FunSearch was tested on the 'cap set problem', a problem evolved from the game Set, and was able to improve on the lower bound for n = 8, generating sets of cards that satisfy all the requirements. The system allows people to see the successful programs created by the LLM and learn from them, setting it apart from other AI applications that function as a black box. This has been highlighted as a new mode of human-machine collaboration, not to replace human mathematicians, but to enhance their capabilities.
Key takeaways:
- An AI system called FunSearch, developed by Google DeepMind, has made progress in solving complex mathematical problems inspired by the card game Set. This is the first time a large language model (LLM)-based system has been able to generate solutions beyond what was previously known by mathematicians and computer scientists.
- FunSearch works by creating requests for a specially trained LLM to write short computer programs that generate solutions to mathematical problems. The system then checks if these solutions are better than known ones, providing feedback to the LLM for improvement.
- The system was tested on the 'cap set problem', a complex version of the game Set. FunSearch was able to improve on the lower bound for n = 8, generating sets of cards that satisfy all the requirements, a solution that goes beyond what was known before.
- Unlike other AI applications that function as a 'black box', FunSearch allows people to see the successful programs created by the LLM and learn from them, promoting a new mode of human-machine collaboration.