Epoch AI

Brief

Frontier models make little progress on an ML innovation test

Epoch AI's early InnovationEval results show little progress on devising a human-matching post-training method despite GPU spend. The team plans to repeat the test.

Epoch AI 2 days ago agent benchmarks

Signal: 32
Facts: 56
Hype: 2

Brief

GPT-6 Astra hits 100% on EBR-bench using a banned time bypass card

Epoch AI says the card is banned in the default setting. Multi-agent scaffolds show no clear score effect on the benchmark.

Epoch AI 2 days ago agent benchmarks

Signal: 38
Facts: 75
Hype: 7

Brief

Epoch tests six models on real tasks: strong on defined work, weak on open-ended

Frontier models do well on well-defined tasks but fail open-ended work, style matching and experiment design. Background for benchmarking.

Epoch AI 2 days ago agent benchmarks

Signal: 42
Facts: 70
Hype: 9