Evals Are the Most Important Skill in AI Today, Argues AI Researcher Omarsar
omarsar0 · x · 2026-10-30
AI researcher Elvis Omarsar argues that building evals is one of the most important skills in AI today. Without evals, he says, there's no real recursive self-improvement — and don't trust agents to generate evals for you, as they're bad at it. Giving agents concrete ways to measure results keeps them on track toward long-term goals: hill climbing only works when things are clearly measured. No amount of agents or intelligence gets you far without proper evals.
More from AGI Musings
- Microsoft CEO Nadella: 'I no longer believe alignment will just fall out' — AndyMasley · 2026-10-30
- Voice is the ultimate interface for humans and agents, investors argue — graceisford · 2026-10-30
- Peer review is plagued by human bias — AI may actually help fix it — pwlot · 2026-10-30
- If AI Becomes Obeisant but Powerful, Every Nation Must Reform Its Command Hierarchy — QuintinPope5 · 2026-10-30
- Integer factorization proven subexponential, with core work done by Claude — burny_tech · 2026-10-30
- A year of using AI as an inexhaustible sounding board to untangle your own thoughts — Takedori · 2026-10-30