We've entered into an agreement to join OpenAI as part of the Codex team.
I'm incredibly proud of the work we've done so far, incredibly grateful to everyone that's supported us, and incredibly excited to keep building tools that make programming feel different.
I posted this week about #MySQL's decline. Percona also writes in https://t.co/oDfLqbyMag: "The overall trend since 2011 shows a sustained decline in the number of commits and a shrinking pool of unique contributors. The trendline is a clear warning that, without intervention, the general development pace is expected to slow further."
👉If you still use MySQL, probably time to switch
We've published a detailed postmortem on three infrastructure bugs that affected Claude between August and early September.
In the post, we explain what happened, why it took time to fix, and what we’re changing:
This is going to be ironic coming from me but the best software engineers I've ever worked with and that you can ever hire are the ones with no online presence, no GitHub activity, and spent the last 10 years of their life working at the most boring company you can ever imagine.
We’re excited to officially announce the premiere of the FULL Python documentary! 🐍
🗓️ August 28th
🕖 7PM CET | 12PM CST
Subscribe on YouTube, set your reminder, & tell anyone who will listen! 📢
https://t.co/ng1IKWKQ2W
Agent 的架构和 Claude Code 的原理,Anthropic 去年发的那篇 《Building effective agents》就讲的很清楚了,上个月的那篇《How we built our multi-agent research system》把 Deep Research 的原理也讲清楚了。
Agent 的底层原理就是模型在收到任务请求后,去调用不同的工具收集上下文信息,必要的话和用户进行讨论,然后规划任务,再借助工具执行任务,并且模型要判断任务的完成情况。
早期的模型没有受过这套流程的强化学习训练,在调用工具、规划 、判断任务完成方面都比较弱,所以像 Auto GPT 这样最早期的先驱,用 GPT-4 也做不出好的效果。
到了 o3 之后,OpenAI 借助强化学习,训练了 Deep Research 这个垂直场景,在搜索任务和搜索相关工具的调用上,做到了极好的效果,接着其他家模型也跟进,所以现在很多模型在 Deep Research 这个场景上都做的不错。
Claude 4 则是把编程相关的工具调用和任务场景借助强化学习对模型进行了大量训练,所以 Claude Code 在编程场景效果很好。但当前的 Gemini 2.5 Pro 没有对 Claude Code 用到的那些工具做过强化学习训练,那么即使可以驱动 Claude Code,效果也会差一截。
Kimi K2 显然是针对 Claude Code 用到的工具做了大量强化学习训练,加上基座模型有不错的编程能力,所以使用 Claude Code 的效果也不错。
有趣的是 K2 不是推理模型,但是执行 Agent 任务也做的不错,看来推理模型并非必须,但 RL 更重要。
选择器是不是在用户手里,取决于模型之间的竞争,一家独大自然没有选择,百花齐放那就有得选,从趋势看,未来很难一家独大了,但是可能会有时间差,就像编程,短期肯定还是 Claude 4 最好。
用户有得选,但 AI Agent 创业团队还是会面临辛苦优化的提示词和工作流被新一代模型能力内化的尴尬处境。比如 Deep Research、AI 播客类产品。
In the long term, much of “resilience” is the ability to fall and stay asleep for 6-8hrs despite daily stress. Learn to “turn off your thoughts”, fall & stay deeply asleep. Non Sleep Deep Rest teaches you how.
멘토링 이야기를 어느 분이랑 하다가 국내에서는 멘토링을 주니어만 이야기하지만, 해외 사례를 보거나 오픈소스 프로젝트 내에서도 어느 정도 경험이 쌓인 엔지니어도 다음 단계를 위한 멘토링을 받기도 한다. "IU도 보컬 트레이닝을 지금도 받는다"라는 말이 굉장히 인상이 남아서 사소하게 적어둔다