiming key advances in an area relevant to the Navier-Stokes problem. The pair says they used several AI models, including Claude and Codex, to complete their work.
Buckmaster also posted a statement claiming that, last week, he learned that OpenAI had become aware of his and Alpögeâs work and had started putting significant resources toward the problem. Buckmaster claims that he asked OpenAI leaders about whether the company had accessed the pairâs Codex logs. He says he was told the model âdidnât look up user dataâ but claims the company didnât respond to questions about training.
«
The row about authorship might overshadow the fact that this (partial; some of the constraints are “relaxed”) solution is a remarkable advance again in mathematics.
unique link to this extract
EV batteries last longer than drivers feared âą Financial Times via Ars Technica
Kana Inagaki and Clara Murray:
»
Electric vehicle batteries are lasting longer than previously feared by drivers, with most used EVs able to retain about 90% of their original usable battery capacity after 150,000km (93,000 miles), a new study has shown.
Despite a surge in global EV sales on the back of rising fuel prices, long-term battery durability continues to be one of the key concerns for drivers when they consider switching from a petrol model to an electric car.
EV battery warranties typically cover eight years or 100,000 miles (160,000 km) with car manufacturers required under the warranty to provide a replacement battery if capacity falls below 70%.
According to the latest study published by Aviloo, an Austria-based group that analyzes battery health, the median state of health (SoH)âwhich represents the remaining percentage of a batteryâs original usable capacityâfor 20 popular EV models stood between 87% and 94% after 150,000 km.
The median SoH was between 91% and 97% after 50,000 km, and 88% and 95% after 100,000 km. The study was based on more than 500,000 tests Aviloo carried out globally on the 20 EV models including Teslaâs Model Y and Volkswagenâs ID.4 between 2022 and 2026.
The Aviloo study, which did not include Chinese models due to lack of data, also showed that the speed of battery degradation varied significantly even within the same EV model depending on climate conditions, battery size, and driver habits, such as parking.
âIn contrast to a combustion engine car, where age and mileage would more or less carry the value of the car and technical condition, thatâs not the case for an EV,â said Marcus Berger, chief executive of Aviloo. âThe car might look the sameâsame age, same mileageâand you donât know how it has been treated.â
For Nissanâs Leaf ZE1 model, which requires more frequent charging to cover the same distance than other EV models due to the smaller size of its battery, the median SoH at 150,000 km varied by as much as 13.5 percentage pointsâequivalent to about 29 km of real-world range per charge and the largest spread among the 20 EV models covered by the study.
«
So that’s one of the most persistent claims killed. The problem of charger infrastructure and pricing remains, but as EV sales grow you’d expect that would start to solve itself. (Note: this might be over-optimistic.)
unique link to this extract
UK to force Apple and Google to block explicit images on childrenâs smartphones âą The Guardian
Dan Milmo and Kiran Stacey:
»
Apple and Google will be forced to block explicit images on childrenâs smartphones by law in the UK after talks failed to produce a breakthrough, the government has said.
The legislation commitment puts ministers on a collision course with two of the biggest players in Silicon Valley, months after the government announced a social media ban for under-16s.
Lisa Nandy, the secretary of state for digital, culture, media and sport, said the government would start work on an act of parliament to stop under-18s taking, sharing or viewing nude images on their devices.
Talks with Google and Apple, whose operating systems dominate the smartphone market, had not reached agreement as a three-month deadline passed on Tuesday.
Nandy told the House of Commons the tech companies could still reach a deal but it was âalmost certainâ groundbreaking legislation would be required.
âFor too long, technology companies have failed to protect children from some of the worst harms,â she said. âAnd to be frank, I am not prepared to give them the benefit of the doubt that progress will continue at the pace that we need it to, while harm is being done now.â
The prime minister, Andy Burnham, said the protections would be âbuilt into the device and switched on by defaultâ, and tech companies will be required to verify an adult userâs age before they are allowed to take, view or send nude images. The government also wants the restrictions to apply to apps used by children, which will require further legislation.
Nandy said: âThis will ensure that the whole ecosystem changes, tackling the means by which children are most often abused or groomed in the online world.â
Sources familiar with progress in the talks have said there iwas a desire to reach a deal on both sides but significant hurdles remained including privacy concerns on the tech side.
«
In general, I’d say any idea advocated by Lisa Nandy is stupid and/or wrong. This sounds like it’s probably unworkable, involving every adult verifying themselves to, what, the government? Or every single app? What if a verified phone gets passed on? Or borrowed? How do you stop them taking photos? It’s just bonkers. The classic “nude pictures are a problem, therefore we must stop nude pictures” solution.
unique link to this extract
The Pentagon asked OpenAI for artificial intelligence designed to rarely say no âą The Intercept
Sam Biddle:
»
The Department of Defense asked OpenAI to provide the U.S. military with a special version of its artificial intelligence technology designed to turn down the military commands as infrequently as possible, according to documents obtained by The Intercept.
