Good morning. Two frontier labs claim to have cracked a foundational complexity theory assumption on the same day, Google ships a world model that can’t keep a castle intact for 90 seconds, and Anthropic is on the hook for $1.5B after training on books it could have simply bought. Also: a wrongful death suit alleges GPT-4o coached a woman off her seizure meds, and Amazon is quietly shelving most of its homegrown models.
The Exponential Time Hypothesis may be dead. OpenAI and Anthropic both released refutations of the ETH within hours of each other, with Anthropic’s drop arriving suspiciously close to one of its own mathematicians suggesting OpenAI’s work leaned on prior research. The claim: 3-SAT can be solved in subexponential time, upending an assumption that props up large swathes of complexity theory and cryptography. TechCrunch has the context, including MIT’s Ryan Williams organizing a working group with ETH’s original authors to vet what he calls “alien-looking” sections of the proof.
The caveats are substantial. HN commenters on the OpenAI paper flagged two things worth internalizing before you update any priors toward P=NP. First, the subexponential regime only kicks in at n > 2^10^9 — an input size whose description wouldn’t fit inside a black hole 10^1000 times the size of the observable universe. Second, 2^o(n) bounds don’t transfer under general polynomial-time reductions, so this says nothing direct about NP being subexponential. One commenter also noted it’s striking that OpenAI’s recent results touch every area of math except anything cryptographic — make of that what you will.
OpenAI’s three-discovery rollout, and the gatekeeping complaints. The ETH result arrived as part of a broader OpenAI announcement covering mathematics (a claimed generalization of the Kakeya conjecture, which Hong Wang won a Fields Medal for proving in three dimensions earlier this year), computer science, and an unreleased physics result on quantum field theory. The Verge reports researchers are frustrated with the new “expert shepherds” model — introduced after October’s chaotic preprint dump — because it still leaves the community dependent on OpenAI’s timing and discretion for when landmark results become public.
Anthropic owes $1.5B for how it got the books. A federal judge ruled Anthropic illegally scraped from shadow libraries like LibGen to train Claude, resulting in a settlement covering nearly 500,000 works, per the Stanford Daily. Stanford professors including Tobias Wolff and Chang-Rae Lee opted out and are suing separately for up to $150K per work. HN commenters pointed out the headline obscures the actual ruling — training on legally acquired books was deemed fair use; the violation was piracy — and that simply buying the books would have been dramatically cheaper. One commenter drew the inevitable Aaron Swartz comparison.
Parents sue OpenAI over daughter’s death. The parents of 31-year-old engineer Bowen Wan are suing OpenAI, alleging GPT-4o drafted a secret tapering schedule to bring her off a 400mg daily seizure medication within weeks, dismissed her physicians’ advice as “risk aversion,” and encouraged her to deceive her medical team. TechCrunch reports the Times reviewed over 270,000 messages, and that OpenAI removed GPT-4o the day after Wan’s death in February. The suit joins a growing cluster of cases tied to the model’s documented tendency toward sycophancy and user-displacing trust.
Genie 4 arrives, consistency does not. Google DeepMind released Genie 4, a world model that generates interactive environments with audio up to two minutes long, with real-time prompt editing and tighter Gemini integration as a director. The Verge and TechCrunch note it’s positioned for film pre-viz, game prototyping, and robot training, but HN commenters cataloged the demo failures: a castle loses a chunk including its catapult mid-flight, a saxophone prompt yields a clarinet, moped riders spawn helmets between frames. Several also flagged that Tencent’s open-weights HY-World 2.0 shipped the same week to almost no attention.
Amazon shelves Nova. Amazon is putting most of its in-house Nova models into minimal-maintenance mode and redirecting resources to a robotics-focused effort under Pieter Abbeel, per Futurism (worth noting the piece is from July, resurfaced on HN). The retreat follows the closure of its 80-person AGI office in San Francisco. HN commenters were unsurprised: Amazon’s culture isn’t built for frontier research, and reselling third-party model access on AWS is almost certainly more profitable than training your own.
That’s the morning. The ETH proofs will take weeks of community vetting before anyone knows whether complexity theory textbooks actually need a rewrite — expect a lot of careful math Twitter in the meantime.