There is a sports concept called “Kissing number“Somewhat disappointing, has nothing to do with the actual kissing; There are how many number Fields It can touch (or “kiss” one field of equal size without crossing it. In one dimension, kissing number is two. In two dimensions, 6 (Think of me New York TimesSpecific bee puzzle Settings). With the growth of the number of dimensions, the answer becomes less clear: for most dimensions that exceed 4, the upper and lower boundaries are known only on the kissing number. Now, artificial intelligence agent was developed by Google Deepmind Alphavolve has made its contribution to the problem, which increased the minimum kissing number in 11 dimensions from 592 to 593.
This may seem to serve as a gradual improvement in the problem, especially since the upper limit for kissing in 11 dimensions is 868, so the unknown range is still very large. But it represents a new sporting discovery by Amnesty International’s agent, and challenges the idea Language models We are unable Original scientific contributions.
This is just one example of what alphavolve has accomplished. We apply Alphavolve through a set of open problems in the search mathematicsAnd intentionally chose problems from different parts of mathematics: analysis, integration, engineering, “he says Matti cDEPMIND research scientist who worked on the project. They found that for 75 percent of the problems, the artificial intelligence model repeated the optimal solution already known. In 20 percent of cases, I found that the new optimum exceeded any known solution. “Every such case is a new discovery,” says Balge. (In 5 percent of cases, artificial intelligence is close to a solution that was worse than optimal cases known.)
The model also developed a new algorithm to strike the matrix – the process behind many Automated learning. There was a previous version of DeepMind’s Ai, called alphatensor, already Defeat The best previous known algorithm, discovered in 1969, to hit 4 of 4 matrices. Alphaevolve found a more general version of that improved algorithm.
DeepMind has made improvements to many practical problems in Google. Google DeepMind
In addition to abstract mathematics, the team also applied its model to the practical problems facing Google every day. Artificial intelligence has also been used to improve the data center coincidence for 1 percent improvement, to improve the next Google design TensionerAnd the discovery of an improvement in the Kernel used in Gemini training, which leads to a 1 % decrease at the time of training.
“It is very surprising that you can do many different things with one system,” he says. Alexander NovkovOne of the major research world in DeepMind who also worked on Alphavolve.
How to work alphavolve
The alphavolve can be very general because it can be applied to almost any problem that can be expressed as a symbol, which can be examined by another symbol. The user provides a preliminary stab in the problem – a program that solves the offered problem, whatever optimal level – and the verification program that checks the extent of part of the code with the required standards.
After that, the Great Language Model comes in this case Gemini, with other filters programs to solve the same problem, and each one is tested by verification. From there, Alphaevolve uses a file Genetic algorithm Like the “most suitable” of proposed solutions remains and develops to the next generation. This process is repeated until solutions stop improvement.
Alphaevolve uses a group of large Gemini language models (LLMS) in conjunction with the evaluation code, all of which are its format Genetic algorithm To improve a piece of code. Google DeepMind
“The big language models came, and we started asking ourselves, is it the situation that will only add what is in the training data, or we can actually use it to discover something completely new and new Algorithms Or new knowledge? Balog says.
Alphavolve comes from long ratios of DeepMind models, and return to Alfaziroany The world surprised By learning to play ChessGo, and other games are better than any human player without using any human knowledge – only by playing the game and using Learning reinforcement To master it. Another Amnesty International, which resolves mathematics based on Learning reinforcementAlphab, procedure On the medical silver level on the International Mathematics Olympics 2024.
For alphavolve, the team exploded from the tradition of reinforcement learning in favor of the genetic algorithm. “The system is much simpler.” “This actually has consequences, it is much easier to prepare a wide range of problems.”
The future (not completely scary)
The team hopes behind Alphavolve to develop its system in two ways.
First, they want to apply it to a wide range of problems, including those in natural sciences. To follow this goal, they are planning to open an early access program for academics interested in using alphavolve in their research. It may be difficult to adapt the system to the natural sciences, because checking the proposed solutions may be less clear. However, Balge says, “We know that in the natural sciences, there are a lot of simulation devices for different types of problems, after which they can be used in Alphavolve as well. We, in the future, are very interested in expanding a range in this direction.”
Second, they want to improve the same system, perhaps by associating it with another DeepMind project: Amnesty International Participants. This artificial intelligence also uses LLM and a genetic algorithm, but it focuses on generating hypotheses in the natural language. “They are developing these ideas and assumptions with a higher level,” says Balge. “Integating this component into alphavolve systems, I think, will allow us to go to higher levels of abstraction.”
These possibilities are exciting, but for some people it may also seem threat-for example, Alphavolve training may be seen on Gemini training as a start from artificial intelligence Worry It would lead to a fugitive intelligence explosion referred to as Uniqueness. Deepmind asserts that this is not their goal, of course. “We are excited to contribute to the progress of artificial intelligence that benefits humanity,” Novikov says.
From your site articles
Related articles about the web
