@ju916 @41414141 @halvarflake If you look at an LLM, then I'd say low-level is rather something like the maths behind it (some linear algebra for instance, and the whole theory behind ML). If you look at a compiler, some automata theory will also help.
@teh_gerg @41414141 @halvarflake that's exactly the point. from what I know (and I might be wrong) we don't have any low level understanding of how a LLM is doing it's "magic". We can't follow the input on its way through the system - it is essentially a black box. So what are the low-level capabilities we need?
@ju916 @41414141 @halvarflake I know what you mean, but my point of view here is a bit different. Our current lack of debugging/inspection capabilities does not mean that we don't understand the magic on a low level. I'd say that when it comes to debugging/inspection, we're currently facing scale issues.
@teh_gerg @41414141 @halvarflake okay - so maybe it's rather that this level is too low. Like: the laws of quantum mechanics do not really help you to understand weather.
@ju916 @41414141 @halvarflake That could be true, but I'm not in the position to argue from a neuroscience perspective; my view is heavily influenced by maths etc., so it might be natural for me to argue from that angle.
@ju916 @41414141 @halvarflake I do like your QM vs. weather example btw. I could now say that while QM will not help you as much here, studying the Navier-Stokes equation _will_ (I believe; not an expert on weather either). I'd argue in the same direction when it comes to machine learning.
@teh_gerg @41414141 @halvarflake if I remember my physics courses correctly they won’t. Because weather is best described as a chaotic system where minor deviations can lead to diverging states.
@ju916 @41414141 @halvarflake Hm. In what sense do you mean that? I've briefly tried to check some sources (e.g., diva-portal.org/smash/get/diva…), which seem to indicate that Navier-Stokes is quite relevant here.
@ju916 @41414141 @halvarflake It can be used to describe some aspects of emergent phenomena - and I suspect there's similar approaches for understanding LLMs. Which also show some emergent phenomena, that we'd like to get better handle on.
@ju916 @41414141 @halvarflake But I now see what you mean. You want to be able to study the emergent properties of LLMs (or similar systems). I initially assumed you were interested in lower-level aspects of such systems.
@teh_gerg @ju916 @halvarflake In short, we are doing the same mistake we did the first time with ignoring Turing: We build something we can provably not predict. #FML