Researchers debate feasibility of compute-governance policy for durable pause on frontier AI training
wfithian · x · 2026-10-31
David Duvenaud, wfithian, alextmallen and others debate a compute-governance policy paper on X:
- The policy targets a durable pause on frontier training, not a mere slowdown; with partial enforcement, systems could still improve slowly via scaffolding gains.
- alextmallen raises a key objection: can inference substitute for training under a ban, given RL training looks largely like inference?
- MaskedTorah argues a true AI slowdown is infeasible — only a strong pause works — since "models get better but slower" would be brutal if developers must fabricate new chips and discard old ones every release.
- wfithian adds that cutting accelerator GPU production could be a standalone pacing measure, though the paper doesn't explore it.
More from AGI Musings
- The 'thesaurus on steroids': an underexplored interaction pattern for generative models — chrissyykat · 2026-11-02
- rauchg: Own Your Intelligence — Embrace Open Source and Local Models — mitsuhiko · 2026-11-02
- rakyll: No Moat Up, Down, or in the Middle of the AI Stack — rakyll · 2026-11-02
- Engineer onboards ad director and surgeon onto Codex — their use cases surprise even her — enggirlfriend · 2026-11-02
- German flagship daily's AI coverage slammed: labs 'anti-democratic,' Meta a thief — kimmonismus · 2026-11-02
- Researchers wonder: is OpenAI using its strongest models to search for a post-Transformer architecture? — SonglinYang4 · 2026-11-02