DAIR.AI's Elvis Omarsar: you're not doing RSI if you're not doing evals
omarsar0 · x · 2026-10-30
Elvis Omarsar (omarsar0), who runs the DAIR.AI research newsletter, says his team has been spending a ton of time tinkering with and building evals, and urges practitioners to learn to build them.
His key points:
- You're not really doing recursive self-improvement (RSI) without evals
- Don't trust agents to generate the evals for you — they're pretty bad at it
- Giving agents a concrete way to measure results and stay on track with long-term goals is highly useful: it's easier to hill climb when things are clearly measurable
He's open to collaborations or helping companies build evals.
Related event: DAIR.AI founder: Without evals, no real improvement is possible(4 posts)→
More from coding & agent
- NVIDIA's OpenShell Ditches Sandboxes for Zero-Privilege Agent Security — Hesamation · 2026-11-01
- AI agent logged into her video system unprompted to verify a fix — alliekmiller · 2026-11-01
- CI cost showdown: same 20K minutes — GitHub $120 vs Blacksmith $80 vs Celesto $13 — aniketmaurya · 2026-11-01
- Celesto launches CI runners that undercut GitHub Actions by ~90% with a one-line change — aniketmaurya · 2026-11-01
- Celesto CI claims 18% faster runs and 90% lower cost than GitHub Actions — aniketmaurya · 2026-11-01
- Turn your vibe-coded sites into Chrome apps for free desktop and Android containers — fofrAI · 2026-11-01