We might not know how the actual distribution works, but we do know how it i s fundamentally structured and designed -- because we did it. We also know that there is something like a representation system inside them. And we also know that human beings do not hold 'internal representations' like any AI system needs to. So there isn't any 'intrinsically magical' in modern AI systems.
None of that helps you understand how exactly LLMs do what they do. Because it describes an interface, not a mechanism.
The inner mechanisms of an LLM are more learned than designed. We know what an LLM does on a low level, but going from that to understanding how they work is like trying to understand how a web browser works by looking at netlists of a CPU. Low level understanding does not grant you high level understanding for free.
But ignoring all of that lets you cling to a very comforting "we understand LLMs because we made them". Ha ha. As if.
> And we also know that human beings do not hold 'internal representations' like any AI system needs to.
Bold fucking claim. Got a source on that?
Because neurobiology has been trying to crack neural representations - the very internal representations brains use - for as long as it existed, and with some success. Both reading and injecting internal representations into the brain is possible now, in narrow cases. The specifics vary region to region, but sparse population coding is a true staple. Today's SOTA for wrangling this mess is ML decoders, and not by a coincidence.
Your analogies about the PC and web browser are not correctly formulated, because in the case of the PC you talk about 'external components' (you should be talking about cpu arch, structure, digital components, interfaces, etc); in the case of the web browser, you should be talking about modules, code, etc.
We do know how LLMs are laid out: layers, att heads, etc. So what we need to look at are the fundamental possibilities of the structure of LLMs, not how the weights are distributed.
> > And we also know that human beings do not hold 'internal representations' like any AI system needs to.
> Bold fucking claim. Got a source on that?
Part of the sources are in the books I mentioned. Nonetheless, you can still fact-check and refute in an adult and serious manner, not in an disrespectful and arrogant way. If my claim sounded arrogant I apologize, but then as I already mentioned, my references back that claim.
Regarding internal representations in the brain: I guess you are referring to areas of the brain being activated when a subject receives a stimuli, and this is tested through MRI. I would be cautious to causally relate stimuli to neuron activations, since you first need to know if the exact configuration of cell involved and their connections allow for such representation (which I think it is still not known -- again, AFAIK, the contrary seems to be the case).
Yeah, no. I'm not walking that chain. If you want to, do it, but for now, I'm filing it as "has no evidence and knows it".
By now, there's plenty of works, up to and including direct neural interfaces. Utah arrays, Michigan arrays. Stab the brain, dump the spike trains, decode. You crack the manifold open by correlating to known stimuli using ML, and generalize from there to unknown stimuli. There is no need to "know the exact configuration", and few bother - you put your hardware into the part of the brain you want (top level map is consistent enough brain to brain), gather a set of reference points, and use them to anchor the rest of the decoding process.
Why use ML? Because you need a very expressive correlator to bridge the gap between known inputs and the products of whatever transformations the brain subjects them to before they show up in spike trains.
> So what we need to look at are the fundamental possibilities of the structure of LLMs, not how the weights are distributed.
And the fundamental possibilities are... what exactly? We know the I/O planes, we know the possible flow of information, now, what does that give us?
We know enough to prove that a transformer LLM can implement a Turing machine, the same way a CPU can implement a Turing machine. So an LLM is capable of performing arbitrary computation within its capacity. That's it. That's the upper bound.
What follows is: if you can represent "thinking" as a computational process, you can implement it with a Turing machine, and thus, an LLM can be made to think. That proves LLMs can think. But not that the existing ones do or don't! Because that's the entire thing about upper bounds!
We've looked at LLM architecture, and learned basically nothing about whether LLMs think, other than "it's not impossible". That's the actual "fundamental possibilities" we derived from knowing the architecture. One step above worthless. Oh fun.
(If thinking requires hypercomputation, then, nope. LLMs are out. Good luck proving that it does though.)
