Jack Clark revisits GPT-2's staged release five years on: it was about setting a norm of caution
gleech · x · 2026-10-28
In Import AI #375, Anthropic co-founder Jack Clark offers a cold-eyed retrospective of OpenAI's 2019 decision to release GPT-2 in stages. Guided by early Scaling Laws work, the team broke from the standard publish-and-release process and instead announced the model via blog post first — an experiment meant to set an industry norm of caution around powerful AI.
Clark argues the decision's value lay in establishing that norm, and that whatever optics damage came from misunderstandings, critics shouldn't judge it with seven years of hindsight. The piece is his response to ongoing X debates over whether the GPT-2 precedent was worth it, offering a first-hand view of early AI safety governance.
Related event: Seven years on, GPT-2's staged release still sparks debate(3 posts)→
More from AGI Musings
- Math is not empirical; intelligence should target cheaper experimental signals — shlomifruchter · 2026-10-28
- Dan Faggella: life is inherently self-overcoming, so expansive AGI shouldn't surprise us — danfaggella · 2026-10-28
- Stewart Alsop: The Singularity Already Happened in the 1850s, and LLM Autonomy May Not Be Near — StewartalsopIII · 2026-10-28
- Dario Amodei must answer four pages of questions before Anthropic's IPO — Hesamation · 2026-10-28
- Big Tech fired hundreds of thousands, then hired again — was AI really the reason? — TansuYegen · 2026-10-28
- Mark Cuban: Run Every Benefits Contract Through AI to Find Where You're Getting Ripped Off — a16z Podcast · 2026-10-28