green-frog-445

what do you mean by "gets"?

As in, it infers correctly what you want, right?

Is this a property that is trained in RL or from scaling pre-training?

Would you say this is an intelligence property or a preference property, as in, can different, equally smart models, "get us" differently when we talk to it?

I'm asking because I've thought about this a lot and it seems to me that the experience of the model "getting" it when we prompt it with increasingly vague prompts is just a glorified auto-complete experience, where what is being inferred by the model is just what it had been trained to infer, not really that it understands our situation.

Thoughts on this? Would really love feedback / any corrections on this thinking! Thanks man.