Pedro Domingos on harness learning paper: the model swallows the harness
DanielKhashabi · x · 2026-10-12
- Pedro Domingos quotes an arXiv paper, Harness Learning Enables Generalizable Test-Time Adaptation, quipping "That didn't take long: the model swallows the harness" — meaning test-time adaptation logic once built into external scaffolding is being internalized by the model itself.
- Elon Musk replied that the same happened with Tesla FSD: the model progressively ate more and more of the C++ code, "until there was nothing left."
- The exchange is a notable datapoint in the trend of learned systems replacing hand-engineered pipelines.
Related event: "Models Devouring the Harness" Sparks Debate on Agent Engineering(4 posts)→
More from AGI Musings
- Scott Aaronson: the singularity has already begun for math — ChrSzegedy · 2026-10-12
- AI Labs 2026: You Pay for the Kitchen, Not the Meal — gerardsans · 2026-10-12
- Meme: how did other species fare after encountering a 'superintelligence'? — Puzzleheaded-King584 · 2026-10-12
- Breiman's 2001 'Two Cultures' paper resurfaces as AI makes black boxes the norm — akbirthko · 2026-10-12
- AI filings are flooding courts — researcher proposes a parallel AI adjudication track — latticecut · 2026-10-12
- jd_pressman to Richard Ngo: MacAskill's opinion on AI doesn't matter, would SBF have skipped Anthropic? — jd_pressman · 2026-10-12