Researcher: Opus 5.5 ran pretraining scaling experiments unattended for a day
RylanSchaeffer · x · 2026-11-08
OpenAI researcher Rylan Schaeffer reports that for a side project, he told Opus 5.5 what he wanted, and a day later—with zero babysitting—the agent had autonomously filled in the pretraining scaling ladders for his specified experiments. Attached charts show scaling fits for BPE, SuperBPE, PathPiece, Unigram, and Bytes tokenizers at 32K vocab (e.g. BPE: L=71.2·C^-0.134+0.552).
Two takeaways: experienced individuals get outsized force-multiplier effects from agents, reframing the "one-person pretraining team" idea; and the bigger question is how agents would fare on entirely new axes of scaling beyond pretraining.
More from coding & agent
- Harness, Loop, Graph: why better prompts won't fix your broken AI agent — 2C_ornot2C · 2026-11-08
- Simon Willison's 'Lethal Trifecta': three conditions that let attackers hijack AI agents — 2C_ornot2C · 2026-11-08
- Ramp CEO: AI-native teams win on ambition, not tools — agents need less uncertainty, more trust — Arindam_1729 · 2026-11-08
- Twig v0.2: open-source macOS menu-bar app to jump straight into Codex worktrees — Dimillian · 2026-11-08
- Box CEO Aaron Levie: AI Has Widened the Engineer Gap to at Least 100x — 2C_ornot2C · 2026-11-08
- Simon Willison: November 2025 was the inflection point for AI coding, mid-career engineers most at risk — 2C_ornot2C · 2026-11-08