Opinion: Most Agentic Tasks Don't Need Post-Training—Evals Are the Moat
abeirami · x · 2026-09-11
Rajeswar Sai amplifies abeirami's argument that post-training is the last step in agentic work and can often be skipped entirely.
The core claim: the real leverage has been sitting in companies' own workflows all along—writing down what "done" means for your key workflows is your eval and your moat, yet almost nobody does it. abeirami ranks priorities as evals ≫ harness optimization (fast learning) ≫ model post-training (slow learning), arguing most agentic tasks need no slow learning loop; even when post-training is the goal, you first need high-SNR evals, then optimize the harness on that signal, and only then distill the resulting system behavior back into the weights.
More from AGI Musings
- mrdoob sparks debate: AI coding traded the joy of flow for new capabilities — brandon_xyzw · 2026-09-11
- iamtrask proposes resignation matching: AI safety researchers quit frontier labs in pairs — iamtrask · 2026-09-11
- Andy Matuschak experiments with weirder representations of 'readings' via cut-up poster-ification — andy_matuschak · 2026-09-11
- AI math skeptics gain ground: advanced math offers little economic value, argues viral thread — voooooogel · 2026-09-11
- Dev jokes he's glad he meditated before diving into ASI risk discourse — JacquesThibs · 2026-09-11
- Meta gives every US adult free AI, yet coverage fixates on a 27-year-old quitter — neil_chilson · 2026-09-11