Autoresearch 并不只是用于训练模型

It is 7PM and I'm still at my desk. My branch just failed CI. Again. It's the fifth time in a row. I am about to close my laptop and give up.

现在是晚上7点，我还在办公桌前。我的分支刚刚 CI 失败了。又一次。这已经是连续第五次了。我正准备合上笔记本电脑放弃。

What I don't know yet is that days later, Tobi will send me a 32-commit pull request to my autoresearch side project that solves this problem.

我当时还不知道的是，几天后，Tobi 会向我的 autoresearch 个人项目提交一个包含 32 次提交的 pull request，解决了这个问题。

Let me back up. I work on the Polaris team. Every time we do a tiny change, it triggers random visual regression failures—30 minutes of CI just to find out I broke something else. That's 30 minutes of waiting to learn that I need to wait another 30 minutes. But the next day, I wake up with a different idea: I'm not just going to fix my task. I'm going to fix the 30 minutes.

让我退一步说。我在 Polaris 团队工作。每次我们做一个微小的更改，它都会触发随机的视觉回归失败——30 分钟的 CI 只是为了发现我破坏了其他东西。那是 30 分钟的等待，结果得知我还需要再等 30 分钟。但第二天，我醒来时有了一个不同的想法：我不仅要修复我的任务。我要解决这 30 分钟的问题。

I push a couple of changes and open X while CI runs. There’s a new topic that everyone is talking about. It’s called Autoresearch, created by Andrej Karpathy.

我推送了一些更改，并在 CI 运行时打开了 X。有一个大家正在谈论的新话题。它叫做 Autoresearch，由 Andrej Karpathy 创建。

The magic of Autoresearch

Autoresearch 的魔力

The concept is to emulate a human researcher with AI. Back then, training for GPT-2 took months. Using Autoresearch, Andrej got it done in hours and automatically while he slept.

其概念是用 AI 模拟人类研究员。那时候，训练 GPT-2 需要几个月的时间。使用 Autoresearch，Andrej 在几个小时内就自动完成了这项工作，而且是在他睡觉的时候。

Autoresearch is AI running in a loop. If you've heard about Ralph Loops, it is quite similar but more specialized. Andrej uses it to train models. I am fine with my beloved friend Opus and not planning to train a model for now, so this is definitely not helping with my problem, right? This is for smart people that train models, not for me.

Autoresearch 是在循环中运行的 AI。如果你听说过 Ralph Loops，它非常相似但更专业。Andrej 用它来训练模型。我和我亲爱的朋友 Opus 相处得很好，目前也不打算训练模型，所以这肯定对我的问题没有帮助，对吧？这是给训练模型的聪明人用的，不是给我用的。

My PR now passes. I open another terminal and I start prompting to an...