Stopping new models won't stop AI progress: today's models with tomorrow's harnesses
wfithian · x · 2026-10-31
In a discussion with David Duvenaud and teortaxesTex, wfithian argues AI will keep improving even after we stop training new models: "we've not seen how capable today's models will be with tomorrow's harnesses, coordination, etc." — the same brains got cavemen to the moon. Duvenaud notes test-time compute scaling is a related channel, and gradient-free tuning of contexts and scaffolds would still enable a "poor-man's RL," just far slower. wfithian agrees.
More from AGI Musings
- Why people overpraise LLMs: they mistake systematicity for intelligence — fkasummer · 2026-10-31
- AI turned everyone into a manager overnight, and most have never managed anyone — tech__unicorn · 2026-10-31
- AI commentator questions why AI CEOs use "you'll own nothing" rhetoric — qtnx_ · 2026-10-31
- Santa Fe Institute publishes interactive history of agent-based modeling — round · 2026-10-31
- Podcast: poet Sasha Stiles on AI poetry, NFTs and transhumanism — Merzmensch · 2026-10-31
- NVIDIA researcher: AI risks come from human steering, not rogue AI — Jsevillamol · 2026-10-31