[logicchains]

They're not running at a loss. Training runs at a loss, but the models are profitable to serve if you don't need to continuously train new models.

[jayd16]

But you do or you're missing current events, right?

[dcre]

Not at all, otherwise models with knowledge cutoffs of six months to a year ago (all current SOTA models) would be useless. Current information is fed into the model as part of the prompt. This is why they use web search.

The main reason they train new models is to make them bigger and better using the latest training techniques, not to update them with the latest knowledge.

[jay_kyburz]

I'm trying to avoid getting into the habit of asking LLMs about current events, or really any events. Or really facts at all.

I think LLMs work best when you give it data, and ask it to try make sense of it, or find something interesting, or some problem. To see something I can't see, then I can go back and go back to the original data and make sure its true.

[fragmede]

There are a number of techniques to modify a model post-training. Some of those techniques allow adding current events to the model's "knowledge" without having to do an entire from-scratch training run, saving money.