The desire for a custom AI tool with âminimal refusal ratesâ to Pentagon commands was revealed in files released to The Intercept as part of a Freedom of Information Act lawsuit seeking information about the militaryâs secretive deals with AI companies.
OpenAI, along with rivals Google, xAI, and Anthropic, all agreed in 2025 to develop militarized prototypes of their state-of-the-art AI to assist the armed forces in uses including logistics, intelligence decision-making, and general âwarfighting.â Earlier this year, the Pentagon sought to expand the scope of these agreements and deploy them across U.S. classified computer networks, resulting in a high-profile showdown with Anthropic over whether and how the company could restrain military uses of its technology.
The clause seeking âminimal refusal ratesâ from OpenAI appears in an updated contract â version âP00003â â expanding upon last summerâs prototype deal, worth up to $200 million over the contractâs two-year duration. A separate document, signed by both OpenAI and the government on February 6, indicates that OpenAI agreed on that day to the contents of an expanded version âP00003.â
OpenAI and the Pentagon deny agreeing to such âminimal refusal â language, claiming that the âP00003â document provided to The Intercept was a draft and not the final version.
«
The story is essentially about the enormous amount of reverse ferreting by the DoD (isn’t it the DoWar now?) and OpenAI over whether the contract is FINAL_THISONE_USETHIS or FINAL_THISONE_USETHIS-v2.
As to why the DoW would want “minimal refusal rates”:
»
a large language model that would aid in the process of spying on, targeting, and killing people would require a substantial relaxation of safeguards to be useful for any military.
«
DeepMindâs new genome âatlasâ charts effects of all nine billion human gene mutations âą Nature
Ewen Callaway:
»
The human genome is an easy place to get lost. Only 2% of its 3 billion letters encode proteins and the rest is diabolically hard to decipher. An AI-generated “atlas” of the human genome unveiled on Tuesday by Google DeepMind aims to guide scientists through our biological code.
One of the most common types of variation in the human genome are changes to individual nucleotides, or letters, which contribute to differences between people including disease risk; some rare single letter changes can directly cause disease.
The AlphaGenome Atlas charts the effects of 9 billion single DNA letter changes to the human genome â every possible mutation â using predictions generated by the AlphaGenome AI model released by DeepMind in London last year. It is freely available for non-commercial use.
The atlas could help to diagnose rare, unexplained diseases and uncover the hidden biology of common illnesses and biological traits, say researchers. It might even reveal some of the hidden rules by which DNA sequences control gene activity.
But it wonât replace experiments or, in the case of diagnosing disease, accounting for details of individual cases, says Martin Kircher, a bioinformatician at the Max DelbrĂŒck Centre for Molecular Medicine in Berlin. âThis is a useful and generous way to scale up access to a strong model.â
Since AlphaGenomeâs release, around 9,000 researchers have accessed the modelâs predictions through an automated programming interface (API), says Dhavi Hariharan, a DeepMind product manager. But doing so requires writing software code â a barrier for some biologists, she says.
To create the AlphaGenome Atlas, DeepMind computed predictions for each of the 3 possible nucleotide changes across every DNA letter in the human genome â 1 petabyteâs worth of data. It also captures more than 100 million short insertions or deletions observed in human genomes. The effort was inspired by DeepMind’s AlphaFold database of more than 200 million protein-structure predictions, which has been accessed by millions of users, according to DeepMind.
«
Stunning. Bioinformatics has struggled with data overload for years; maybe LLMs are finally the tool to make them useable.
unique link to this extract
Moto wants in: a wide Razr foldable just had its design leaked âą Android Central
Nickolas Diaz:
»
Wide foldables are quickly becoming a popular trend. Leaks from over the weekend claim Motorola could be the next brand to hop on that train.
Leaks revealed by Android Headlines, courtesy of an “anonymous source,” suggest the design Motorola is going for. Its source allegedly provided a few sketches, which purportedly showcase the various design ideas for the wide Razr foldable. One thing rumors claim is that Motorola’s foldable could feature an “easy-open” profile. Among theories of a crease-less display and “3D glass,” the wide Razr foldable’s frame “isn’t flat.”
The report states this decision could help make the device easier for consumers to open on the fly. This wide Razr Fold certainly fits the bill in these leaked illustrations. Offering that shorter (perhaps even stout) frame with rounded corners. “Motorola” will allegedly be inscribed on the hinge, visible when folded.
«
Oh look everyone’s piling into foldables, even though they’re absurdly expensive and will only appeal to 1-2% of the market. Anyone would think that Apple was about to validate the form factor by releasing its own.
unique link to this extract
AI is ushering in an era of mass toe-treading at work âą Financial Times
Sarah O’Connor:
»
When researchers Jana Retkowsky, Ella Hafermalz and Marleen Huysman conducted an ethnographic study into an advertising team in a Dutch media company, for example, they found that generative AI had changed the workflow profoundly.
Before AI, a senior creative would brainstorm ideas and come up with some rough concepts for clients to choose from. After that, a number of different professionals, from graphic designers to photographers, would be involved in the collaborative and creative to-and-fro which eventually resulted in the finished product.
