What’s next for AI and math

This year, a number of LRMS, which tries to solve a step -by -step problem instead of making the first result to them, achieved high degrees in the American Da`wah mathematics exam (AIME), a test that conducted the best 5 % of the American high school students.

Meanwhile, a handful of new hybrid models that combine LLMS with a type of fact -verification system, also a penetration. Emily de Oliveira Santos, a mathematician at the University of São Paulo, Brazil, refers to Google Deppings Alphabab, a system that combines LLM and a DeepMind game model to play Alphazero, as one major pork. Last year, AlPhaproff became the first computer program that matches the performance of the silver medal in the International Mathematics Olympics, and it is one of the most prestigious mathematics competitions in the world.

In May, the Google DeepMind model called alphavolve discovered better results than anything that people had reached so far for more than 50 mathematics puzzles that have not been solved and many computer science problems in the real world.

The rise in progress is clear. “The GPT-4 cannot do mathematics beyond the first university level,” says De Oliveira Santos. “I remember testing it at the time of its release with a problem in the topology, and he was unable to write more than a few lines without being completely lost.” But when it gave the same problem to Openai’s O1, LRM, which was released in January, she listened to it.

Does this mean that such models have been appointed to become a kind of DARPA common author? “The problems of the mathematics Olympics often do not include the ability to implement smart tricks, while search problems are more exploring and often have a lot of moving pieces,” she says. Success may not be transmitted in a kind of problem solving to another.

Others agree. Martin Bridson, a mathematician at Oxford University, believes that the Olympic Olympic result is a great achievement. He says: “On the other hand, I do not find it outperforming the mind.” “It is not a change in the model in the sense that” Wow, I thought the machines would never be able to do this. “I expected that the machines would be able to do this.”

This is because although the problems in Math Olympiad – and secondary school tests or similar university stage such as AIME – is difficult, there is a pattern of many of them. “We have training camps to train high school children to do,” says Bridson. “And if you can train a large number of people to do these problems, why shouldn’t you be able to train a machine to do?”

Sergey Gokov, a mathematician at the California Institute of Technology, who coaches the Sports Olympic Games, notes that the question style does not change much between the competitions. New problems are set every year, but can be solved with the same old tricks.