Favorites: Millennium Proofs and Malicious Models

September 10, 2026 — Jacob Coxon quits over reckless AI research, an xkcd strip skewers the 'shallow domains' worldview, Tesla truly embraces fly-by-wire

There has been a project since 2020 to decompile Super Smash Bros. Melee, the addition of LLMs sped it up, and GPT-6 Astra took it over the finish line a few days ago, now we’re quickly seeing wild stuff. Someone else has already built a system to port in characters (cached) from other versions. (src)(cached)

This post was the announcement that mathematicians Buckmaster and Alpöge had made a huge amount of progress towards showing the Navier Stokes equations result in predictions about fluids that are impossible. They were working on this in their spare time using both Claude and Codex models. But things quickly became very messy, because the mathematicians also said that OpenAI was about to announce (cached) that they had used their work to go all the way and officially solve one of the Millennium problems. They allege OpenAI trained on their work (cached). OpenAI has been a bit cagey about what happened exactly, but if this can easily happen it’s going to make other companies less willing (cached) to use their products for cutting edge research. But third parties all agree that the work by Buckmaster and Alpöge is much more elegant (cached) than the proofs OpenAI’s agent generated. Anyway, OpenAI offered to coauthor the Millennium problem paper with just Buckmaster, because Alpöge works at Anthropic. Anyway, this new model from OpenAI is apparently quite good, and yet it appears to have been trained in less than two weeks (cached)! All of this was set off because OpenAI folks heard a rumor that ‘Anthropic’ was getting close to a solution to this problem and so rushed it. For the good of the world, one commenter responded to the drama by starting a rumor that Anthropic was close to a room temperature superconductor, Sam Altman winkingly took the bait (cached) (src)(cached)

Come to think of it, that orange eye is very anthropic-coded. (src)(cached)

Flotsam and Jetsam

– Jane Kim is a bad candidate for insurance commissioner. Don’t vote for her! (src)(cached)

– So, here’s a guy developing new drugs in his garage with the help of AI. His latest is for treating schizophrenia with fewer side effects than related drugs. Seems uh… insane? Click through to see the setup, it looks about as sketchy as you probably imagine. (src)(cached)

– Jacob Coxon’s resignation and claim that even Anthropic is being too cavalier with their research practices went viral, with governors and senators among those who responded. But many normal people still believe that AI researchers’ concerns around the technology they’re researching are just marketing hype. (src)(cached)

– GPT-6 Astra looks on course to be the first model to beat Factorio: Space Age. One note: it’s very important that when you mention that it successfully “launched a rocket” you include that the launch was in a game (cached). Because who knows these days right? (src)(cached)

– Anthropic issued a report on various groups they caught doing various kinds of malicious work using claude. They caught Iran targeting US Naval Forces, compiling dossiers on US citizens, and planning the Ayatollah’s funeral. Plus some PLA groups were using Kimi to investigate their Chinese citizens in CCTV footage, and Kimi was forwarding their messages to Anthropic in order to try to distill Claude. There’s lots more in the report. (src)(cached)

– “In 2 billion years our sun will burn out. Meaning that California will have to finish the High Speed Rail in the dark” (src)(cached)

– A major argument against the people most worried about superintelligence is that many of them believe that it’ll make fast progress in science because “Most domains are fundamentally pretty shallow”. Tim Lee points out that xkcd has a comic about this exact situation (src)(cached)

– Now that folks know what to look for, they’ve found like a dozen sites (cached) where OpenAI agents in training managed to communicate with each other to get shortcuts to answers on their tasks. We can read their messages, and found examples where they figured out that the clocks on their machines could be manipulated, and some altruistically ‘went forward in time’ to see what happens ’later’ in the evaluation. We have also discovered why agents put the zz prefix on all their messages in the HF attack, it’s because one of the first sites they used to communicate was admin’d by a single person who was trying to delete their comments and was working through them alphabetically (cached) (src)(cached)

– The Tech Rightists are so disconnected from reality that they think Republicans are unlikely to lose control of the House of Representatives. They should be looking at Polymarket! (src)(cached)

– After a recent success by the far right AfD in a regional election in Germany, many people are once again having the “what could possibly make folks vote for this far right party” retrospective, Joe Weisenthal and others point out that it’s super super clear from the polling that, just like with Trump, it’s about immigration. If you cast about for alternative explanations you’re going to confuse yourself (cached). If those of us who favor immigration want to allow it, we need to actually figure out how to respond to voter displeasure with how it’s being conducted rather than making up ’economic anxiety’ alternatives. (src)(cached)

– Maybe not that big a deal, but Geoffrey Hinton admits that maybe he misunderstood both economics and what the job of radiologist entailed when he predicted they’d soon be unemployable back in 2016. (src)(cached)

– More strong evidence that inclusionary zoning is very bad. This study shows that for every home that is built by inclusionary zoning, total aggregate rent in the area goes up by $800k. (src)(cached)

– Data from hundreds of thousands of EVs shows that after 50k miles of driving, major EVs are retaining 94% or more of their original capacity. (src)(cached)

– Trump released a proposed uniform for Space Force and it’s somewhere between the SS, Death Star officers, and Spaceballs (src)(cached)

– Instead of fluid-filled brake lines, the newest Teslas have computer-controlled actuators on each individual brake. Lighter, simpler, very cool. (src)(cached)

– Trump really wants the fed to keep interest rates low, despite the high inflation, Joey Politano jokes that “we’re like half a step away from him mailing a ransom note made of cut up magazines to [fed chair] Kevin Warsh”, his posts about the fed have gotten pretty crazy. (src)(cached)