Yoav Goldberg coins 'harness distillation': agent scaffolding skills will be absorbed into models
yoavgo · x · 2026-10-04
In the harness-vs-model debate, NLP researcher Yoav Goldberg coins the term 'harness distillation': capabilities start in the harness, and once they work well, the next model generation is trained on harness-assisted outputs and trajectories, internalizing the scaffolding. He adds that real harnesses may still beat distilled models on edge cases, but the gap mostly shows in rare scenarios—not worth the operational cost for common cases. Well-funded domains will converge; niche domains will keep dedicated harnesses.
Related event: AI Researchers Argue Harness Capabilities Will Be Internalized by Models(2 posts)→
More from coding & agent
- Replit CEO Amjad Masad: general models should JIT-train their own smaller replacements — amasad · 2026-10-04
- Startup claims it will 'kill a $1T industry'; agent dev Jason Kneen mocks the approach — jasonkneen · 2026-10-04
- Running Claude-dev + Codex-review loop, but can't wire Gemini Pro in as reviewer — bocondo · 2026-10-04
- New edition of 'An Introduction to Programming Languages' adds browser runtime and AI chapters — alfcnz · 2026-10-04
- pg-jev: an open-source Postgres extension for querying rows in plain language — Born_Excuse_5610 · 2026-10-04
- New research prototype tests whether defensive mechanisms can stop autonomous web agents — Admin-ABC-XYZ · 2026-10-04