Vercel Workflow SDK 5 与 durable-harnessVercel Workflow SDK 5 and durable-harness
Vercel CEO Guillermo Rauch 宣布 Workflow SDK 5 和 Vercel Workflows 是 agent 基础设施的重大进步:首步延迟和步骤间延迟大幅降低,加密与安全进一步加强,系统可靠性在大规模场景下得到验证,并抽出可复用的 durability log,催生了 durable-harness。他同时表示 Vercel Sandbox “不可避免”。
Vercel CEO Guillermo Rauch described Workflow SDK 5 and Vercel Workflows as a monumental step for agents: dramatically lower time-to-first-step and inter-step latency, stronger encryption and security, proven reliability at massive scale, and a reusable durability log that gave birth to durable-harness. He also declared “Vercel Sandbox is inevitable.”
查看原文 →查看原文 →
Garry Tan 的 harness 评测与模型成本对比Garry Tan’s open-source harness eval and model cost comparison
Y Combinator 总裁 Garry Tan 分享了自己的开源 harness 评测仓库。在 Claude Code 的 52 个编程任务上,Fable 5.2 解题数量与 Opus 5.5 接近(82.7% vs 84.6%),但每个成功任务的成本却是后者的 2.80 倍。因此他建议继续把 Opus 5.5 作为 Claude Code 的默认编码模型,并指出真正拉开差距需要更难的任务。
Y Combinator President & CEO Garry Tan released his open-source harness eval repo. On 52 coding tasks in Claude Code, Fable 5.2 solved almost as many problems as Opus 5.5 (82.7% vs 84.6%) yet cost 2.80× more per solved task. He therefore recommends keeping Opus 5.5 as the coding default and notes that harder tasks will be needed to differentiate the models further.
查看原文 →查看原文 →
Peter Yang 的 /fuck-cancer AI skillPeter Yang’s free /fuck-cancer skill
Peter Yang 发布了一个可免费使用的 AI skill:把患者全部病历上传到文件夹,然后让 ChatGPT 或 Claude 安装该 skill,即可生成一页摘要、检索临床试验等。他鼓励有需要的人 fork 并适配其他长期疾病,用 AI 全力对抗恶性疾病。
Peter Yang released a free AI skill that lets anyone upload a patient’s full medical records, point ChatGPT or Claude at the skill, and instantly generate a one-page brief, research clinical trials, and more. He invites people to fork it for other long-term conditions and use the full power of AI against serious disease.
查看原文 →
反对盲目刷榜,真正用起来才知道Stop benchmark-maxxing; real workflows matter
Nikunj Kothari 批评在 Llama 4 之后本该停止的“刷榜”又卷土重来。他强调真正重要的只有极少数基准,其余多为营销姿态;唯一可靠的方法是把模型放进自己的真实工作流里去感受。
Nikunj Kothari observed that benchmark-maxxing has returned after Llama 4. He argues only a handful of benchmarks truly matter and the rest is mostly posturing; the only reliable way to know which model works is to run it on your actual workflows.
查看原文 →