Artificial intelligence (AI) researchers at Google DeepMind have announced what they believe to be the world's first scientific discovery facilitated by a large language model (LLM), marking a significant leap in the capabilities of technologies like ChatGPT. Pushmeet Kohli, the head of AI for science at DeepMind, expressed astonishment at the project's unexpected outcome, stating that it represents the first instance of a large language model producing a genuinely new scientific discovery.

Large language models, such as those powering ChatGPT, learn language patterns from vast amounts of text and data. While these models have been widely popular for various applications, they were previously thought to lack the ability to generate new knowledge. However, DeepMind's project, named "FunSearch," challenges this notion by employing an LLM to create computer programs solving problems, subsequently evolving and improving them through an automatic ranking system.

FunSearch tackled two puzzles, including a complex problem in pure mathematics known as the cap set problem, where it outperformed existing human-derived solutions. Additionally, it addressed the bin packing problem, providing a more efficient solution than conventional methods, as reported in Nature.The impact of this discovery extends beyond mathematics, with potential implications for computer programmers. Kohli emphasized the transformative nature of LLMs in pushing the boundaries of algorithmic possibilities, marking a shift in how people approach computer science and algorithmic discovery. While the research showcases the potential of FunSearch, challenges remain, particularly in applying it to scientific problems that require manual verification, limiting its application in certain fields like biology.

Jordan Ellenberg, a professor of mathematics at the University of Wisconsin-Madison and co-author on the paper, expressed excitement about the prospects for human-machine interaction in math, emphasizing that FunSearch's ability to generate programs that find solutions provides a new avenue for human creativity and problem-solving in the future.