Scientists managed to make a discovery of mathematics with the help of DeepMind

Scientists managed to make a discovery of mathematics with the help of DeepMind

Artificial intelligence researchers claim to have made the world’s first scientific discovery using a large language model. This discovery suggests that the technology behind ChatGPT and similar programs can generate information beyond human knowledge.

This discovery was made at Google DeepMind, where scientists are studying whether the large language models that underpin modern chatbots such as OpenAI’s ChatGPT and Google’s Bard can do more than just process information learned through learning, and come to new discoveries.

“When we started the project, there was no indication that it would produce anything really new,” says Pushmeet Kohli, head of AI science at DeepMind. “To our knowledge, this is the first time that a large language model has made a truly new scientific discovery.”

Large-scale language models, or LLMs, are powerful neural networks that learn patterns in language, including computer code, based on vast amounts of text and other data. Since ChatGPT’s launch last year, the technology has been fixing broken software and creating everything from college essays and travel itineraries to Shakespearean poems about climate change.

But while chatbots have proven extremely popular, they don’t generate new knowledge and often spout nonsense, resulting in verbose answers that look like the truth but are often delusional.

To create “FunSearch”, short for “searching in the function space” [поиск в функциональном пространстве], DeepMind used LLM to write problem solutions in the form of computer programs. LLM works in tandem with an “evaluator” that automatically ranks programs according to their level of effectiveness. The best programs are combined and fed back to the LLM for improvement. This forces the system to constantly improve bad programs, turning them into more powerful ones capable of discovering new knowledge.

The researchers launched FunSearch to solve two puzzles. The first is an ancient and somewhat tricky problem in affine geometry, known as the problem of many covers. It is related to finding the largest set of points in space in which three points do not form a straight line. FunSearch has created programs that generate new large sets that surpass anything that mathematicians have come up with.

The second puzzle was the task of packing containers, which allows you to find optimal ways to pack objects of different sizes into containers. Although it applies to physical objects, for example, the most efficient way to place boxes in a shipping container, the same mathematics is applicable in other areas, for example, when planning computing tasks in data centers. Usually the problem is solved either by packing the items in the first container that has room, or in the container with the smallest free space that will still fit the item. According to the results published in the journal Nature, FunSearch found the best approach, which avoids leaving small gaps that are unlikely to ever be filled.

“In the last two or three years, there have been some interesting examples of human mathematicians working with artificial intelligence to achieve breakthroughs in solving unsolved problems,” says Sir Tim Gowers, a professor of mathematics at the University of Cambridge who was not involved in the research. “This work potentially gives us another very interesting tool for such collaborations, allowing mathematicians to efficiently search for clever and unexpected designs. Even better, these designs are amenable to human interpretation.”

Currently, researchers are studying the range of scientific problems that FunSearch can handle. A major limiting factor is that problems must have solutions that can be tested automatically, which precludes many questions in biology where hypotheses often need to be tested through laboratory experiments.

More immediate consequences may be programmers. Over the past 50 years, coding has mostly improved due to people creating more and more specialized algorithms. “It’s really going to be transformative in how people approach computer science and algorithmic discovery,” says Kohli. “For the first time, we see that masters are not taking responsibility, but are definitely helping to push the boundaries of what is possible in algorithms.”

Jordan Ellenberg, professor of mathematics at the University of Wisconsin-Madison and a co-author of the paper, said: “What I find really exciting, even more than the specific results we found, are the prospects this opens up for the future of human-machine interaction in mathematics.” .

“Instead of generating a solution, FunSearch generates a program that finds a solution. Solving a particular problem may not give me any idea how to solve other related problems. But a program that finds a solution is something that a person can read, to interpret and, hopefully, in this way generate ideas for subsequent tasks.”

Related posts