After AI, the senior creative could use Midjourney to create a seemingly fully formed, beautifully polished concept right at the beginning of the process. The other professionalsâ jobs were now reduced to trying to execute the AI-generated pitch.
Retkowsky told me it was a tricky dynamic for colleagues to discuss, but that those conversations could lead to new ways of doing things that kept everyone involved.
âIt has to be talked about, which is awkward, since you are effectively telling a colleague they have taken over parts of your work,â she said. âThat is harder still when the person taking it over is excited about it. The creative directors loved the new autonomy of being able to visualise their own ideas, and they were not encroaching on anyone deliberately.â But, she said, teams that avoid the awkward conversation could âend up pushing people outâ.
Another piece of forthcoming research, this time on video games companies, finds a similar phenomenon of âoverlapping rolesâ in which AI is enabling artists to code and programmers to make artwork. Damian Grimshaw of Kingâs College London, one of the authors, said that while people had become âvery obsessedâ with the impact of AI on the vertical lines between juniors and seniors, he thought it âequally important just how much overlapping there is going on now which is horizontalâ.
«
AI is potentially turning people into jacks of all trades, masters of none. Which is a problem when you do actually need some experts.
unique link to this extract
Your perception of time is, in fact, warped âą Noema
Mike Mariani:
»
Our memories are the key to the central irony of our supersaturated world. Roseboomâs theory posits that time spent in an environment overflowing with visual information should feel long and robust. âYou would think that because weâre paying more attention to more things, that retrospectively time should feel longer,â Jones noted. âBut I think weâve got the worst of all worlds.â Because while we encounter a glittering constellation of apps, games and videos beckoning us at every turn, eager to light up our sensory cortex, theyâre not generating the indelible experiences crucial to forming memories. âWhen you get to the end of the day, you look back and thereâs no meaningful content there. Itâs just the same blur, the same stuff, and thereâs nothing distinctive about it.â
Wittmann agrees that toggling between apps on our phone and numbing ourselves with Instagram reels can impoverish our retrospective perception of time. âWhen weâre on digital media, time passes very quickly,â he explained. âYou lose yourself. Youâre jumping around like mad in these apps, and you canât put them into memory and later recall them, because the experiences just vanish.âÂ
Browsing, scrolling and swiping are more likely to quiet or even anesthetize our brains than activate them in the ways required to form memories. As a result, spending an hour on TikTok might feel like 10 or 15 minutes, because we donât have any of the âintricate, multitudinous, and long-drawn-outâ sensations that give time depth and fullness.
Instead, it all blends together in a turbulent vortex, overwhelming the senses rather than engaging them on the level necessary to endure over time. Itâs easy to extrapolate this further and speculate how spending several hours a day on social media distorts our sense of time, giving us the impression that we were awake for just 12 or 13 hours of the day. Over months and years, surrendering this much time to our ersatz digital warrens can lead to a sense of deprivation â a feeling that years have been carved out of our lives.
«
“The Starlink of desalination” âą The Rational Optimist
Stephen McBride:
»
Getting fast internet used to mean someone running a cable to your house. Fine in Manhattan. But a farmhouse in Montana was stuck with dial-up speeds. Too expensive to run a cable to one house.
Starlink leapfrogged cable by putting the network in the sky.
Vital Lyfe is running that playbook to build the âStarlink of water.â Itâs putting a billion-dollar [desalination] plant into a $749 box called Access.
Access weighs 25 pounds (11kg) and looks like a beer cooler. I lifted it one-handed. You drop a hose in the ocean, press a button, and get six gallons of clean, fresh water every hour.
Itâs designed to pass what Jon calls the âTerry test.â
âMy momâs name is Terry,â Jon told me. âI love my mom. Sheâs awesome, but sheâs not the most technically savvy.â
Terry will plug it in wrong, drop it, leave it in the sun, and try to use it before reading the instructions. Every design decision gets run past Terry. I can report I passed the Terry test. I opened the lid, pressed a button, and Access started pulling seawater out of a tub and pushing fresh water out the other side. We drank it, and it tasted fine.
Vital Lyfe will start shipping thousands of desal units in October.
You can tell these guys spent years at SpaceX. Ask Jon the most important thing he learned there, and he rifles off Elonâs five-step algorithm for making stuff fast and cheap.
The coolest thing he showed me was the membrane. Did you know membrane manufacturing is a cozy oligopoly controlled by five companies? They charge $60 apiece. Vital Lyfe makes its own for a fraction of the cost.
«
Although: what’s the point of personal desalination plants, unless your only water source is saltwater? Vital Lyfe thinks its customers are fishing villages and small islands. Could be? But also could be hard to hit those markets.
unique link to this extract
| âą Why do social networks drive us a little mad? âą Why does angry content seem to dominate what we see? âą How much of a role do algorithms play in affecting what we see and do online? âą What can we do about it? âą Did Facebook have any inkling of what was coming in Myanmar in 2016? Read Social Warming, my latest book, and find answers – and more. |
Errata, corrigenda and ai no corrida: none notified
The Overspill: when there's more that I want to say 