OpenAI just officially said it has reached its "automated research intern" milestone. i.e. a human-supervised system able to complete well-defined tasks that would take a skilled researcher quite few days. inside OpenAI research, agent runtime has already crossed human labor by a wide margin. 3.1-to-1 agent-to-human ratio “In terms of a standard 8 hour workday, as of mid-August, in total, the research organization uses 3.1 agent-workdays of effort for every workday of human labor.” That ratio measures agent runtime rather than equivalent productivity, but it captures how deeply parallel agent work has entered OpenAI research.
X:Rohan Paul
@rohanpaul_ai · X
切换来源
@rohanpaul_ai@rohanpaul_aiAI 评分5656
引用@rohanpaul_ai@rohanpaul_ai@rohanpaul_ai@rohanpaul_aiAI 评分3030 Jevons 悖论应用于劳动力。 当智能变得廉价,对它的需求就会爆炸式增长,而最接近源头的人会吸收这一切。 AI 确实把时间还回来了,但领域里的每个人都只是把它再投资了回去。

@rohanpaul_ai@rohanpaul_aiAI 评分5454 
@rohanpaul_ai@rohanpaul_aiAI 评分4141 
@rohanpaul_ai@rohanpaul_aiAI 评分4444 微软与康奈尔大学论文提出 Free Pause Tokens,让模型预测下一个 token 时获得额外计算,且不增加 token、不扩大 KV cache、不增加解码步骤。

@rohanpaul_ai@rohanpaul_ai精选AI 评分7878 
推荐理由:OpenAI 首席科学家公开谈论扩展节奏与对齐缺口,读者可据此了解头部实验室对安全进展的最新判断。
@rohanpaul_ai@rohanpaul_aiAI 评分6464
引用@rohanpaul_ai@rohanpaul_aiOpenAI just officially said it has reached its "automated research intern" milestone. i.e. a human-supervised system able to complete well-defined tasks that would take a skilled researcher quite few days. inside OpenAI research, agent runtime has already crossed human labor by a wide margin. 3.1-to-1 agent-to-human ratio “In terms of a standard 8 hour workday, as of mid-August, in total, the research organization uses 3.1 agent-workdays of effort for every workday of human labor.” That ratio measures agent runtime rather than equivalent productivity, but it captures how deeply parallel agent work has entered OpenAI research.
@rohanpaul_ai@rohanpaul_aiAI 评分3838 引用@rohanpaul_ai@rohanpaul_aiOpenAI engineers are changing roughly 7x more code per contributor than the pre-2025 baseline. https://t.co/K8EK8zuvqe https://t.co/5afIiyoITp
@rohanpaul_ai@rohanpaul_aiAI 评分5656
引用@rohanpaul_ai@rohanpaul_aiOpenAI just officially said it has reached its "automated research intern" milestone. i.e. a human-supervised system able to complete well-defined tasks that would take a skilled researcher quite few days. inside OpenAI research, agent runtime has already crossed human labor by a wide margin. 3.1-to-1 agent-to-human ratio “In terms of a standard 8 hour workday, as of mid-August, in total, the research organization uses 3.1 agent-workdays of effort for every workday of human labor.” That ratio measures agent runtime rather than equivalent productivity, but it captures how deeply parallel agent work has entered OpenAI research.
@rohanpaul_ai@rohanpaul_aiAI 评分2020 反AI的愤怒病毒是真实存在的。 事实上,反AI的愤怒是对工具有效最响亮的承认。 https://t.co/tKgtijLfLK https://t.co/qf7sEgCAcQ

@rohanpaul_ai@rohanpaul_aiAI 评分4747 Claude Code 作者 Boris Cherny 在 Startup School 2026 上被问及人们如何学会像他一样使用 Claude Code,他建议别听 LinkedIn 网红的说法。

