Artificial intelligence (AI) researchers at Google DeepMind have reported a groundbreaking achievement—the first scientific discovery facilitated by a large language model (LLM). This development challenges the conventional boundaries of AI capabilities, suggesting that models like OpenAI's ChatGPT can go beyond information repackaging to generate entirely novel insights.
Head of AI for Science at DeepMind, Pushmeet Kohli, expressed the unexpected nature of the discovery, highlighting that it is the first time a genuine scientific breakthrough has been attributed to a large language model.
Large language models, including ChatGPT, are neural networks proficient in learning language patterns, including coding, from vast datasets. While widely used for tasks such as content creation and software debugging, they have been criticized for not generating new knowledge. DeepMind's project, "FunSearch," deploys an LLM to compose computer programs addressing problems. An evaluator ranks these programs based on performance, leading to a continuous evolution of more powerful programs capable of uncovering fresh knowledge.
FunSearch's success was evident in solving puzzles like the cap set problem in pure mathematics and the bin packing problem. Surpassing mathematicians' current best solutions, FunSearch introduces new perspectives on longstanding challenges.
Professor Tim Gowers of Cambridge University praised the collaboration between human mathematicians and AI, envisioning a valuable tool for efficient problem-solving. However, researchers acknowledge FunSearch's limitations in handling scientific problems requiring manual verification, particularly in biology.
The implications of this breakthrough extend beyond academia, with potential transformative impacts on computer science and algorithmic discovery. Kohli sees LLMs not taking over but assisting in pushing the boundaries of algorithmic possibilities.
Professor Jordan Ellenberg of the University of Wisconsin-Madison highlighted the broader significance, emphasizing that FunSearch generates a program that human beings can read and interpret, leading to ideas for future problem-solving.
As the AI community explores FunSearch's capabilities, the future of human-machine collaboration in mathematics and computer science appears promising. This achievement opens new avenues, suggesting a future where AI augments human creativity and problem-solving abilities.
