Form ideas without the use of language.
For example: imagining how you would organize a cluttered room.
For example: imagining how you would organize a cluttered room.
We don't apply that criteria when evaluating animal intelligence. We sort of take it for granted that humans at large do that, but not via any test that would satisfy an alien.
Why should we be imposing white-box constraints to machine intelligence when we can't do so for any other?
We don't know, but it's completely plausible that we might find that the cost of analyzing LLMs in their current form, to the point of removing all doubt about how/what they are thinking, is also unbearably high.
We also might find that it's possible for us (or for an LLM training process itself) to encrypt LLM weights in such a way that the only way to know anything about what it knows is to ask it.
You can't have it both ways. If your test for whether something is intelligent/thinking or not isn't applicable to any known form of intelligence, then what you are testing for is not intelligence/thinking.
Don't LLMs already do that? "Language" is just something we've added as a later step in order to understand what they're "saying" and "communicate" with them, otherwise they're just dealing with floats with different values, in different layers, essentially (and grossly over-simplified of course).
Humans got by for millions of years with our current biological hardware before we developed language. Your brain stores a model of your experience, not just the words other experiencers have shared with yiu.
I guess if we limit ourselves to "one-modal LLMs" yes, but nowadays we have multimodal ones, who could think of a duck in the way of language, visuals or even audio.
If you didn't know the word "duck", you could still see the duck, hunt the duck, use the ducks feather's for your bedding and eat the duck's meat. You would know it could fly and swim without having to know what either of those actions were called.
The LLM "sees" a thing, identifies it as a "duck", and then depends on a single modal LLM to tell it anything about ducks.
I don't think you can generalize like that, it's a big category, not all multimodal models work the same, it's just a label for a model that has multiple modalities after all, not a specific architecture of machine learning models.