4 ms·
> I think a lot of the misconceptions about these LLMs (which I actually think are pretty cool and useful) stem from not really understanding what they are doin
by Closi 3y ago
> I think a lot of the misconceptions about these LLMs (which I actually think are pretty cool and useful) stem from not really understanding what they are doing and how they arrive at their answers.
I think the misconception is actually the opposite - focussing too much on the method rather than the output. Bread is not just yeasty-wheat - if you mix bread and wheat together and apply some heat you get something a little surprising. I think the same is true for LLM's - in the task of training a next-word guesser there are emergent capabilities which end up going beyond what people expect when looking at the method.
Yes, it is designed to statistically guess the next word, but in order to do that the LLM has (surprisingly) gained internal representations of things like 'what is a giraffe'. In order to do the task of 'what is the next word' most effectively, it has had to understand/learn about about the world.
Can you provide an example where you think GPT-4 performs badly that proves your point? (i.e. something not in the training corpus that I could probably answer but a leading LLM could not)
- svieira 3y agoJesting Pilot's question is one, for a start. "Quid est veritas" can only be answered by one who sees.
- Closi 3y agoThat’s not a specific example question I can answer… and if it is I don’t understand it? If the question is just “what is truth?” then that’s something that’s something that’s pretty heavily in the training corpus and also something an LLM can pretty easily define.
- svieira 3y agoDoes it define it correctly? Can it tell? Can you?
- Closi 3y agoWell GPT responds by stating that the concept of truth is a complex and ambiguous philosophical concept, and that there are multiple definitions depending on the context (e.g. mathematical truth is different to scientific truth which is different to pragmatic/constructionist philosophical theories of truth etc). It provides a much better, more balanced and more refined definition of truth that I can manage... "Can it tell?" implies sentience - which I don't believe is required for AGI.
- flimsypremise 3y agoWhat exactly in that reply makes you think that GTP-4 has an internal representation of "what is a giraffe"? You basically fed it the context in which to answer the question in your prompt, by asking it to consider a world were giraffe brains could be networked. The creative bit came from the human, the LLM model is a recitation of factual information that is present in the training data.