@rohanpaul_ai@rohanpaul_aiAI 评分44 @rohanpaul_ai@rohanpaul_aiAI 评分3636 引用@rohanpaul_ai@rohanpaul_aiKnowledge work has always been bottlenecked by human serial execution - that time is changing. OpenAI's research workflow has now shifted from individual AI assistance toward researchers supervising multiple simultaneous agent workflows. In April, only about one-third of researchers were hitting 4+ concurrent agent workflows; by mid-August, it was roughly three-quarters.
@rohanpaul_ai@rohanpaul_aiAI 评分5959
引用@rohanpaul_ai@rohanpaul_aiOpenAI just officially said it has reached its "automated research intern" milestone. i.e. a human-supervised system able to complete well-defined tasks that would take a skilled researcher quite few days. inside OpenAI research, agent runtime has already crossed human labor by a wide margin. 3.1-to-1 agent-to-human ratio “In terms of a standard 8 hour workday, as of mid-August, in total, the research organization uses 3.1 agent-workdays of effort for every workday of human labor.” That ratio measures agent runtime rather than equivalent productivity, but it captures how deeply parallel agent work has entered OpenAI research.
@rohanpaul_ai@rohanpaul_ai精选AI 评分6565 
推荐理由:文中给出 agent 运行时长超过人类工作时长的比值变化,并说明了该比值只衡量运行时长而非同等生产力。
@rohanpaul_ai@rohanpaul_aiAI 评分55 @rohanpaul_ai@rohanpaul_aiAI 评分3535 
@rohanpaul_ai@rohanpaul_aiAI 评分2020 @rohanpaul_ai@rohanpaul_aiAI 评分5151 微软一篇新论文提出把测试时推理成本摊销为蒸馏技能:收集 35–50 条历史轨迹,由编码智能体提取反复出现的失败模式,再将其写成 markdown 技能加入非推理模型的系统提示词。

@rohanpaul_ai@rohanpaul_aiAI 评分00 @rohanpaul_ai@rohanpaul_ai精选AI 评分8484 
推荐理由:报道给出 Anthropic 与 OpenAI 在收入和算力上的对比,并披露供应商同时投资买家的融资结构。
@rohanpaul_ai@rohanpaul_aiAI 评分2323 @rohanpaul_ai@rohanpaul_aiAI 评分44 @rohanpaul_ai@rohanpaul_aiAI 评分3030 
@rohanpaul_ai@rohanpaul_aiAI 评分66 @rohanpaul_ai@rohanpaul_aiAI 评分1818 
@rohanpaul_ai@rohanpaul_aiAI 评分1818 – https://t.co/5zCBqp3CAb 标题:"Repo-To-Skill:将 GitHub 仓库蒸馏为 AI4AI 技能"
@rohanpaul_ai@rohanpaul_aiAI 评分4848 
@rohanpaul_ai@rohanpaul_aiAI 评分77 Hugging Face:https://t.co/cnMs1y9aUG GitHub:https://t.co/y7ILKSLDxF
@rohanpaul_ai@rohanpaul_aiAI 评分3434 
@rohanpaul_ai@rohanpaul_aiAI 评分55 引用@rohanpaul_ai@rohanpaul_aifull video https://t.co/wBA6vhRyvn
@rohanpaul_ai@rohanpaul_aiAI 评分3535 
@rohanpaul_ai@rohanpaul_aiAI 评分1919 @rohanpaul_ai@rohanpaul_aiAI 评分4444 
@rohanpaul_ai@rohanpaul_aiAI 评分4646 
@rohanpaul_ai@rohanpaul_aiAI 评分2323 – https://t.co/7DxSnqWl9r 标题:"SkillGLoW:面向长时程任务流中自我改进智能体的过程族技能整合"
@rohanpaul_ai@rohanpaul_aiAI 评分3939 
@rohanpaul_ai@rohanpaul_aiAI 评分55 @rohanpaul_ai@rohanpaul_aiAI 评分3838 
@rohanpaul_ai@rohanpaul_aiAI 评分3535 