AI Generates New Scientific Discovery: A Breakthrough for Language Models

Artificial intelligence researchers at Google DeepMind have achieved a groundbreaking feat by using a large language model to make the world's first scientific discovery. This development suggests that technology like ChatGPT has the potential to generate information that surpasses human knowledge.

DeepMind's project, known as "FunSearch" (short for "searching in the function space"), explores the capability of large language models (LLMs) to go beyond rehashing existing knowledge and produce new insights. LLMs, such as ChatGPT, are neural networks trained on vast amounts of text and data, enabling them to understand and generate human language.

To create FunSearch, DeepMind paired an LLM with an "evaluator" that ranks computer programs based on their performance in solving problems. The top-performing programs are then combined and fed back into the LLM, gradually improving the quality of generated solutions.

FunSearch was put to the test on two challenging puzzles. The first, the "cap set problem" in pure mathematics, involves finding the largest set of points in space where no three points form a straight line. FunSearch produced programs that generated new, larger cap sets that exceeded previous mathematicians' best efforts.

The second puzzle was the "bin packing problem," which has applications in logistics and scheduling. FunSearch discovered a more efficient approach to packing items into containers, reducing wasted space.

This achievement has significant implications for both mathematics and computer science. Mathematicians can now collaborate with AI to explore new and unexpected solutions to longstanding problems. In computer science, this breakthrough is expected to revolutionize algorithmic discovery by assisting programmers in pushing the boundaries of what is possible.

While this development is promising, it has limitations. FunSearch can only handle problems with verifiable solutions, excluding many questions in biology that require lab experiments for validation.

In summary, the success of FunSearch showcases the potential of LLMs to contribute to scientific discovery and algorithmic advancements, opening up exciting possibilities for human-machine collaboration in various fields.