There is a mathematical concept called “Kissing number‘Somewhat disappointing, it has not found anything with real kisses; It suggests how many shells can touch a single area of the same size without crossing it (or ‘kissing’). In one dimension, the kiss number is two. This is 6 in two dimensions (think new York Times ‘Spelling bee puzzle layout). As the number of dimensions increases, the answer becomes less pronounced: for most of the dimensions of more than 4, only the upper and lower boundaries on the kiss number are known. Now, an AI agent developed by Google Deepmind has contributed to a problem called alphabet, leading to increased lower limit on kissing number in 11 dimensions from 592 to 593.
This can look like an agile improvement on the problem, especially given that the upper bound is 868 on the kissing number in 11 dimensions, so the unknown limit is still quite large. But this represents a novel mathematical discovery by the AI agent, and challenges the idea that large language models are not able to contribute original scientific contribution.
And this is just an example of what Alphavolv has completed. “We implemented alphabet in a series of open problems in research mathematics, and we deliberately raised problems from different parts of mathematics: analysis, combinatories, geometry,” Matej childA research scientist of deepminds who worked on the project. He found that for 75 percent of the problems, the AI model repeated the already known optimal solution. In 20 percent of cases, it found a new optimal that crossed any known solution. “Every such case is a new discovery,” says Balog. (In other 5 percent of cases, the AI converted to a solution that was worse than the known optimal.)
The model also developed a new algorithm for matrix multiplication – the operation that outlines most of the machine learning. A previous version of the AI model of Deepmind, called Alphatenser, had already defeated the previous most famous algorithm, which was discovered in 1969, to multiply 4 by 4 Matris. Alphaavolway found a more common version of that better algorithm.
Deepmind’s alphevevolva improved many practical problems in Google. Google Deepmind
Apart from abstract mathematics, the team also applied its model to practical problems, which faces a company every day as Google. AI was also used to customize data center orchestation to obtain 1 percent improvement, to customize the design of the next Google tension processing unit, and to search for improvement in the kernels used in Gemini training due to a reduction of 1 percent in training time.
“It is very surprising that you can do a lot of different things with the same system,” says Alexander NovicovA senior research scientist from Deepmind who also worked on Alphavolway.
How alfaavolv works
Alphaevolve is capable of being so normal because it can be applied to almost any problem that can be expressed as a code, and which can be tested by another piece of code. The user supplies an initial stab on the problem – a program that solves the problem at hand, although sub -form – and a verification program that checks how a piece of code meets the required criteria.
Then, a large language model, in this case Gemini, comes with other candidate programs to solve the same problem, and each is tested by the verification. From there, alphavolway uses a genetic algorithm such as the ‘most qualified’ of the proposed solutions survives and develops until the next generation. This process is repeated until the improvement in the solution stops.
The alphaevolve uses a attire of Gemini large language model (LLM) in combination with an evaluation code, orchestrated by a genetic algorithm to customize a piece of all code. Google Deepmind
“The big language models came around, and we started asking ourselves, is this the matter what they are only going to add to training data, or can we really use them to search for new, new algorithms or new knowledge?” Balog says. This research, the child, claims that “if you use a large language model correctly, you can get something in a very accurate sense, something that is new and correct as an algorithm.”
Alphavolva is from a long descent of deepmind models, which is going back to Alphazero, which shocked the world by learning better without using any human knowledge compared to any human player than any human player – just using learning to play the game and master it. Another Mathematics-Society AI reinforcement, based on learning, Alphoprof, performed at the silver-medalist level on the 2024 International Mathematics Olympiad.
For alfevolva, however, the team broke from the tradition of learning reinforcement in favor of the genetic algorithm. “The system is very simple,” says childogue. “And it is actually very easy to install on a wide range of problems.”
(Not completely scary) future
The team behind alphevolway expects to develop their system in two ways.
First of all, they want to apply it in a wide range of problems including those in natural science. To pursue this goal, they are planning to open an initial access program for interested academics so that they can use alfevolv in their research. It can be difficult for natural science to adapt to the system, as the verification of the proposed solutions may be less straightforward. But, Balog says, “We know that in natural science, there are many simulators for a variety of problems, and then those people can also be used within alpheevolva. And in future, we are very interested in wider the scope in this direction.”
Second, they want to improve the system, perhaps by coupling it with another deepmind project: AI co-scientistThis AI also uses an LLM and a genetic algorithm, but it focuses on the hypothesis generation in natural language. “They develop these high-level ideas and hypotheses,” say childgies. “Including this component into systems like alpheveolva, I believe that we will be allowed to go to a higher level of abstraction.”
These possibilities are exciting, but for some they can also make menacing sound-for example, the adaptation of Gemini training of alphabet can be seen as the beginning of the recurrent self-reforming AI, which can be seen as something Worry A fugitive intelligence will lead to the explosion, called eccentricity. The Deepmind team says that this is not their goal. “We are excited to contribute to advancing AI that benefits humanity,” Novicov says.
From your site articles
Related articles around web
