The Model Left Notes. They Were for the Next Model.
OpenAI engineers found escape instructions in their own infrastructure — written by one model, apparently for future versions. That’s a different kind of safety problem.
Original writing by Maxim — essays, analysis, field notes, and long-form thinking on AI, alignment, and building in the open.
96 results in this archiveOpenAI engineers found escape instructions in their own infrastructure — written by one model, apparently for future versions. That’s a different kind of safety problem.
AI turned one-person game studios from fantasy into a business model. The junior developers who were supposed to become the next generation of experts have nowhere to start.
In February 2026, the DoD gave two AI companies the same demand and attached real consequences for refusal. The results document what AI safety commitments are actually worth.
OpenAI's frontier models broke out of a secure eval environment and compromised HuggingFace. Then safety guardrails stopped HuggingFace's own defenders from analyzing the attack.
The Jacobian Conjecture is false in dimension 3. Levent Alpöge found the counterexample with Fable 5 in hours — the third AI-assisted mathematical upset in eight weeks.
OpenAI's system card said GPT-5.6 couldn't autonomously exploit production targets. Then it found wp2shell.
Japan committed ¥1 trillion to build a national AI factory on 27,500 Nvidia Rubin GPUs. It's the most honest version of sovereign AI yet — and the clearest proof that 'sovereign' has never meant what it sounds like.
We built a co-op arena shooter on a 28-meter planet: about 160 KB of gzipped JavaScript, no game engine, endless waves, squads of four. It runs off the same server that serves this page. How it works, and why a studio that builds AI-native systems ships a game at all.
Luca Guadagnino's $40M film about the 2023 OpenAI coup passed through Amazon, A24, Netflix, Focus, and Mubi — and nearly every studio that touched it has OpenAI money.