In a significant development, DeepMind, a subsidiary of Google, has introduced an innovative Large Language Model (LLM) named “FunSearch,” which is revolutionizing the field of mathematics.
Unlike traditional AI models that occasionally generate inaccurate or fictional results, FunSearch specializes in finding precise solutions to complex mathematical problems, often revealing entirely new solutions never before conceived by humans.
FunSearch: A pioneering mathematical marvel
FunSearch, aptly named for its focus on mathematical functions rather than amusement, is setting new standards in the realm of AI-driven mathematics. At the heart of this groundbreaking model is a two-tiered architecture.
The first layer is a variant of Google’s PaLM 2 called “Codey,” a large language model. The second layer acts as an error-checking mechanism, meticulously scanning Codey’s output and eliminating incorrect information.
DeepMind’s research team, spearheading this exceptional project, embarked on a journey of uncertainty, unsure of whether this approach would yield remarkable results. Even today, they remain mystified about the underlying mechanisms that drive FunSearch’s extraordinary capabilities, according to DeepMind researcher Alhussein Fawzi.
Solving the enigmatic cap set problem
One of the key mathematical conundrums FunSearch tackled is the infamous “cap set problem.” This puzzle has confounded mathematicians for years, primarily due to the lack of consensus on the best approach to solving it.
However, FunSearch has transcended this challenge by generating entirely new and, crucially, accurate solutions to the cap set problem—solutions previously unattainable through human endeavors.
To accomplish this feat, DeepMind engineers constructed a Python representation of the cap set problem, omitting the lines that defined the solution. It was then Codey’s responsibility to add lines that would correctly resolve the issue.
The error-checking layer rigorously assessed Codey’s solutions for accuracy and quality, recognizing that in high-level mathematics, equations might have multiple solutions, but not all are deemed equally valuable. Over time, FunSearch’s algorithm identifies the optimal solutions generated by Codey and integrates them back into the model.
DeepMind allowed FunSearch to operate for several days, during which it produced millions of potential solutions. This extended runtime allowed FunSearch to refine its code and generate increasingly superior results. The results of this research highlight FunSearch’s ability to produce previously unknown, yet mathematically sound solutions to the cap set problem.
Beyond the cap set: Tackling the bin packing problem
In addition to the cap set problem, FunSearch demonstrated its prowess in addressing another formidable mathematical challenge known as the “bin packing problem.” This problem entails determining the most efficient way to pack bins, a task laden with complexity and practical applications. Remarkably, FunSearch outperformed human-calculated solutions by uncovering a faster and more optimized approach.
The capacity of FunSearch to excel in such diverse mathematical domains underscores its potential utility in assisting mathematicians and researchers in various fields.
While the integration of Large Language Models (LLMs) into the realm of mathematics continues to present challenges, DeepMind’s FunSearch offers a promising path forward. What sets this approach apart is its generation of computer code, as opposed to delivering raw mathematical outputs. This distinction simplifies the understanding and verification process, making it more accessible to human mathematicians and researchers.
The advent of FunSearch represents another significant stride in DeepMind’s ongoing contributions to artificial intelligence. Their earlier projects, including AlphaFold (protein folding), AlphaStar (StarCraft), and AlphaGo (Go), achieved remarkable feats but were not based on LLMs. Nevertheless, they uncovered new mathematical concepts, foreshadowing the groundbreaking potential of FunSearch.
As mathematicians grapple with the evolving landscape of LLM technology, DeepMind’s latest innovation offers a glimmer of hope and potential solutions to long-standing mathematical enigmas. FunSearch’s unique approach and its ability to provide novel, verified solutions could reshape the way mathematicians approach complex problems.