Google DeepMind uses artificial intelligence to solve two mathematical problems – Futura

Researchers at Google DeepMind have managed to show that generative artificial intelligence can produce completely new information. Their FunSearch tool successfully created programs to solve two tricky math problems.

This will also interest you

[EN VIDÉO] Discover the job of engineer for calculation and numerical simulation!

In a paper published in the journal Nature, researchers at Google DeepMind managed to solve two mathematical problems using a large language model (LLM). LLMs, the artificial intelligences that power new chatbots like ChatGPT or Google Bard, are good at repeating already known information, but are also prone to hallucinations. This is the first time that they are producing new information that did not exist before.

The researchers have developed a tool called FunSearch (Functional Search) based on Codey, a version of the Google PaLM 2 LLM trained on computer code, and an algorithm that evaluates the generated code and rejects failed attempts or hallucinations. His first task was to create a program to solve the “cap set” problem, which involves determining the maximum number of points in a graph without them being aligned.

The first “real scientific discovery”

The second task was to solve the bin packing problem, which involves storing items of different sizes in as few boxes as possible. This problem occurs regularly in the real world, for example when filling shipping containers or assigning computing tasks in data centers. After repeating the process millions of times in a loop, FunSearch managed to find original solutions to both of these problems. In the case of the bin packing problem, its solution was even more efficient than the reference heuristic approaches.

“To our knowledge, this is the first time a real scientific discovery has been made using a large language model,” said Pushmeet Kohli, vice president of research at Google DeepMind. Algorithms have been used to solve problems before, but machine learning was specialized in a very specific area. Additionally, instead of offering a solution for a specific situation, FunSearch has developed a program that can be reused for other situations and is understandable to people.