METR warns AI progress now outpaces experts' ability to evaluate safety
The Decoder · rss · 2026-11-22
METR (Model Evaluation & Threat Research) warns that leading AI models such as Opus 5.5 and GPT-6 Astra are improving rapidly at research tasks while becoming harder to measure reliably.
- Current data cannot rule out a dangerous acceleration in AI development
- However, METR also notes that common tests may overestimate what the models can actually do
The core conclusion: model capability progress has outpaced the evaluation infrastructure needed to verify safety, making measurement itself a bottleneck.
More from AGI Musings
- Living cells aren't theorems awaiting type-checks: a debate on why biology has no Lean — JosephJacks_ · 2026-11-22
- GTA 6 Cost $1B — In 5 Years a Game Like It Might Cost $30 in API Calls — flowersslop · 2026-11-22
- Hugging Face's Lewis Tunstall Maps the Road from LLMs to Agent Swarms — _lewtun · 2026-11-22
- Mathematician Ken Ono publishes 2,500-word essay on staying in math in the AI era — Dr_Atoosa · 2026-11-22
- Commentator predicts AI will consume every field within 1-2 years, starting with math — haider1 · 2026-11-22
- Interview: Medical AI is only as good as human consensus labels — CurieuxExplorer · 2026-11-22