Skip to content

Lauren Tan’s AI Engineer Talk Outlines How the Grok Bot Team Ships Code

Baoyu · XRelayed· @dotey·· 1 day ago

9.1K views · 113 likes · 9 reposts (at collection)

Original

What happened

Lauren Tan’s AI Engineer talk, “All Code Is Slop (Except Mine),” outlines practices used by the Grok Bot team, where she leads the desktop version. Her recommendations include turning corrections into lint rules, keeping the codebase clean, banning error-prone features and most comments, and having AI run the program and provide recordings and screenshots for review. The post reports that the team merged 11,000 PRs last month, with about 0.5% rolled back, and says the app is faster than at launch.

Read next

More changes in today’s briefingGemini Connects to Google Wallet for Personal Finance Insights
Body · English edition pending (original content)

Lauren Tan 在 AI Engineer 大会上的演讲《All Code Is Slop (Except Mine)》要点。她在 SpaceXAI 负责 Grok Bot 桌面版。

  1. 每纠正 AI 一次,就把这次纠正写成 lint 这类自动检查规则,少往 agents.md 这类说明文档里加话。违反 lint 规则的代码会直接报错、提交不上去,文档里的话 AI 不一定照做。
  2. 代码库是 AI 最主要的学习材料。里面的坏写法、临时补丁,会被下一个 AI 当成惯例照抄,像杂草一样越长越多,所以代码库要保持干净。
  3. 容易用错的东西,直接禁掉。React 里有个功能人都常用错,AI 更用不好,他们在产品代码里禁用了它,这类问题从此不再出现。
  4. 禁掉 99% 的代码注释。AI 爱在注释里给自己走的捷径找理由,后来的 AI 就照着学。她在 Cursor 时,这种“有理由的补丁”扩散开来,把应用拖得很卡。
  5. 让 AI 自己检查自己的活。AI 把程序跑起来像用户一样点一遍,交上录屏和截图,人审代码就变成看这些材料。她的 Token 大部分花在这一步。
  6. 以前用不起的检查方法,现在用得起了。用数学方法证明程序没问题这类严格验证,过去太费人力,AI 把成本降了下来。
  7. 结果:团队上个月合并了 1.1 万个 PR,被撤回的约 0.5%,应用比刚发布时还快。
  8. 先打好地基,再谈全自动。AI 写代码这一步本身出烂代码的话,全自动化只会出更多烂代码,还更费钱。
Quoted postlauren@poteto

my AIE talk is up! come hear how the grok @bot team ships so quickly without compromise on quality

youtu.be/O_zcc6Xl12g

View the quoted post on X
This story is in 11/16 AI DailyRead the whole issue