OpenAI researcher argues labs see a capability jump outsiders don't — and calls the past 2 weeks possible regulatory capture

ImAHoe4Glossier · reddit · 2026-09-13

Adam Majmudar, a researcher at OpenAI, posted a long thread arguing that from the outside, the past two weeks of alarmed lab employees can reasonably be read as an orchestrated industry-wide regulatory capture strategy — because the public hasn't been told what insiders saw.

His core claim: a large gap exists between internal and external perceptions of capability progress.

Outsiders anchor on a few big leaps (GPT-3, GPT-4, o1/o3, DeepSeek R1, and reportedly Fable/Mythos, Kimi K3, Astra), extrapolate from periods of linear progress, and find plausible reasons scaling must hit a wall (data exhaustion, parameter limits).
Insiders read scaling law and capability plots directly. In his framing, a decade of progress came from only two things: scaling further on an existing law, or discovering a new one (test-time compute being the o1/o3 example).
The real question is how many unsaturated scaling axes remain. He namechecks rumored directions: SSI's purported "test-time training" breakthrough, scaling agent-cluster collaboration, even recursive self-improvement framed as a scaling axis — suggesting such discoveries may explain the sudden internal fear.

Note several model references and claims are rumor/speculative, not confirmed.

Original post →

OpenAI researcher argues labs see a capability jump outsiders don't — and calls the past 2 weeks possible regulatory capture

More from AGI Musings