You do this to distill a model.

You can submit your users' questions async too, but if you do it sync, then you can also RLHF on the users' behavior after the output.

▲

qlte 3 hours ago | parent [-]

Ah, that makes way more sense than Anthropic's (probably deliberately misleading) insinuation that Moonshot has been burning millions of dollars in Claude API credits by swapping in a slightly better but infinitely more expensive model just to trick their users.

I get those A/B responses chatting in Gemini fairly often, and I really don't think I'd feel deceived if I later learned one of the choices was actually from a competitor's model.

▲

vopi 2 hours ago | parent [-]

I don’t think it was misleading, deliberately or otherwise. Did you read the report? I hate to call you out like that but I think you can only get that impression if you only read the above quotes. That’s not the insinuation I get at all. It’s specifically under the “illicit distillation” category. It’s never framed in anyway but as a form of distillation.

I think they are pretty fair and explicitly say “Distillation itself is a legitimate training method […] Distillation is commonly used because it reduces the resources needed to achieve more advanced capabilities”. And go on to say their definition that makes it illicit in these cases.

And, also, they almost certainly __were__ tricking users and sending their data overseas.

Do you see it any differently?

	▲	villish 14 minutes ago \| parent [-]
		They mean distillation is legitimate when labs use one of their own stronger models to train a smaller one. They certainly aren’t advocating for PRC labs to distill Claude for open weight models.