Codex hits 7M users, GPT-5.6 lands, and math still beats the frontier
A busy midweek: OpenAI's coding agent crossed a growth milestone, a new model tier arrived with commentary from Zvi, and a fresh math benchmark reminded everyone the frontier still has hard ceilings. Here's what moved.
OpenAI Codex reaches 7M users with 10x growth in 6 months, potentially overtaking Claude Code
10x growth to 7M users in six months makes Codex a real market signal — and the absence of comparable Anthropic numbers is itself a data point.
Zvi Mowshowitz commentary on OpenAI GPT-5.6-Sol, Terra, and Luna releases
Three new models across different price/capability points signals OpenAI is actively tiering its lineup — worth tracking for what it implies about inference economics.
AdvancedMathBench: New benchmark suite for advanced mathematical proof generation and verification
Frontier models topping out at ~75% on undergrad math and struggling to verify proofs shows advanced reasoning is still a genuinely open problem.
Mechanistic interpretability study reveals geometric structure of LLM-as-judge bias in hidden states
Finding that judge bias lives in predictable low-dimensional subspaces gives practitioners a concrete handle on when and why LLM-as-judge evals go wrong.
Requential coding: self-generated training data enables orders-of-magnitude model compression and tighter generalization bounds
Generalization bounds that tighten with scale — rather than loosen — challenge a long-standing assumption about why big models generalize at all.
Theoretical framework reveals low-dimensional invariant manifolds governing Transformer inductive reasoning dynamics
Reducing circuit formation to a low-dimensional manifold makes mechanistic interpretability more tractable — a useful theoretical foothold for the field.
Get the next edition in your inbox
Midweek and Sunday. Primary-sourced AI news paired with the guides that explain it.