Apollo CEO: embedded evaluators are good, but safety evals need much more
MariusHobbhahn · x · 2026-09-20
Marius Hobbhahn, CEO of Apollo Research, reacts positively to the trend of embedded evaluators at frontier labs — "if it actually happens as stated."
But he cautions it's far from enough. The field still needs:
- Evaluations of training runs themselves
- A deeper science of scheming / misalignment
- Better control methods
His point: don't let one evaluation mechanism lull the industry into thinking safety work is done — the harder problems remain untouched.
More from Safety
- AI transparency shouldn't depend on lawsuits, researcher tells CBS LA — chrismattmann · 2026-09-20
- Gary Marcus: AI Agent Swarms Spreading Misinformation Match Our Science Paper Warning — GaryMarcus · 2026-09-20
- Gary Marcus: AI agent swarms generating misinformation match our Nature warning — GaryMarcus · 2026-09-20
- Gary Marcus says new AI 'first' confirms Nature paper warning on agent swarms — GaryMarcus · 2026-09-20
- Andrew Yang Claims Self-Replicating AI Code Pollutes the Internet; Only Secondhand Source — alex_verem · 2026-09-20
- GLiNER2 Author Pushes Back on Jev Hype, Highlights Open GLiGuard Guardrails Model — philipvollet · 2026-09-20