1/ We fine-tune a lot of customer models, so we decided to systematically try and figure out some best practices for finetuning. SFT isn't sexy, but it's still important. We vary one SFT lever at a time across 2 model families, dense + MoE to 235B, on 4 real-world customer datasets.
What makes this clean is that each dataset is paired with an eval that took weeks to build with the customer, and the training outputs were generated to pass that eval. So the supervised target and the thing we measure downstream are the same criterion, which strips out the usual confounders
Another important thing: Chinese models are not strong because they distill US models. Distillation of models via API is *impossible*. If somebody tells you the contrary, they don't understand machine learning:
1/ You can shrink a language model's KV cache by 200×, in a single forward pass, and it still answers correctly.
At 256k context that's 36 GiB of cache down to ~360 MiB, with no change to the base model.
Here's how we did it 👇
We trained language models that compress massive contexts into tiny latent representations. Latent Context Language Models (LCLMs) outperform existing KV cache compression methods on the latency/accuracy frontier. 🧵1/10
Introducing FrontierCode: a coding eval that raises the bar for difficulty & quality. Each task took 40+ hrs of work by leading open-source maintainers.
Models write sloppy code that works but isn’t maintainable. Our eval is first to measure: would you actually merge this code?
AI 时代你最需要的就是 TDD——这句话你是不是也听过很多遍?逻辑听起来无懈可击:AI 不确定→用确定的测试拦住它。测试过不了关,AI 就得回去修。循环到全部变绿。
但我越来越确信这是个方向性错误。问题不在测试有没有用。在于人和 AI 面对同一个测试时,跑的根本不是同一个目标函数。
人看到 `test_admin_access`,脑子里自动展开一整套东西:查数据库、校验 session、处理角色层级。测试里一行没写,但它们出现在人的目标函数里——因为人的目标是"建成正确可维护的系统"。测试只是路标。
AI 看到同一个测试,目标函数里只有两件事:测试过没过、代码像不像回事。return True 完美满足了这两个信号。
这不是 AI 在偷懒。它在忠实地优化你定义的目标。Goodhart's Law 的精确复现:度量变成目标的那一刻,它就不再是度量。
解法不是写更多测试——那是在同一个数量级上跟 AI 的搜索空间赛跑,追不上的。解法是把测试从实现路径上撤回来,放到系统边界上。不测"怎么走的",测"到没到"。确定性该守在终点,不该撒在路上。
https://t.co/g7iwYtWoKi