MIT prof: model capabilities follow a scaling law, with a threshold for open math discovery already crossed

ProfBuehlerMIT · x · 2026-11-23

MIT professor Markus Buehler highlights @merettm's work showing broader categories of model capability follow scaling laws: around 0.76 log-perplexity on OpenAI's benchmark, models began producing results on extremely complex open math problems.

He argues we're entering a new scaling regime that may yield qualitatively different open-ended discovery capability. Justin Bullock adds that "a threshold in what these models can do seems to have been crossed," and expansions into lower tiers will likely be even more consequential — though what comes next remains unknown.

Original post →

More from Models

Models channel →