Astra tops ErdosBench as OpenAI skips math on purpose

OpenAI had put math wins front and center in its first announcement, which makes Pachocki's admission in "An Alien Mind" notable: the company believes it could make its models better at math research with additional focus, but chooses not to, citing urgency around recursive self-improvement and automated alignment research. That trade-off is the article's central evidence for what it calls the "spiky" thesis of AI development — extreme strength in select domains rather than broad, gradual improvement across all tasks.

Cambridge researcher Adam Hunt's visualization contrasts the "mainstream AGI" path of broad gradual improvement with an increasingly spiky trajectory, where coding and math surge while language quality, common sense, or social reasoning stagnate or regress. The article presents Pachocki's statement as evidence for the latter, noting that even the leading AI lab can't push maximum progress in every direction at once. Behind the RSI priority, however, is the hope that the model will eventually make those optimizations itself, scaling faster across the board — including in math.

Whether Astra is 5 or 50 percent better than its predecessor, the deeper question for mathematics remains: as Terence Tao raised at the 2026 International Congress of Mathematicians, if AI models keep producing proofs faster than humans can check them, the field risks shifting from proof scarcity to proof overload, and the critical task becomes deciding which results actually matter. The hardest problems remain unsolved for now, which may buy the discipline some time, but the direction still seems set — even if some mathematicians doubt language models can deliver real breakthroughs without human-like creativity.