什么是 epoch.ai？

Epoch AI推出Benchmark Reviews项目，揭示当前AI基准测试存在严重缺陷，15个基准中9个被标记为有缺陷。

FrontierCode 是一项新的代码质量评估基准，专注于衡量代码是否可合并，而非仅通过单元测试。

China's open-source models lag behind the US frontier models by 3-7 months according to CAISI evaluations.

China's open AI ecosystem reduces redundant R&D compute costs, enhancing model development efficiency and sustainability.

AI evals are fundamentally broken—over-reliance on objective metrics misleads—but they remain critical when built, interpreted, and embedded properly in agent workflows.

Author predicts 2026 will be a key year for AI development, with open models facing both challenges and opportunities.

Memory now accounts for 63% of total AI chip component costs, making it the largest single cost driver.

FrontierMath evaluation found fatal errors in ~33% of problems; Epoch AI will release corrected dataset with updated scores.

AI在数学研究中的应用正在加速渗透，预计一年内将成为主流工具。