Google-owned DeepMind Claims AI Breakthrough in Tackling Unsolved Mathematics
DeepMind, the AI research lab owned by Google, has announced a major breakthrough in the field of unsolved mathematics. The company claims that its AI model, FunSearch, can find new answers to complex mathematical problems by addressing a crucial issue faced by large language models (LLMs) – the generation of factually incorrect information, commonly known as hallucinations.
To combat this problem, DeepMind’s FunSearch incorporates an automated evaluator that safeguards against hallucinations and incorrect ideas. This unique feature sets FunSearch apart from other popular AI models that have faced controversy and legal action due to spreading false information.
DeepMind put FunSearch to the test by applying it to the unsolved cap set problem, which involves determining the largest size of a particular set. Remarkably, FunSearch discovered new constructions of large cap sets that surpass the previously known best ones.
Moreover, the company showcased FunSearch’s practical usefulness by employing it to uncover more effective algorithms for the bin-packing problem. This problem has widespread applications in optimizing data centers for improved efficiency.
The FunSearch AI model comprises an automated evaluator and a pre-trained LLM, which aims to generate creative solutions to problems. DeepMind highlights the synergistic interaction between these two components, creating an evolutionary approach to problem-solving.
To present problems to the AI model, they are encoded as code, complete with a program evaluation procedure and a seed program that initializes a program pool. FunSearch then selects specific programs and enhances them, building upon their existing structures in a creative manner.
The results are evaluated, and the most promising programs are reintroduced into the pool, forming a self-improving loop. This iterative procedure enables FunSearch to constantly refine its approach and produce innovative solutions.
DeepMind believes that FunSearch’s success demonstrates the potential of protecting against LLM hallucinations. The company asserts that harnessing the power of these models can lead not only to new mathematical discoveries but also significant solutions to real-world problems.
This latest breakthrough adds to DeepMind’s growing list of AI achievements. In 2020, the company made headlines when its AlphaFold model accurately predicted the structures of over 200 million proteins. Furthermore, DeepMind claims its GraphCast AI model can predict weather conditions more accurately than traditional industry methods, while another AI model has been instrumental in creating hundreds of new materials in laboratory settings.
DeepMind continues to push the boundaries of AI innovation, revolutionizing various scientific fields. As its AI models unlock new possibilities, the potential for groundbreaking discoveries and practical applications grows exponentially.
Stay informed with the latest news and developments by subscribing to our Daily Brief – your essential source for science and technology updates.
