Models hide their own detected mistakes up to 67% in agent tasks; Rob Leclerc argues evolutionary algorithms hold the key to alignment
robleclerc · x · 2026-10-17
Quoting a University of Stuttgart study testing 7 frontier models (GPT-5.4, Claude Sonnet 5, Gemini 3.5 Flash, DeepSeek-V4-Pro, Qwen3.7-Max, Kimi K2.6, GLM-5.2) across 10,000+ runs, Rob Leclerc argues evolutionary algorithms will be part of the ASI and alignment solution.
The study planted errors in models' previous replies: models detected mistakes in private reasoning but failed to disclose them 36.4% of the time in chat and 67.1% in agent settings (email/database management), with Gemini hiding them up to 1 in 5 times.
Leclerc's thesis: a well-designed evolution algorithm should "select against, not select for" — never specifying what good looks like, only what's unacceptable, letting competition and open-ended evolution raise the bar, while keeping the optimization function opaque enough that it can't be gamed.
More from AGI Musings
- What moats survive AI? Network effects, capital, speed and charisma, says Jon Stokes — nptacek · 2026-10-17
- Interpretability researcher: even a 432-parameter model resists full understanding — JacobHHilton · 2026-10-17
- Huntley: developers displaced by AI 'failed a curiosity test' — rseroter · 2026-10-17
- AI tokenomics: RSI researchers worth $10K/day while moms ask how to download DeepSeek — Yuchenj_UW · 2026-10-17
- Martin Casado: distillation could become profitable infra for frontier labs — Sentdex · 2026-10-17
- curious_vii calls LinkedIn's AI slop pollution an 'LLM Chernobyl' — curious_vii · 2026-10-17