Good morning. GPT-5.6 was only the opening act — overnight, a variant of the same model appears to have produced a proof of the Cycle Double Cover Conjecture, a graph theory problem open since the 1970s. Meanwhile OpenAI’s head of safety walked out the door, Meta yanked an AI feature within days of launch, and a developer got a 744B-parameter model running on a home PC.
GPT-5.6 Sol Ultra reportedly cracks a decades-old conjecture. OpenAI published a PDF proof of the Cycle Double Cover Conjecture, apparently produced by GPT-5.6 Sol Ultra in about an hour. The proof is unusually short, which suggests either a clever shortcut experts had missed or a subtle error waiting to be found — expert verification is still pending. One HN commenter drew the obvious chess-engine parallel, while others fixated on the released prompt, most of which is spent begging the model to actually solve the problem instead of filing “vague optimism” status reports.
As we covered yesterday, GPT-5.6 shipped — the reviews are in. A day into hands-on testing, the community verdict on GPT-5.6 is that Sol genuinely leads on benchmarks (including that ARC-AGI-3 first), while Terra feels roughly comparable to GPT-5.5 and behind Sonnet 5 on real coding tasks. Someone also spotted that OpenAI excluded Fable 5 from its GeneBench and LifeSciBench comparisons because the model “refuses the majority of questions” — winning by default is one way to top a chart. The Sol/Terra/Luna naming continues to confuse basically everyone.
OpenAI’s head of safety is out. Johannes Heidecke is leaving following a reorg that folds safety systems under VP Mia Glaese, Wired reports, joining a string of safety-focused departures including chief futurist Joshua Achiam. The framing from OpenAI is that integration will speed things up; the framing from anyone counting exits is that GPT-5.6 reportedly showed misaligned behaviors at launch and the safety org is thinning at exactly the wrong moment.
Meta pulls its Instagram deepfake feature within days. The Muse Image tagging feature we flagged Monday — the one that let anyone generate AI images of any public Instagram account by @mentioning them — is gone. The Verge and TechCrunch both note the National Center on Sexual Exploitation, SAG-AFTRA, and CAA all condemned it before Meta admitted it “missed the mark.” Critics point out that opt-out-by-default was the actual problem, and Meta hasn’t addressed that part.
A 744B MoE model, streamed from disk. A developer released Colibri, a ~2,400-line dependency-free C inference engine that runs GLM-5.2’s 744B parameters on ~25GB of RAM by keeping dense layers in memory and streaming routed experts from an NVMe SSD. The HN thread is largely about whether tokens-per-second is usable enough to matter, with several noting llama.cpp already does mmap-based streaming — but the trajectory (RAID SSD arrays as a cheap RAM substitute for local inference) is what people are actually excited about.
AI videos tuned to hijack specific brain regions. EPFL researchers built a system that generates videos designed to maximally activate targeted brain areas, using a “digital twin” model that predicts neural response to visual stimuli. The stated purpose is reducing experimenter bias in neuroscience; the HN reaction went somewhere darker fast, with commenters noting that social platforms already A/B test videos for maximal engagement and this collapses the search space directly onto the brain. One researcher called it “the most dangerous and morally wrong” idea in their notebook.
Two from the money-and-metal side. SK Hynix raised $26.5B on Nasdaq, the largest foreign IPO in U.S. history, with demand seven times the shares on offer — Commerce Secretary Lutnick is already leaning on both SK Hynix and Samsung to build U.S. fabs, matching Micron’s $250B commitment. Separately, Hugging Face CEO Clem Delangue argued on TechCrunch’s podcast that Fortune 500 companies are increasingly done renting AI as scaling costs push them toward open weights, and noted Hugging Face turned down a large Nvidia investment to stay capital-efficient.
That’s a lot for one Friday. If the CDC proof holds up under review, we’ll be talking about it for a while — if it doesn’t, we’ll be talking about that instead.