News for hackers. Stuff that matters.
Re:AlphaGeometry: An Olympiad-level AI system for geometry by killerstorm on Wednesday January 17, 2024 @03:24PM (#39032938)
> it's still pretty striking how far away the structure of it is from the usual idea of automated reasoning/intelligence.

How so? Reasoning is fundamentally a search problem.

The process you described is exactly the process humans use: i.e. make a guess about what's useful, try to work out details mechanically. If get stuck, make another guess, etc. So the process is like searching through a tree.

People figured out this process back in 1955 (and made a working prototype which can prove theorems): https://en.wikipedia.org/wiki/Logic_Theorist but it all hinges on using good 'heuristics'. Neural networks are relevant here as they can extract heuristics from data.

What do you think is "the usual idea of automated reasoning"? Some magic device which can solve any problem using a single linear pass?

Re:AlphaGeometry: An Olympiad-level AI system for geometry by somewhereoutth on Wednesday January 17, 2024 @04:57PM (#39034105)
> Some magic device which can solve any problem using a single linear pass?

You mean like a calculator?

Re:AlphaGeometry: An Olympiad-level AI system for geometry by gowld on Wednesday January 17, 2024 @05:15PM (#39034328)
A calculator that can solve any problem human could, and more? Sure.
Re:AlphaGeometry: An Olympiad-level AI system for geometry by jhrmnn on Wednesday January 17, 2024 @05:04PM (#39034174)
I think the intuition here is that an “intelligent” system would lean much more heavily on the heuristics than on the search, like humans. It’s hard to quantify in this case, but certainly in the case of chess, when engines were about as good as best human players, they were doing orders of magnitude larger search than humans. Which made them feel more like chess _engines_ than chess AI. AlphaZero certainly made a step in the right direction, but it’s still far from how humans play.
Re:AlphaGeometry: An Olympiad-level AI system for geometry by killerstorm on Wednesday January 17, 2024 @05:25PM (#39034437)
The comment above mentions "brute force" incorrectly. It's fundamentally impossible to brute force googol combinations...
Re:AlphaGeometry: An Olympiad-level AI system for geometry by nybsjytm on Wednesday January 17, 2024 @08:30PM (#39036320)
The search space here is not so large. Read this paper, a key part of the present work: https://doi.org/10.1023/A:1006171315513
Re:AlphaGeometry: An Olympiad-level AI system for geometry by tsimionescu on Thursday January 18, 2024 @02:17AM (#39038930)
The key insight is this whole thread is that this Alpha Geometry only works because the search field is not a googol combinations. So, it doesn't really generalize to many other fields of math. We shouldn't expect an AlphaCategoryTheory or AlphaNumberTheory anytime soon.
Re:AlphaGeometry: An Olympiad-level AI system for geometry by killerstorm on Thursday January 18, 2024 @06:45AM (#39040613)
"Brute force" = going through all combinations indiscriminately.

Using heuristics, beam search, etc, = "not brute force".

Calling something smart "brute force" is wrong.

> We shouldn't expect an AlphaCategoryTheory or AlphaNumberTheory anytime soon.

There's already a number of papers demonstrating use of LLMs for math in general:

"Autoformalization with Large Language Models" https://arxiv.org/abs/2205.12615

"Large language models can write informal proofs, translate them into formal ones, and achieve SoTA performance in proving competition-level maths problems!" https://twitter.com/AlbertQJiang/status/1584877475502301184

"Magnushammer: A Transformer-based Approach to Premise Selection" https://twitter.com/arankomatsuzaki/status/16336564683345879...

I find it rather sad that the reaction to amazing research is "but it's just..." and "it might not work on ...". Like, maybe appreciate it a bit and consider a possibility that something similar to it might work?

Re:AlphaGeometry: An Olympiad-level AI system for geometry by tsimionescu on Thursday January 18, 2024 @11:06AM (#39043433)
AlphaGeometry works, in simplified terms, in two stages:

1. Heuristically add a candidate construction using an NN (a transformer)

2. Brute force search through all possible deductions from the current set of constructions using a symbolic solver

If you don't find a solution after step 2, repeat. There may be some backtracking involved to try a different set of constructions as well.

This approach only works because, in geometry, the set of possible deductions from a given construction is actually quite small.

Also, note that this approach overall is essentially an optimization, not amazing new capabilities. Replacing step 1 with a random construction still solves 10 problems on average in the given time, compared to 30 with the new approach. The existing algorithm, relying mostly on brute force search, is probably able to solve all of the geometry problems if given, say, 10 times as much time as the olympiad students (so not some absurd amount of time).

Re:AlphaGeometry: An Olympiad-level AI system for geometry by hxypqr on Saturday January 20, 2024 @01:38AM (#39065275)
Key insight is the finiteness of reasoning parts in planar geometry that can be quickly solved by the SAT, which often does not exist in most first-order and second-order logics, such as number theory, algebra, or functional analysis
Re:AlphaGeometry: An Olympiad-level AI system for geometry by amelius on Wednesday January 17, 2024 @05:13PM (#39034291)
Well, what's different is that humans invent new abstractions along the way such as complex numbers and Fourier transforms.
Re:AlphaGeometry: An Olympiad-level AI system for geometry by MrPatan on Wednesday January 17, 2024 @05:16PM (#39034333)
Not a lot of humans do
Re:AlphaGeometry: An Olympiad-level AI system for geometry by skepticATX on Wednesday January 17, 2024 @07:46PM (#39035908)
Every single human has abstractions that are unique to them. Your world model isn’t the same as mine.

It’s just that usually these abstractions are fuzzy and hard to formalize, so they aren’t shared. It doesn’t mean that they don’t exist.

Re:AlphaGeometry: An Olympiad-level AI system for geometry by sdenton4 on Thursday January 18, 2024 @10:40AM (#39043075)
This is laughable... Neural networks also have basically unknowable abstractions encoded in their weights. There was some work not long ago which taught an ANN to do modular arithmetic, and found that it was doing Fourier transforms with the learned weights....
Re:AlphaGeometry: An Olympiad-level AI system for geometry by amelius on Wednesday January 17, 2024 @05:21PM (#39034395)
When a human has done the same thing many times they tend to try to generalize and take shortcuts. And make tools. Perhaps I missed something but I haven't seen a neural net do that.
Re:AlphaGeometry: An Olympiad-level AI system for geometry by mitthrowaway2 on Wednesday January 17, 2024 @07:49PM (#39035939)
Is that very different than the distillation and amplification process that happens during training? Where the neural net learns to predict in one step what initially required several steps of iterated execution.
Re:AlphaGeometry: An Olympiad-level AI system for geometry by WalterSear on Wednesday January 17, 2024 @08:22PM (#39036244)
IMHO, yes. It's not an (internal) invention occurring as a result of insight into the process - it's an (external) human training one model on the output of another model.
Re:AlphaGeometry: An Olympiad-level AI system for geometry by littlestymaar on Wednesday January 17, 2024 @05:35PM (#39034548)
True, but we're talking about Olympiad level math skills here.
Re:AlphaGeometry: An Olympiad-level AI system for geometry by gowld on Wednesday January 17, 2024 @05:15PM (#39034321)
A neural network is nothing but a heap of new abstractions from data.