Revisiting 2019: OpenAI withheld GPT-2 as 'too dangerous to release'
olalonde · hn · 2026-09-13
- A 2019 TechCrunch article resurfaced on Hacker News: OpenAI's 1.5B-parameter GPT-2 was deemed capable enough to generate convincing fake news that the lab staged its release, withholding full weights over misuse concerns.
- The decision sparked a landmark debate on open-source vs. safety and became a classic case study in staged AI model releases.
More from Models
- 2B open-source MiniCPM5 tops small-model index, tested as an offline iPhone iMessage agent — SimplyAnnisa · 2026-09-13
- OpenAI pauses Pro subscriptions, leaving a Codex power user stranded after 10B tokens — andimarafioti · 2026-09-13
- InternLM Releases Intern-S2-397B: Multimodal Model for Scientific AI and Long-Horizon Agents — jacek2023 · 2026-09-13
- Writer Tests Own Articles on GPTZero: AI Hasn't Learned His Style Yet — michalmalewicz · 2026-09-13
- GPT-6 Astra Clears All 25 ARC-AGI-3 Games, Hitting ~80% of Optimal Play — i_dg23 · 2026-09-13
- OpenAI claims its model found a Navier-Stokes blowup counterexample, cracking a Millennium Problem — victor_explore · 2026-09-13