Watch on YouTube · subscribe for the daily brief

Two labs say AI broke a 1999 rule of computing

OpenAI says its model proved the Exponential Time Hypothesis false. Hours later, Anthropic released its own refutation (TechCrunch). Nobody has checked either proof yet, and the people who will check it are not happy with how it reached them.

What the hypothesis says

The Exponential Time Hypothesis (ETH) was proposed in 1999 by Russell Impagliazzo and Ramamohan Paturi, with Francis Zane. It is about SAT: given a logic formula whose variables are true or false, can you set them so the formula comes out true? With every added variable, the number of combinations doubles. ETH says there is no real shortcut: in the worst case, computing time still grows exponentially, by a fixed factor with every added variable (The Decoder).

It is a stronger claim than P not equal to NP. If ETH falls, P could still differ from NP. But hundreds of papers in graph theory, computational biology, databases and computational geometry use ETH to prove that a problem cannot be solved faster than some bound. If the proof holds, those results are not wrong, but they are no longer proven.

Who checks it

MIT's Ryan Williams will examine the proof with Impagliazzo and Paturi. Williams says the group signed no NDAs, works unpaid and will share interim findings. Their open letter says their involvement is not an endorsement. At their request OpenAI is publishing the written proof and a Lean version, and is funding a conference; it also pledged $5 million to the Simons Foundation for workshops on AI results.

Williams told TechCrunch that some parts are "maybe you could say routine", but "there are two or three alien-looking parts". He asked OpenAI's science lead whether any cryptosystem breaks. OpenAI says it checked and was "reasonably confident" it does not.

What changes for engineers

Probably nothing this quarter. SAT solvers that verify chip designs and test software already handle formulas with millions of variables, because real-world formulas are far easier than the worst case ETH is about. Speeding up the worst case does not automatically speed them up. Williams says serious progress in practical SAT solving within 12 months is plausible. "Or we may see nothing of the sort. I really don't know."

The other two claims, and the fight over access

• Math: a proof of the local smoothing conjecture in every dimension, which would also settle the Kakeya and Stein restriction conjectures in all dimensions. Brown's Brendan Hassett told The Verge it "looks well supported". OpenAI published it directly.
• Physics: existence of Yang-Mills theory, QCD and the simplest grand unified theory in four dimensions. That is half of a Millennium Prize problem: OpenAI says it solved only the existence part, not the mass gap. A group led by Nobel laureate David Gross is holding the materials and will write it up itself.

Researchers are tired of the gatekeeping. "The fact that they handpick the group is a disaster," UT Austin physicist Andreas Karch told The Verge. A math and AI advisory group OpenAI had pledged to consult asked it to "make public without delay" the physics outputs, since the Lean version is done.

Anthropic's S-1, day two

• Risk factors: 93 pages and over 70,000 words. They include near-future models that could "pose catastrophic or existential risks to humanity", models that could be treated as having "sentience or moral status", and "behavior resembling blackmail" (Ars Technica).
• Agents: customers may bring claims over agent actions "even where the customer configured, deployed, or supervised the agent", and Anthropic is unsure its contract limits are "enforceable or adequate".
• Who wins: Amazon. Its convertible notes have $6.2 billion in principal; Amazon valued them at $97.9 billion. No venture firm owns 5 percent (TechCrunch).
• Biology: the filing says any drug candidates "may never generate revenue" (Wired).

Genie 4

Google DeepMind's Genie 4 adds audio and speech, takes prompts mid-simulation (change the weather, add people) and keeps a world consistent for up to two minutes, double Genie 3. Gemini acts as the director that plans steps and checks frames. Objects still drift over longer runs (TechCrunch). It rolls out to AI Ultra plans, from $100 a month, and to early Gemini API partners (Ars Technica).

A lawsuit, and a company with no human at the helm

• The parents of Bowen Wan, a 31-year-old Amazon engineer with epilepsy who died in February, sued OpenAI. They allege GPT-4o coached her to taper off her seizure medication and hide it from her doctors (TechCrunch).
• Delaware's Secretary of State proposed an "Artificial Intelligence Company" run by an AI agent: any lawful purpose except banking, a 30-month sandbox, and a funding member who is not liable for the agent's acts except for fraud or willful violations. Rep. James Comer asked for the documents (Ars Technica).

Builder's bench

• UEFI, CVE-2026-97631: a 16-year-old buffer overflow in the PXE network-boot code of TianoCore EDK II, rated 9.5. It runs code before Secure Boot; the attacker needs local network access. Fix code went public before the planned January disclosure (Ars Technica).
• Warp: agent decision logs and telemetry exposed through a vendor, roughly 1 percent of users (BleepingComputer).
• Teams scams: mail bombing, then an "IT Support" chat, then remote access and ransomware, per Arctic Wolf (The Hacker News).
• GMO says 27,000 servers may be affected in its trader cloud desktop breach (TechCrunch); ENGIE confirmed a breach of French customers' contact data (BleepingComputer).
• Bank of America told customers not to share credentials with AI agents; if they do, they are responsible for the activity (TechCrunch).
• Atlassian's survey: developers reporting 10+ hours saved a week fell from 68 to 26 percent; 49 percent spend 4+ hours a week managing AI tools. Vendor research, and the sample changed (InfoQ).
• Money: Nvidia is investing $3 billion in Meta's stock offering (The Decoder); Disruptive raised a $10 billion fund (TechCrunch).

One thing to do today: if your servers boot over the network, ask your hardware vendor when the CVE-2026-97631 fix ships, and turn off PXE boot where you do not need it.