Cognition 用真实工程检验 Claude Fable 5 长程编程能力Cognition Tests Claude Fable 5's Long-Horizon Programming Abilities with Real-World Engineering Projects


AI free token/AI免费积分
Cognition 用真实工程检验 Claude Fable 5 长程编程能力ginobefun@hongming731 · X
2026-07-11 07:56·5小时前
在 X 看原推· x.com
AI 摘要

Cognition 基于 Devin 背景,用内部工程师真实任务检验 Claude Fable 5 的长程编程能力。其内部评测 Frontier Code 不仅要求补丁通过测试,还排除不可维护的结果,最难子集成绩从 Opus 约 10% 提升至 Fable 5 约 30%。团队强调模型能在杂乱日志中持续检索、区分已知与未知,并先明确约束再执行。文章指出“长程”更实用的定义是代理是否记得任务边界、检查操作、承认未知,并建议团队将信任拆解为可观察的过程证据。

http://x.com/i/article/2075729990013132800

BestBlogs 早报 · 07-11|真实工程检验长程编程,训练系统扩展智能体,产品团队重审判断与推荐透明度

在线阅读本期早报

BestBlogs.dev 是 AI 驱动的私人阅读助手。这是面向所有人的每日早报内容,如果你希望它基于你的兴趣和阅读习惯整理,可以体验「我的早报」。

导语

把一个完整的 issue 交给代理,真正令人不安的地方不在于它能否写出补丁,而在于你要在什么时候相信它已经完成。

今天的三篇精讲分别给出三个不相同的切面:Cognition 用工程师日常工作检验模型能否在长任务里保持判断;快手把可运行仓库变成训练场;Instagram 负责人 Adam Mosseri 则把视线放回产品团队仍要做的选择。

7 月 9 日我们还在讨论长任务模型如何进入工作流。今天更适合把问题收窄:模型能跑得更久之后,评测、训练环境与人的策略判断分别该承担什么责任?

ginobefun@ hongming731 · XCognition Tests Claude Fable 5's Long-Horizon Programming Abilities with Real-World Engineering Projects

48 Export Markdown

2026-07-11 07:56 · 5 hours ago

Cognition, drawing on Devin's background, tested Claude Fable 5's long-horizon programming ability using real tasks from internal engineers. Its internal Frontier Code evaluation requires not only that patches pass tests but also excludes unmaintainable results, raising performance on the hardest subset from ~10% with Opus to ~30% with Fable 5. The team emphasized that the model can continuously retrieve information from messy logs, distinguish known from unknown, and clarify constraints before executing. The article argues that a more practical definition of "long-horizon" is whether an agent remembers task boundaries, checks its operations, and acknowledges uncertainty, and recommends that teams decompose trust into observable process evidence.

http://x.com/i/article/2075729990013132800

BestBlogs Morning Report · 07-11 | Real Engineering Tests for Long-Horizon Programming, Training Systems to Scale Agents, Product Teams Revisit Judgment and Recommendation Transparency

Read this morning report online

BestBlogs.dev is an AI-powered personal reading assistant. This is the daily morning report content for everyone. If you'd like it tailored based on your interests and reading habits, try "My Morning Report."

Lead

Handing a complete issue to an agent—the truly unsettling part isn't whether it can write a patch, but when you decide to trust that it's finished.

Today's three in-depth pieces offer three distinct angles: Cognition uses engineers' daily work to test whether models can maintain judgment over long tasks; Kuaishou turns runnable repositories into training grounds; and Instagram head Adam Mosseri turns attention back to the choices product teams still must make.

On July 9 we were still discussing how long-task models enter workflows. Today it's more useful to narrow the question: once models can run longer, what responsibility should evaluation, training environments, and human strategic judgment each bear?

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry