#FunSearch

2024-01-22

I have thought that Google's apparent lagging in applying #ML to its services has been partly the result of being realistic: #llms just is not a good model for search, because of hallucinations, and Google engineers know this.

But Deep Mind keeps finding ways to make ML useful: #alphafold is now showing some potential worth through collaborations with Lilly and Novartis.

#funsearch is showing that the combination of traditional machine reasoning (good, old-fashioned AI, or #gofai) might work as a proof-assistant, and now they've combined symbolic reasoning with a LLM to tackle International Math Olympiad geometry problems :

deepmind.google/discover/blog/

Here, the LLM suggests adding constructs to the problem, and the reasoner explores the implications of that construct for solving the problem.

One cool thing: the proofs produced are perfectly readable and human verifiable.

Another cool thing: it trains itself by producing random diagrams, deriving relationships (I suspect there's a lot of algebra involved in this step), then producing proofs for these relationships.

Anyway, this work seems real, and plausibly hype-reduced, and looks like it earns the sobriquet of "#AI"

rexirexi
2024-01-11

technologyreview.com/2023/12/1

ICYMI
LLM discovers solution to long-standing scientific puzzle—producing verifiable new info…“It’s not in the training data—it wasn’t even known.”

combines a LLM called Codey, a version of Google’s PaLM 2 that is tuned on

research scientist at , Alhussein Fawzi: “To be very honest with you, we have hypotheses, but we don’t know exactly why this works. In the beginning of the project, we didn’t know whether this would work at all.”

scinexx - das wissensmagazinscinexx@nrw.social
2023-12-18

KI schreibt Mathe-Programme. Künstliche Intelligenz entwickelt innovative mathematische Lösungsalgorithmen. #Mathematik #KI #AI #FunSearch #kuenstlicheIntelligenz
scinexx.de/news/technik/ki-als

Slash Insiderslashinsider
2023-12-16

slashinsider.com/funsearch-dee

DeepMind's FunSearch, an AI language model with a fact-checking layer, claims to have solved an unsolvable math problem by generating and verifying millions of responses. The breakthrough is seen as a promising paradigm for solving complex mathematical challenges.

Client Info

Server: https://mastodon.social
Version: 2025.04
Repository: https://github.com/cyevgeniy/lmst