Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Ask HN: What is the current state of "logical" AI?

news.ycombinator.com

Post reply on HN
Print view View original

1–10 of 54 posts

Ask HN: What is the current state of "logical" AI?

#1

Post by mtlb »

The kind of AI that gets the public attention right now lacks a quality that can be described as "formal correctness", "actual reasoning", "rigorous thinking", "mathematical ability", "logic", "explainability", etc.

This is the quality that should be studied and developed in symbolic AI approach. However, the actual symbolic AI work I know of seems to fall in one of the two buckets: 1. "Let's solve a mathematical problem (e.g. winning at chess) and say that the solution is AI" (because humans can play chess, and now computers can too!) 2. "Let's make something like Prolog but with different solver algorithm / knowledge representation". Products like Cyc and Wolfram seem to work essentially in this manner, although with lots of custom coding for specific cases to make them practical. There's lots of work on separate aspects of this as well, like temporal and other modal logics.

I see the first bucket as just applied maths, not really AI. The second bucket is actually aimed at general reasoning, but the approaches and achievements in it are somewhat uninspiring, maybe because I don't know many of them.

So my broad question is: what is happening in such "logical AI" research/development in general? Are there any buckets I missed in the description above, or maybe my description is wrong to begin with? Are there any approaches that seem promising, and if so, how and why?

I would be grateful for suggestions of the books/blogs/other resources on the topic as well.

Re: Ask HN: What is the current state of "logical" AI?

#4

Post by PaulHoule »

See https://cacm.acm.org/magazines/2023/6/273222-the-silent-revo... and also modern production rules engines like https://drools.org/

Oddly, back when “expert system shells” were cool people thought 10,000 rules were difficult to handle, now 1,000,000 might not be a problem at all. Back then the RETE algorithm was still under development and people were using linear search and not hash tables to do their lookups.

Also https://github.com/Z3Prover/z3

Note “the semantic web” is both an advance and a retreat in that OWL is a subset of first order logic which is really decidable and sorta kinda fast. It can do a lot but people aren’t really happy with what it can do.

Re: Ask HN: What is the current state of "logical" AI?

#5

Post by mtlb »

bkovacic wrote: post #2

Gemini Ultra should show good progress according to Google - it's supposed to perform better than 85% of computer science competitors, which requires a lot of logical reasoning. Lets see it once it goes live, but sounds promising.

Their previous model was better than 46% of such competitors (according to them), so 85% seems achievable by throwing more compute resources at typical ML training. After all, training on millions of examples of logical reasoning will undoubtedly store logical rules in the model in some shape or form (it does so even in ChatGPT), yet the results are still more "convincing" rather than "correct", or "probably correct" at best, usually achieved with lots of postprocessing on top. GPT-4 is better than 90% of lawers at the bar exam, yet still manages to fail at reasoning on much simpler domains.

Re: Ask HN: What is the current state of "logical" AI?

#8

Post by mtlb »

PaulHoule wrote: post #4

See https://cacm.acm.org/magazines/2023/6/273222-the-silent-revo... and also modern production rules engines like https://drools.org/ Oddly, back when “expert system shells” were cool people thought 10,000 rules were difficult to handle, now 1,000,000 might not be a problem at all. Back then the RETE algorithm was still under development and people were using linear search and not hash tables to do their lookups. Als…

Thanks, modern rule engines and description logic formalisations are something for me to explore! Are there any other practical applications of such advanced SAT solvers?

Re: Ask HN: What is the current state of "logical" AI?

#10

Post by jschveibinz »

"Formal reasoning" or "logic" as you suggest is a model for finding "truth" from static inputs and simple operations. However, if the inputs are random variables (they have an associated distribution) then so (likely) are the outputs, and "truth" is still a random variable. The world we live in is better modeled by the latter than the former, and as such the "decision tree" approach of AI seems like a more reasonable approach and model to finding "truth" than a strictly mathematical approach.
Post reply on HN