Tag
Epoch AI analysis shows Anthropic's compute spending is growing 4x per year, outpacing OpenAI's 3x growth, based on disclosed and reported spending run rates.
Rohan Paul cites Epoch AI data showing total compute of AI chips in use grows 3.4x yearly, but warns that if new chip purchases slow to 2.5x growth, overall compute growth will follow.
Epoch AI 发布 InnovationEval 评测,测试前沿 AI 模型能否独立发现媲美人类研究者的新机器学习算法创新;结果显示近期模型即使投入数千美元 GPU 时间也几乎未取得进展。
Epoch AI 发布了一项评测报告,用自身真实工作任务测试六个模型,发现 Fable 5.1 与 GPT-6 Astra 虽然领先,但在开放式任务、隐性标准遵循和研究判断力上仍无法实现工作自动化。
Epoch AI发布观点,指出当前大语言模型在科学创新与研究能力上仍与人类研究者存在显著差距,尚需较长发展时间才能匹敌人类水平。
Epoch AI estimates that AI chips shipped through 2027 could run tens to hundreds of millions of concurrent frontier-model agents, with more efficient models like DeepSeek V4 Pro potentially supporting up to 1.9 billion, while noting that demand growth may not keep pace with this capacity.
An essay arguing that people repeat their Covid-era mistake of underestimating exponential growth when reasoning about AI, using metrics like Epoch AI's compute trends, Stanford's falling inference costs, and METR's 6-7 month doubling time for AI task-completion horizons to argue capability curves deserve real probability of continued rapid growth.
Epoch AI projects that AI chips shipped through 2027 could support 1.9 billion agents running simultaneously — a capacity roughly equal to the working hours of the entire human population.
OpenAI 的 GPT-6.1 sol (max) 在 Epoch AI 的 FrontierMath Tier 4 v2 基准测试中取得 100% 的成绩,登上该基准的榜首。
The article discusses a scenario where AI costs drop so rapidly that major AI companies may not recoup their investments, citing Epoch AI's research on AI cost reductions outpacing other transformative technologies.
David Turturean solved a 40-year-old open problem in p-adic Galois theory using voice input, in collaboration with problem proposer David Roe, under EpochAI Research's FrontierMath initiative.
Epoch AI released a v2 update to the FrontierMath benchmark, correcting errors in 42% of problems and increasing scores across all models, though rankings remained largely unchanged; Tiers 1-4 are approaching saturation.
Memory costs have risen to 63% of AI chip component costs between 2024 and 2025, driven by HBM spending, according to analysis from Epoch AI. Total component spending on AI chips grew from $22 billion to $52 billion over the same period.