OpenAI's Former Safety Lead: We Now Ship 'New Capability and Risk Every Tuesday'

sourdub · reddit · 2026-10-08

Former OpenAI safety lead David Robinson told Ezra Klein that reasoning updates, new tools, and coding agents now let OpenAI ship new AI risk weekly.

Key points:

Previously each new generation started with a massive pretraining run, followed by post-training and safety testing; OpenAI once touted "months" of safety work on GPT-4 after the model was done.
The base model is now just one layer: pretraining is "the baking," while reasoning training and other post-training steps can be redone quickly—a better reasoning recipe can be layered on without retraining from scratch.
Release gaps are collapsing: GPT-6.1 Sol arrived at DevDay just a week after GPT-6 Sol.
"It's not just a chat anymore"—tools and affordances wired into models change capabilities and risks without any new training run.

His warning: safety evaluation cadence hasn't kept up, and risk now compounds on a weekly basis.

More from Companies & People