I half-jest but I envision the direction of LLM research to head towards a parser-oriented setup where LLMs merely extract the entities and relations and the actual logic is done by a logical engine such as Prolog.
https://m.youtube.com/watch?v=PdE-waSx-d8
He’s very much on this kind of beat. In general I have a feeling that there are orders of magnitude to gain by successfully applying computer science to “language algorithms”.
Feels like we are exploring very narrow paths of computations that can be performed with language. Like we have an x86 cpu and we are building pocket calculators. So much untapped
Re prolog I had a similar intuition at some point and tried to make a stack based programming language that uses a language model as a kind of control/logic unit
https://github.com/LachlanGray/silas
I was missing a bunch of cs background at the time so I didn’t get very far, but I feel like there’s a lot to be done for this kind of thing
Combining LLM’s with rewriting logic, like Maude or K Framework, would be the most, powerful option. The rewriting tools plus LLM’s could probably rapidly develop static analyzers and code porting tools.
Quote:
One Cyc application aims to help students doing math at a 6th grade level, helping them much more deeply understand that subject matter... Unlike almost all other educational software, where the computer plays the role of the teacher, this application of Cyc, called MathCraft, has Cyc play the role of a fellow student who is always slightly more confused than you, the user, are about the subject.
Cyc is the huge reasoning engine. You can call it Prolog, if you want. I won't.
I just gave it to Claude: https://imgur.com/a/fQQOy1d
2) The "action" text gives me such I put on my wizard hat vibes
[1] https://www.cyc.com/mathcraft/
[2] https://www.youtube.com/watch?v=pbrp7MzBDm0It creates foundation for reinforcement learning without human feedback - a missing piece of puzzle.
Simplifying: propose plausible theorem, try to find provable solution, reinforce reasoning/solution path, move proved statement into axioms, repeat.
(super)intelligence has many dimentions. One of less explored ones is exploiting concurrency in thought chains. It's something very un-natural to us, but there is a lot of gain if you're able to branch and collect feedback from dead ends and progress from different directions being taken at the same time.
Graph RAG is an emerging design pattern where factual information is retrieved from knowledge graphs to augment the textual information retrieved from document stores in classic RAG before submitting to an LLM. What you are proposing is essentially using the LLM to build the knowledge graph in the first place. Would be interesting to see the two techniques combined along with some sort of planner/optimizer in the middle.
For example, if you tell it to reply in JSON (and it obeys), you've just constrained its search space in a particular way. There is space for very interesting informal programming that can be done from this perspective, setting up constraints and then allowing inference to solve within them. I've been using this heavily.
When I was first getting deep into LLM stuff a few months ago and contemplating latent space my main characterization was that much of its high level behavior can be usefully grappled with by viewing it as a kind of 'learned geometric prolog'.
I did a bunch of illustrations and talked about some of these ideas here if anyone's curious: https://x.com/Westoncb/status/1757910205478703277 (I think I mostly dropped the prolog terminology in that presentation because not everyone knows about it)
That feature by itself is not enough, but can be a very effective glue to be used with other components of an intelligent system. The analogy with the human brain would be the broca area vs. the rest of the brain.
Now, there are open questions about whether the _architecture_ that underpins the LLMs is also good enough to be used as a substrate for other functions and what's the most effective way for having these different components of the system communicate between each other.
The analogy with the human brain can guide us (as well as lead us astray), in that our brain, like biological systems often do, re-purposes the basic building blocks to create different subsystems.
It's not clear to me at which level we'll find the most effective re-purposable building blocks.
It's easy to try (and people do) to use the top-level LLM system as such a building block and have it produce plans, connect it to external systems that feed information back and have it iterate again on it (ab)using it's language processing as an API with the environment.
The human analogy of that is when we use external tools to extend our cognitive capacity, like when we do arithmetic using pencil and paper or when we scribble some notes to help us think.
I think this level is useful and real but I wonder if we also need to give more power to some lower levels too.
Granted, some of that "power" can already be emerging during the training of the LLMs but I wonder if some more specialized blocks might enhance the effectiveness