Post

this is really cool

Okay, I guess that's interesting as an experiment, but you just showed us a video of it not answering the prompt correctly. The prompt reads, "Who are all the grandfathers"? The "Solution" identifies one grandfather (Orville) but not the other (Abe).

github.comGitHub - FFrydman/eXa-LM: eXa-LM: A Controlled Natural Language Bridge between Large Language Models and Logic SolverseXa-LM: A Controlled Natural Language Bridge between Large Language Models and Logic Solvers - FFrydman/eXa-LM

Also, for whatever it's worth, in a few minutes of searching I found several papers and GitHub repos regarding using LLMs as a natural language bridge to actual logic solvers. For example: github.com/FFrydman/eXa...

You seem like you've got a pretty decorated history with tech companies. If you know anyone looking for a dev who knows both ML/NLP and also broader AI topics like logic solvers and planning systems, who can cut through the hype and make things like this actually work, hmu.

Yeah, I decided just to record a quick video. Didn't realize this many people would reshare it. If you look closely, your screenshot shows "step 16 of 84". I can assure you that by step 34 it finds all the grandpa relations

Gotcha, fair enough. The part I shared an image of was the last thing shown in the video, which it paused on, so it seemed to be the final result.

But I should also just clean up the code and open source it so you can try it out for yourself

github.comGitHub - narphorium/nl-logic-interpreter: A step-through logic interpreter for natural-language facts and rules, unified with JevA step-through logic interpreter for natural-language facts and rules, unified with Jev - narphorium/nl-logic-interpreter

Here it is github.com/narphorium/n...

Amber 🌸

@shimmermathlabs.com

this is amazing! love the look too, it's gorgeous

Thanks Amber!

Have you tried hooking it up to Prolog (or maybe even Lean) to create ontologies so you can gradually offload queries into a more deterministic (and less compute intensive) system?

Not yet. In that case Prolog would be the system one model. What I've been trying to get to is a combination of the search and backtracking algorithm from prolog with the natural language understanding of LLMs but until now its just been too slow

1A ⋈ locria

@1a-insec.net

insane. where is the full article with the full-steps playback?

Thanks! The code is open source now so you can run it. What would you like to see if I were to do a blog post? github.com/narphorium/n...

1A ⋈ locria

@1a-insec.net

20 Sept

Just explain how it works roughly and give a interactive playback in HTML5. I already know miniKanren, so maybe you want to reference that?

iamwil

@interjectedfuture.com

Wait, I don’t get what Jev’s role is here. It seems like this is a datalog/prolog for reasoning, and Haiku converts question to query for the reasoning system…what does Jev have to decide?

Jev decides if two statements can be unified. So it can do fuzzy matching on natural language statements

Amber 🌸

@shimmermathlabs.com

does it still do prolog style order of matching and stuff with normal backtracking? i want jev to decide which way to go in the resolution

Good idea! 😍

That would be awesome but this version just does simple exhaustive search through the facts. I have some ideas for smarter search. I think it would probably need a proper generative model to predict the best search queries though.

ernesto

@ernesto.bsky.social

a bit of the way toward joe armstrong’s entropy reverser

youtu.be"The Mess We're In" by Joe ArmstrongYouTube video by Strange Loop Conference

Great talk youtu.be/lKXe3HUG2l4?...

To match a goal against a fact or a rule, the interpreter asks Jev questions like: Could "Abe is the father of Y" and "Abe is Homer's father" state the same fact? Jev answers with a probability. The rest is ordinary proof search, written in code.

Jev doesn't generate text, so I used Claude Haiku to turn the question into a logic query first. After that, this query took 100 Jev calls in 8.5 seconds, for roughly a quarter of a cent.