Google DeepMind Uses Large Language Model to Crack Unsolved Math Problem

In a momentous achievement, Google DeepMind has successfully utilized a large language model to solve an unresolved mathematical problem, marking a groundbreaking moment in artificial intelligence research. This significant accomplishment represents the first instance in which a large language model has been employed to unravel a long-standing scientific enigma, generating valuable new knowledge that was previously unknown. Pushmeet Kohli, vice president of research at Google DeepMind and coauthor of the study, emphasized the importance of this breakthrough, highlighting that the discovered solution was not present in the model’s training data, signifying a significant advancement in the field of artificial intelligence.

The tool responsible for this remarkable accomplishment, affectionately known as FunSearch, challenges the traditional perception of large language models as generators of fabricated content by demonstrating its potential to make genuine discoveries. This groundbreaking potential has reshaped the prevailing narrative, providing insight into the untapped capabilities of large language models.

FunSearch, unlike its predecessors, AlphaTensor and AlphaDev, is not limited by predefined problem-solving methodologies, allowing for a more adaptable approach. By integrating a large language model, Codey, with systems designed to eliminate erroneous outcomes and retain valuable solutions, FunSearch has showcased its unprecedented capacity to uncover complex solutions to perplexing problems in mathematics and computer science.

The development of FunSearch represented a departure from established problem-solving methodologies, with Alhussein Fawzi, a research scientist at Google DeepMind, acknowledging the initial uncertainty surrounding the feasibility of the project. However, despite initial skepticism, the research team pursued a novel approach, leveraging FunSearch to derive solutions by employing Codey to systematically complete programming tasks, leading to the generation of previously unknown correct solutions to intricate mathematical problems.

One such instance of FunSearch’s success is its ability to provide a solution to the cap set problem, a cryptic mathematical puzzle that has baffled mathematicians for years. This solution, previously elusive and the subject of ongoing debate within the mathematics community, represents a significant breakthrough and a testament to the potential of large language models in addressing complex problems.

Furthermore, FunSearch’s versatility was further validated through its successful application to the bin packing problem, a challenging mathematical dilemma with far-reaching applications in computer science. By producing a solution that surpassed existing human-devised methods, FunSearch has underscored the transformative potential of large language models in tackling diverse computational challenges.

The implications of FunSearch’s achievements extend beyond the realm of mathematics and computer science, with prominent mathematicians expressing optimism regarding the emerging paradigm. Terence Tao, a distinguished mathematician and influential figure in the field, commended the promising prospects of leveraging large language models and their innovative capacity to unlock new frontiers in problem-solving methodologies.

As the scientific community continues to explore the integration of large language models into diverse research domains, FunSearch serves as a testament to the immense potential of AI in reshaping our understanding of complex mathematical and computational problems.