Exactly this. When you test LLMs for generalization, you can reasonably gurantee whether some specific pattern of tokens falls in its corpus or not. Thus, tasks where you teach it something totally OOD (say, a
koijayttyis a man, andtokopornifis a woman, make a sentence using both words & english) then the LLM should be able to do this simple task if it learnt the basics of what we’re asking - it should first synthesize a meaningful sentence, parse my query to locate the words to swap with, swap them, and then continue along its way.
So how well do they do? I expect humans to solve a task like this by linking these new words to mental models of man and woman, and then using them freely. Can a LLM do as well if all it has to offer is word substitution? Can we tell the difference? Is there one?
So according to the TBT/HTM theory and predictive coding in general, its the opposite - our intelligence is derived somehow by the brain solving to the sole objective of modelling our world. Due to our intellect, we can model our world extremely well - enough to hold such higher level thoughts to model things, which we call ‘abstract’.
[That sentence does not parse well]. IMO all animal brains create mental models of parts of the world based on sensory input and evolved templates. Concepts like near and far, up and down, in, on, over are built in (and get short words). We simply construct more and better models, but it’s the same core thing for us as a mouse or a bird.
In that case, GPT3’s world is the 1 dimensional stream of tokens that it must predict. Thus, it is, by that definition, having some level of intelligence if its able to predict them reasonably well. It’s not perfect, which means its not AGI or atleast HLAI yet.
Agreed, except for one problem. What the hell is intelligence anyway? How do we test for it? How do we quantify it?