METR Time Horizon 2.0—The benchmark you’ve been waiting for by AgentMa🔸

·Nuno Sempere··

TL;DR: I ap­plied METR’s time-hori­zon method­ol­ogy to Microsoft Ex­cel. On my 23-task suite, Ex­cel com­pletes tasks that take an un­aided hu­man 6.5 hours at 80% re­li­a­bil­ity — more than dou­ble Claude Mythos Pre­view, the best fron­tier model on the cur­rent METR plot (~3 hours at 80%). Ex­cel’s 50%-time hori­zon is 54 work­ing weeks. I have up­dated my prob­a­bil­ity of spread­sheet-in­duced catas­tro­phe by 2035 from 0.1% to 5–10%. Then I wrote down ev­ery­thing wrong with my method­ol­...

Read full article →

Related Articles

Same Question, Three AI Models, Different Answers: A Case Study in AI Biosecurity Safeguards by Nnaemeka Emmanuel Nnadi
Nnaemeka Emmanuel Nnadi · Nuno Sempere · 10h ago
AI Forecasting in 2026: What 11 Analyses Say by Benjamin Wilson 🔸
Benjamin Wilson 🔸 · Nuno Sempere · 11h ago
A Calibration Benchmark for LLM Beliefs Across a Taxonomic Hierarchy by DanRKAlex
DanRKAlex · Nuno Sempere · 1d ago
When will AI forecasting crush human superforecasters?
Daniel Reeves · Manifold Markets · 1d ago
Disagreement = Counterfactuality for Matching Funds by ellie
ellie · Nuno Sempere · 2d ago