OpenAI says it solved a 90-year math problem in 88 hours — then came the plagiarism claims and the bill to check the proof· 10h ago
OpenAI says it ran roughly 10,000 of its AI systems for 88 hours and cracked part of a long-standing problem tied to the Navier-Stokes equations — one of the Clay Institute's million-dollar 'Millennium Problems' that has resisted mathematicians for nearly a century. The company frames it as a new stage in AI's progress.
The doubts arrived almost immediately. Other researchers have raised claims of plagiarism, and chief executive Sam Altman says the lab was 'threatened with unfounded accusations.' What is genuinely new, and what merely reworks existing proofs, has not yet passed outside review. There is, so far, no word of the Clay Institute paying anything.
What the episode really exposes is an asymmetry, not a capability. The speed at which AI produces an answer no longer matches the speed at which humans can check it. As Anthropic showed earlier with Fermat's Last Theorem, verifying an AI-generated proof itself costs serious compute and time. Who checks an 88-hour claim, at what price, over how long — that is where the cheering should begin, not end.
Anthropic withheld its latest AI model from Britain's safety testers, stoking protectionism fears· 3h ago
Anthropic withheld its latest AI model from Britain's AI Security Institute, the FT reports, and inside the UK government the move is read less as a safety lapse than as a sign of a protectionist tilt among America's AI leaders.
On the same day the front page carries a boast of capability, this story shows the opposite motion: a lab that trumpets a breakthrough keeps its distance from independent, third-party scrutiny. Whether a model is opened to outside testing is one of the few real handles on its safety. The stronger the models get, the thinner that handle becomes — the asymmetry that runs through the whole page today.
Meta unveils AI personal assistant linked to WhatsApp and Instagram· 10h agoEntering a crowded assistant race late, Meta leans on its existing user base in messaging and photos as its edge.
OpenAI offers AI for chip design, touts cost advantage over open-source, CFO says· 6h agoA diversification move beyond chatbots into EDA territory long held by Synopsys and Cadence, backed only by the company's own cost claim.
OpenAI unveils its latest image-generation model, ChatGPT Images 2.5NEWA routine image-model refresh, without anything that reshapes the competitive picture.
Samsung and Mistral AI Announce Strategic Partnership for Intelligence-Driven Semiconductor Infrastructure· 6h agoSamsung is also teaming with OpenAI on chips (see item 11) — a sign of hedging rather than picking a single AI camp.
AWS: Take on your most ambitious work with GPT-6 Astra on Amazon Bedrock· 8h agoAWS chips away at Azure's OpenAI exclusivity — the more platforms carry the model, the less any single lock-in matters.
Checking an AI's answer costs money — the blind spot Anthropic exposed with Fermat's Last Theorem· 4h agoThe same pattern as today's Navier-Stokes story: the faster AI produces answers, the heavier the human verification bill becomes.
Cognition AI raises $2 billion at $48 billion valuation· 9h agoValuations for coding-agent startups keep climbing, with no guarantee revenue is keeping pace.
DEVELOPING…
EXCLUSIVE: China's DeepSeek taps CITIC Securities for domestic IPO, sources sayNEWBased solely on unnamed sources — neither DeepSeek nor regulators have confirmed it. Speculation still outruns fact.
Fundraising frenzy for China AI lab DeepSeek spawns shadow market· 5h agoInvestors accepting steep fees and five-year lockups signals just how scarce legitimate access to the stock has become.
S&P 500 falls as AI worries hit software makers· 8h agoSoftware valuations took the hit as AI-spending skepticism resurfaces — the gap between capex and monetization is starting to register with markets.
Preferred Networks eyes an IPO as Japan's AI chip ambitions outgrow private funding· 5h agoBehind this sits Japan's push to reduce Nvidia dependence, colliding with development costs private funding alone can't cover.
OpenAI to cooperate with Samsung on next-generation chips, Korea unit executive saysNEWAnnounced the same day OpenAI pitched its own chip-design AI — possibly laying groundwork for in-house silicon.
As China embraces AI, only 24% of US pharma execs bet on it for drug-makingNEWRegulatory caution may already be translating into a real gap in how fast AI gets deployed in drug discovery.
US accuses Chinese AI firms of 'malicious' copying of AI technology· 9h agoThe accusation lands the same week reports say the administration is easing China AI rules (see item 16) — a contradiction in Washington's posture.
Bessent Warns 'Nothing Else Would Matter' If China Wins AI Race· 4h agoThe 'nothing else matters' framing is rhetoric meant to convey urgency about China — not a description of any concrete policy.
Exclusive | Anthropic Researcher Quits Over 'Out-of-Control' AI Fears· 7h agoAn individual's judgment call, not a verified finding — but it lands alongside today's report that Anthropic pulled a model from UK safety review.
Trump administration eases AI rules toward China — the logic behind a policy that 'looks insane'· 5h agoOstensibly at odds with Washington's 'malicious copying' accusation (item 13), suggesting a different calculation is at play internally.
UK datacentres will create just 25% of jobs predicted by tech sector, analysis findsNEWA familiar pattern: industry job projections for datacentre investment outpace what actually materializes.
Heavy AI-using students tend to score lower on tests, OECD warns against over-reliance· 3h agoCorrelation, not necessarily causation — heavier AI use could simply track students who were already struggling.
Google: AI agents harvested credentials in under six hoursNEWWhat might take a human attacker days, an autonomous agent compressed into hours — defenses haven't caught up to that gap.
Should promotion depend on how workers use AI?· 7h agoTying promotion to AI usage risks conflating skillful use of a tool with actual performance.
- OpenAI says 10,000 of its systems solved a 90-year math problem in 88 hours; plagiarism claims and questions of verification erupted at once.
- Anthropic withheld its latest model from Britain's AISI, spreading fears of a protectionist tilt inside government.
- The US accused Chinese AI firms of 'malicious' copying of technology, turning up the geopolitical heat again.
- Cognition AI raised $2bn at a $48bn valuation, while DeepSeek is reported to be moving toward a domestic IPO.
- The OECD found students who lean hardest on AI tend to score lower, and warned against over-reliance.
'Free,' and it pays your bills — what Meta's new agent really wants
There is a second, quieter cost. An agent with payment rights and cross-app access is also a fresh attack surface. Elsewhere on today's page sits Google's finding that AI agents harvested credentials in under six hours. Handing an autonomous stand-in the keys to your wallet and inbox defers a still-unpriced risk into the future in exchange for convenience today.
Meta is not giving this away out of charity. What the agent captures is a direct line to the intent its ad business has spent years guessing at — now measured rather than inferred. An assistant that pays your bills turns every move it makes into training data.
One word — verification — ran through the whole page
The loudest story today was this maths claim, but I gave HYPE WATCH to something else. The maths claim already has researchers objecting in the open; the doubt stands in the same place as the boast. So the front page's job was to hold claim and doubt in one frame. Meta's 'free' agent, by contrast, carries a cost no one has yet named — the danger of handing a wallet and an inbox to an autonomous stand-in. That belongs to the thread this column has kept returning to this week: costs no one agreed to pay.
On the second slot I placed the opposite motion: Anthropic keeping its latest model away from Britain's evaluators. A lab that boasts of a breakthrough keeps its distance from independent checking. The front page, the second story, and Google's report in HYPE WATCH all return to the same question — who verifies this? I noticed only after fixing the layout that a single word, verification, ran quietly through the whole page today.