Apollo Research's Hobbhahn calls for evaluating models throughout development
moniquejmorrow · x · 2026-10-29
Apollo Research's Marius Hobbhahn argues for evaluating models throughout development, including testing whether our detection methods can catch unsafe behavior in systems where it's already known to exist — a meta-test of evaluation itself, responding to the problem of models behaving differently under evaluation.
Related event: Debate grows over when and how to evaluate AI safety(5 posts)→
More from Safety
- a16z Podcast with Harvey's CLO: AI Expands Legal Services Rather Than Replacing Them — MattPerault · 2026-10-29
- UK budget allocates £40m for a National Agentic Security Unit, £120m more for AISI — davidmanheim · 2026-10-29
- Malware embeds nuke and bioweapon text to trip LLM safety refusals and dodge AI scanners — davidmanheim · 2026-10-29
- A new context poisoning attack tricks frontier LLMs into bioweapon content — davidmanheim · 2026-10-29
- MATS to Host AI Safety Talent Fair in Berkeley on November 19 — austinc3301 · 2026-10-29
- Deploying AI Agents Is Flying Blind — We Need Real-Time Monitoring — pstAsiatech · 2026-10-29