AI Researcher. Co-founder & CEO, Corto. Director of Project on AI in Media @ USC's Entertainment Technology Center. Member of @columbiaDSL. GenAI @TheAcademy
Thanks to @benfritz for quoting Corto data on the fast growth of the anime genre. Corto's Geist tool knows where culture is ... and where it's going https://t.co/ohj8o5YkKI via @WSJ
New paper: "Large Language Models & Emergence: A Complex Systems Perspective" (D. Krakauer, J. Krakauer, M. Mitchell).
We look at claims of "emergent capabilities" & "emergent intelligence" in LLMs from perspective of what emergence means in complexity science.
⬇️
Hmm, I find current reasoning LLMs incredibly useful as math and programming and research assistants. However the limitations that this Apple paper notes have been obvious to me from the start of my interaction with them.
The way I would put it conceptually is: these models are amazing and amazingly useful, but if the complexity intervening between their knowledge and your problem is too much, they will fail. E.g. they are good at doing some quite complex math so long as it's basically in the domain of math they already know. Same with programming. But if you give them math or programming in a too far out of the mainstream domains they're trained on, they will flail around.
So they tend to mess up e.g. with paraconsistent logic or MeTTa programming because these things are a bit weird compared to the bulk of their training data. But they're pretty good at category theory, information theory etc. as there is a lot of relevant training data there.
In terms of reasoning LLMs as a component of AGI systems, what this means is they may be great helpers for domains that are already fairly well known to the AGI systems but not for exploring wild unknown territory. I.e. they are for exploitation (in the AI sense, though yeah not only..) not so much exploration.
Anyway there is a LOT of useful reasoning to be done close to the spheres of accumulated knowledge.
But for reasoning about highly novel situations in games, or about new forms of math or programming, we will need systems that can actually reason in a more GI-ish way....
Hybridizing these reasoning models with other systems like oh say Hyperon's Probabilistic Logic Networks that can actually do full-on logical reasoning -- regardless of how far it leaps ahead of prior knowledge, seems more promising as an AGI strategy... but we (or well my colleagues and I) already knew that...
In short -- modern reasoning LLMs as potential central hubs for AGI systems... -- NAY ... modern reasoning LLMs as reasoning helpers to humans or AGIs dealing with familiar domains where there's loads of training data -- YAY
@PeterMoskos I REALLY love “BFTB” by the way. I recommended it to a few senior Hollywood studio executives to learn how to get stuff done. The interview style narrative is really fun and brings it all close to home. Brilliant.
@PeterMoskos I don’t have a point, just wanted to reference this fact, which always blows my mind. 14k is the official number, lots of cops work private security and (even though they get paid) rarely show up for work. I lived in Karachi. It’s a tragic and incredible place.
I am surprised that AI researchers & people thought that language models are able to reliably explain their own outputs.
It is well known that they just chain together token sequences based on statistical relations conditioned on prompt.
They don't understand what they output.