DeepMind's FunSearch AI Aims to Solve Complex Math Problems

DeepMind, the renowned AI research company, has heralded one of its AI models, FunSearch, for its potential to resolve intricate mathematical challenges. The FunSearch AI model is designed with an innovative feature — an automated 'evaluator' that critically assesses the model's outputs to mitigate the risk of producing factually incorrect or illogical information, commonly known as 'hallucinations'.

Addressing Hallucinations in AI Models

Large language models (LLMs), such as OpenAI's ChatGPT, have often been under scrutiny for their tendency to generate false statements. This limitation has even led to legal actions in some instances. DeepMind's FunSearch, however, is said to counteract this propensity by incorporating the evaluator that screens for and prevents such errors, thereby enhancing the reliability of its solutions.

Evaluating the Unsolvable

DeepMind has put its AI to the test with the enigmatic cap set problem, a notorious problem in mathematics that revolves around determining the maximum size of a specific set configuration. The company asserts that FunSearch not only addressed this problem but also formulated new, more considerable cap set constructions than those previously known.

The AI's capabilities extend beyond theoretical mathematics. DeepMind has highlighted the practical application of FunSearch in optimizing algorithms for the 'bin-packing' problem, which offers significant benefits, such as improving efficiency in data centers.

The Method Behind FunSearch

FunSearch consists of a pre-trained LLM along with its automated evaluator. This combination is engineered to foster 'creative solutions' by establishing a dynamic, evolutionary process to sift through and enhance problem-solving methods.

Problems are fed to the AI as code comprising evaluation procedures and an initial seed program. FunSearch then progresses iteratively, selecting programs, refining them, evaluating the outcomes, and reintroducing the most effective solutions to the program pool. This cycle leads to a self-enhancing system functioning towards continually improved outcomes.

DeepMind envisions that, with measures to avert the hallucinatory tendencies of LLMs, such models could not only pioneer new mathematical discoveries but also unlock innovative solutions to pressing real-world problems.

DeepMind's Other AI Triumphs

In the realm of AI breakthroughs, DeepMind is no stranger. The company has already made headlines with AlphaFold, an AI that has predicted the structure of a vast number of proteins, and its successor which could potentially map nearly all molecules catalogued in the Protein Data Bank. Additionally, their GraphCast AI model boasts enhanced weather forecasting capabilities, predicting conditions up to 10 days ahead with greater precision than conventional methods. DeepMind AI has also aided researchers in synthesizing hundreds of new materials within lab settings, showcasing the versatile applications of AI across diverse scientific domains.

DeepMind, AI, Mathematics