Google DeepMind has achieved a groundbreaking feat by unraveling a longstanding mathematical enigma through a sophisticated approach known as FunSearch. This methodological triumph centers on the resolution of the renowned cap set problem in the realm of pure mathematics, a conundrum that has perplexed even the most brilliant human mathematicians.
Marking a historic moment, this represents the inaugural instance wherein a protracted scientific puzzle has been unraveled leveraging the capabilities of a Large Language Model (LLM). The monumental discovery, accomplished by the adept team at Google DeepMind, has been meticulously documented in the esteemed peer-reviewed journal Nature.
In their comprehensive peer-reviewed study, the researchers expounded, “To the best of our knowledge, this signifies the maiden scientific revelation—a novel piece of empirically substantiated knowledge pertaining to a notorious scientific problem—facilitated by an LLM.”
The key innovation lies in the amalgamation of a pre-trained LLM with an automated “evaluator.” Large Language Models, exemplified by the likes of GPT-4, have showcased unparalleled prowess in tackling intricate problems necessitating quantitative reasoning and the generation of predictive text. These models exhibit the capacity to assimilate vast quantities of information, generating responses indicative of a profound comprehension across diverse domains.
FunSearch, elucidated by authors Alhussein Fawzi and Bernardino Romera Paredes, research scientists at Google DeepMind, employs a dynamic synergy between a pre-trained LLM and an automated evaluator. This collaboration endeavors to transcend the limitations inherent in existing LLM-based approaches. Their joint mission is to deliver inventive solutions in the form of computer code while mitigating the risks of hallucinations and inaccuracies.
Despite the remarkable capabilities of LLMs, they are not without their imperfections, often manifesting as confabulations or hallucinations—instances where the model produces responses that, though plausible, are factually inaccurate. Such inaccuracies pose challenges to the utilization of LLMs in scientific discovery, where precision and reliability are paramount.
The FunSearch methodology introduces a continuous iterative interplay between the LLM and the evaluator. This iterative dialogue metamorphoses initial solutions into novel knowledge, thereby fostering innovation. Contrary to the misnomer, the “Fun” in FunSearch does not denote amusement but rather ‘functions,’ emphasizing the core pursuit of identifying and decoding functions in computer language.
Fawzi and Paredes underscored the application of FunSearch in uncovering more efficient algorithms for the “bin-packing” problem—a complex optimization challenge involving the efficient allocation of items of varying sizes into a finite number of bins or containers, each with predefined capacity. The primary objective of this computational problem is to minimize the total number of bins employed in the packing process, enhancing the efficiency of data centers, as emphasized by the researchers.