What's this about?
Building a self-hosted LLM that covers the corporate request mix, using the company's own production traffic as the basis for post-training. 32 upvotes.
Why is "the corporate request mix" the right target?
Because a company does not need a model that is good at everything — it needs one that is good at the specific distribution of things its employees actually ask. That distribution is narrow, idiosyncratic, and completely invisible from outside: the internal jargon, the recurring document types, the five workflows that account for most requests. Post-training on real production traffic targets exactly that, and it is the argument for self-hosting that is not about privacy: a small model tuned on your distribution can beat a frontier model on your distribution.
Why does this matter to you?
- It is a documented recipe for the "should we self-host" question that most organisations answer with vibes.
- The privacy story is real too — production traffic is the most sensitive corpus a company has, which is why the Bocconi-style prompt studies are so rare.
- 32 upvotes for an engineering account rather than a novel method suggests practitioners found it useful.