Mubit 推出可自训练小模型与 Banyan Sapling

Aakash Gupta @aakashgupta原推发布于 2026.10.23

在 X 上打开原推

Every bank, airline, and insurance company you called this year handed you the same robot. Different hold music, same brain, rented from the same couple of vendors. Now think about the exec who signed that contract. Her competitor signed the same one and gets the same answers back. The only thing she owns is the prompt, and prompts get copied fast. So where does an edge come from? I keep landing on the same place. Your agents do a handful of jobs over and over, sorting tickets and checking claims against policy, run after run, each one graded by whether a human had to step in afterward. No one outside your building has that record, and most teams let it expire in a logging tool. Mubit's bet is that those repeated jobs should become small models you own, retrained every few minutes, while the big frontier model handles the weird stuff. You change one line, the base URL, and keep paying OpenAI or Anthropic for everything the small model hasn't earned yet. What sold me is the open model they shipped with it. Banyan Sapling reads a policy, some evidence, and a claim, then says whether the claim holds up. On their eval it answered 779 of 1,731 cases and got all 779 right. For the other 952 it said "unsure." 45% coverage sounds weak until you ask who was reviewing all 1,731 yesterday. So 779 skip review, and the other 952 go to the humans who were already doing every single one. If you work anywhere regulated, you know what a model that says "unsure" at the right moment is worth, because you've cleaned up after the confident kind. There's also a benchmark where models plan a latte pour, get measured feedback, and try again, to check whether agents actually learn anything. A longer context window makes an agent better at your job about as much as a longer recipe makes someone a better cook. Your competitor can rent the same model you do, and the logs they'd need to catch up only exist in your building.

这条是引用着另一条写的：x.com/1832067056428015616/status/211365182313113…