Fascinating discovery by artificial intelligence researchers using chatbots to solve math problems

Artificial intelligence researchers have claimed the world's first scientific discovery using a large language model. A breakthrough that shows the technology behind ChatGPT and similar programs can produce information beyond normal human knowledge.

This finding comes from Google DeepMind. Currently, scientists are investigating whether large language models, which form the basis of modern chatbots such as OpenAI's ChatGPT and Google's Bard, can be improved by making new changes to the learning process rather than rewriting the learning information. Achieve new performance and achieve new results?

“When we started this project, there was no indication that a truly innovative result could be achieved,” says Pushmeet Kohli, head of the AI ​​research team at DeepMind. “We have, this is the first time that a truly innovative scientific discovery has been made by a large language model.”

“Large-scale language models, or LLMs, are powerful neural networks that learn language patterns, including computer code, from large amounts of text and other data. Since the sudden rise and fall of ChatGPT last year, the technology has improved broken software and produced content ranging from academic papers and travel itineraries to Shakespeare-style poems about climate change.

But even though chatbots have received a lot of attention, they tend to repeat their learning rather than generate new information, and seem to lead to responses similar to the best coffee shop characters. which may seem attractive and smooth at first, but face serious problems.

To build the “FunSearch” system, which stands for “searching in the function space”, artificial intelligence researchers used an LLM to write solutions to problems as computer programs. This LLM is coordinated with an “evaluator” that automatically ranks programs based on their performance. The best programs are then combined and returned to the LLM to improve. This continuously stimulates the system to transform weak programs into stronger programs that may discover new knowledge.

Artificial intelligence researchers received a better result from the Funsearch system than all existing solutions!

Artificial intelligence researchers used this system to solve two puzzles. The first puzzle was a long-standing and somewhat uncommon challenge in pure mathematics known as the Kapp set problem. This problem is about finding the largest set of points in space such that no three points form a line. FunSearch produced programs that produce larger new cup sets that are better than the best solutions that mathematicians have found so far.

The second puzzle was the issue of packing boxes, which seeks the best methods for packing items of different sizes in containers. This applies to physical objects as well, such as the best way to arrange boxes in a shipping container, but the same mathematics is used in other areas as well, such as scheduling computing jobs in data centers. This problem is usually solved by packing items in the first empty box or in the box with the least amount of space that still has room for items. According to the results published in the journal Nature, FunSearch found a better way than leaving spaces. Small that will probably never be filled, prevented.

“In the last two or three years, there have been some fascinating examples of mathematicians working with artificial intelligence to make progress on unsolved problems,” says Sir Tim Gowers, professor of mathematics at Cambridge University, who was not involved in the research. It gives us another interesting advantage for such collaborations, allowing mathematicians to efficiently search for subtle and unexpected structures. Even better, these structures are human interpretable.”

Artificial intelligence researchers are currently exploring the range of scientific problems that FunSearch can deal with. One of the main limitations is that problems must have a solution that can be automatically verified. In this way, many questions in the field of biology cannot be used by these methods. Because their assumptions often require laboratory tests.