Toby Ord says RL post-training compute neared parity with pretraining; bet on broad compute-capability trends
ben_j_todd · x · 2026-10-14
Toby Ord revisits his claim that RL post-training compute had reached >10% of pretraining compute — with a dubious xAI chart even suggesting parity in their latest model. His October 2025 essay suggesting post-training compute was approaching parity drew skepticism, though the chart indeed shows RL nearing that mark that quarter.
Ben Todd adds that forecasts of when specific drivers run out almost always fail; it's better to bet on the broadest relationship between compute and capabilities.
Related event: Toby Ord: RL Post-Training Compute Now Matches Pretraining(2 posts)→
More from AGI Musings
- "Hitting the bullseye doesn't make you a good shooter": an AI proofs debate — gerardsans · 2026-10-14
- Humans on AI: possible power-seeking incidents, White House AI order, RL at 63% of Reflection's run — gleech · 2026-10-14
- Data center panel asks 'Are We Screwed?' as Gary Marcus says AI industry may have screwed itself — GaryMarcus · 2026-10-14
- "Everything is now a token": ML vocabulary is being rewritten by the LLM paradigm — jongranskog · 2026-10-14
- Distrusting experts passes as critical thinking, distrusting conspiracies as closed-mindedness — DanielWhiteson · 2026-10-14
- Gemini 4 agents crack open problem in plasma physics, new exact solutions within hours — Dr_Singularity · 2026-10-14