Anthropic delayed pre-training over a month for security — but it won't slow AI, researcher argues
thlarsen · x · 2026-10-30
Anthropic's S-1 reveals it delayed pre-training for over a month for security reasons. Researcher thlarsen argues this accomplishes little: R&D continued at full speed, so algorithmic progress simply piles up and a smarter model ships later anyway — and the algorithms could be stolen regardless of whether a run started. He favors pacing proposals that bite on the production of new algorithms (the "can train" line) rather than the output of final training runs (the "has trained" line), while conceding that delaying runs matters more during an intelligence explosion, when earlier model access compounds algorithmic progress.
More from AGI Musings
- Diamandis: Measure the AI revolution by whether it makes homes, kids and healthcare affordable — PeterDiamandis · 2026-10-30
- Altman: OpenAI barely knows how to use agent swarms, answers in 3 months — 141_1337 · 2026-10-30
- Carmack: AI writing software from scratch isn't optimal — well-regularized building blocks matter — ID_AA_Carmack · 2026-10-30
- Economist Walks Through 'Unserious AI Use' in a Published Saez-Zucman-Adjacent Paper — anshulkundaje · 2026-10-30
- 'Agents Are the New Buyers': User Solves App Problem Without Knowing It Runs on Supabase — gabe_ragland · 2026-10-30
- "If everyone's convinced AI is a bubble, there's a good chance it isn't" — justalexoki · 2026-10-30