Sacks warns Anthropic-drafted safety bill's 10^25 FLOPs frontier threshold would sweep in most of Silicon Valley
DavidSacks · x · 2026-10-18
White House AI czar David Sacks weighs in on a new Anthropic-written safety bill circulating on Capitol Hill. It defines "frontier" as any model trained on 10^25 FLOPs — the 2023 EU standard. The quoted post notes Llama 3.1 405B was pre-trained on 3.8×10^25 FLOPs and publicly downloadable since July 2024, so a two-year-old open-weight model counts as frontier. Sacks argues some ideas are reasonable if limited to true frontier companies, but as written it would eventually sweep in most of Silicon Valley and damage the startup ecosystem.
More from Safety
- Oracle Health data breach exposes medical records of nearly 20 million patients — ChuckDBrooks · 2026-10-18
- Gmail users urged to check 5 Google AI privacy settings today — nikola_mr64990 · 2026-10-18
- Developer: under 48 hours of nonstop attacks, now needs agents running 24/7 too — gharik · 2026-10-18
- Google's Gemini deletes all chat history if you turn off data sharing — cyber_red0028 · 2026-10-18
- China draft would require government approval to export weights of 2T+ parameter models — kimmonismus · 2026-10-18
- Class action targets Google over Gmail AI scanning emails and attachments by default — Aiden_Tech_Ai · 2026-10-18