History of AI doomer warnings: from OpenAI withholding GPT-2 to Amodei's slowdown calls
AjdDavison · x · 2026-09-15
Reflecting on Dario Amodei's latest warnings that AI development may need to slow, the author traces the history: OpenAI withheld full GPT-2 in 2019 over misuse fears, then released it nine months later having found little evidence of harm; in 2023 Amodei told the US Senate that open-source AI was on a "very dangerous path." He agrees Amodei's concrete proposals (pacing capabilities, independent evaluator access) are the right instruments, but warns that repeatedly escalating catastrophe claims risks the public tuning out — so when genuinely compelling evidence of extreme danger appears, no one will listen.
More from AGI Musings
- Clarifying the AI safety split: safety testing time vs long internal deployment — JacquesThibs · 2026-09-15
- a16z partner puts P(abundance) at 99.99% — nptacek · 2026-09-15
- AI safety debate: external model releases may be safer, internal deployment the real risk — JacquesThibs · 2026-09-15
- Rival AI Labs Rally Behind Anthropic's Slowdown Proposal — The AI Daily Brief · 2026-09-15
- AI arms-race narrative misleads policy, says researcher: cooperation, not competition, is the only path — GregCook2011 · 2026-09-15
- One hour of vibe coding flips 'AI is useless' skeptics, from professors to execs — pwlot · 2026-09-15