Top AI labs have stopped publishing research

A thread on AI's top startups barely publishing their research is generating resigned but pointed commentary. The most-upvoted observation: none of these companies would exist without Google publishing 'Attention Is All You Need.' The irony is not lost on anyone. A commenter noted the shift is 'driven by greed' though labs will frame it as safety. Another pushed back on the safety framing directly, arguing that the race is not toward cancer-curing AI anyway, so the urgency used to justify secrecy is hollow.

Anthropics's cryptanalysis thread adds a wrinkle. Claude Mythos, their most advanced model, produced novel cryptanalysis results and the model itself will not be released publicly. Commenters are parsing whether 'unreleased' means 'never' or 'not yet.' The distinction matters a lot for researchers trying to reproduce or build on the work.

The pattern is a steady closure of the research commons that bootstrapped the field. For founders building on top of these models, this matters practically: fewer published architectures means less ability to understand what you're building on, predict capability jumps, or catch failure modes before they hit production.