Despite the excitement around the discovery, DeepMind researchers admit they are unsure why the AI was able to generate this solution. The AI's solution to the cap set problem will need further research for verification and to understand how FunSearch arrived at its solution. However, the researchers are optimistic, stating this could be the first time a new scientific discovery has been made by an AI.
Key takeaways:
- DeepMind claims that its AI tool, FunSearch, has generated a new solution to the "cap set problem," a famously difficult math problem, marking the first time AI has made a unique scientific discovery not part of its training data.
- The tool was built on the back of DeepMind's AlphaZero AI and uses a large language model (LLM) called Codey, which is trained on computer code and programmed to reject incorrect answers.
- DeepMind researchers admit they do not fully understand why the AI was able to generate this new solution, stating it was not in the training data and was previously unknown.
- While further research is needed to verify the claims and understand how FunSearch generated its solution, DeepMind's VP of research, Pushmeet Kohli, believes this is the first time a genuine, new scientific discovery has been made by a large language model.