Google DeepMind just introduced a new AI system called AlphaEvolve. AlphaEvolve is not just another chatbot or coding assistant. It is something more powerful. It can write code, test that code, and improve it over time. It does all this to solve complex problems. Think of it like an AI engineer that never gets tired and can try thousands of ideas until it finds the best one.

This is a big deal. AI is no longer just helping us do tasks faster. It is starting to take on creative and technical challenges that once needed expert human engineers.

How AlphaEvolve works

AlphaEvolve works like an automated research assistant. First, a human gives it a task. This could be anything from optimizing a schedule to designing a math algorithm or improving computer chip performance. Along with the task, the human provides a way to measure success. This is called an evaluation function. It is like a scorecard. The AI uses this to judge how well each solution works.

AlphaEvolve uses two versions of Gemini. One is Gemini Flash. It is fast and explores many ideas. The other is Gemini Pro. It is smarter and focuses on better solutions. These two models work together to create lots of different versions of code for a specific problem.

Once AlphaEvolve creates the code, it tests each one using a scoring system. This score checks how good the code is based on the rules the user provides. It might be speed, memory use or accuracy. The top-scoring programs move on to the next round. From there, AlphaEvolve creates new code based on the winners. Think of it like survival of the fittest, but for software.

Not just short scripts

One of the most impressive things about AlphaEvolve is that it can create and improve entire systems, not just single pieces of code. It can write and refine long, multi-part programs. That means it can redesign how different parts of a computer system work together. For example, it can write and test parts of a new AI training routine, or optimize hardware design in chip manufacturing.

The code AlphaEvolve creates is readable and structured. This makes it easier for engineers to review and use in real-world systems. That is a big step forward. Many AI tools today are hard to trust because their output can be confusing or unreliable. But AlphaEvolve’s work is clear and auditable.

Big wins at Google

DeepMind has already used AlphaEvolve inside Google with impressive results. For example, it helped improve the company’s massive scheduling system, called Borg. AlphaEvolve came up with a new scheduling trick that saves about 0.7 percent of computing power across Google’s global data centers. That may not sound like much, but for a company the size of Google, it is a huge win. More jobs get done with the same amount of hardware.

In another case, AlphaEvolve looked at some low-level chip designs. It rewrote part of a circuit in Verilog, a hardware design language. After engineers checked the work, they used the improved design in an upcoming TPU chip. That is the kind of chip Google uses to train its biggest AI models.

AlphaEvolve also helped speed up some of the math tools used in DeepMind’s own models. It found ways to restructure complex code, like matrix multiplication kernels. One change sped up training by 1 percent. Another improved an attention function used in transformer models by over 30 percent. These are serious improvements that save time and money.

Smarter than it looks

AlphaEvolve does not just recycle ideas. It creates new ones. In a set of 50 math challenges, AlphaEvolve found the best known solution 75 percent of the time. Even better, in 20 percent of cases, it beat the best known human solution.

In one standout example, it found a new way to multiply two 4×4 matrices using only 48 basic multiplication steps. That beat a famous result from 1969. In another, it improved the known lower bound in a complex geometry problem called the kissing number in 11 dimensions. These are problems that top researchers have worked on for decades.

That kind of progress shows AlphaEvolve is not just fast. It is truly innovative in how it solves problems.

Where AlphaEvolve could go next

Right now, AlphaEvolve is being tested mostly within DeepMind and Google. But the company plans to expand access. A friendly user interface is in the works. Selected researchers will get early access soon.

The idea is to let more people use AlphaEvolve to solve tough problems in other fields. For example, materials science, logistics, and drug discovery. Any task that can be written as code and tested with a clear metric is fair game.

This could change how companies and scientists approach problem solving. Instead of writing and testing ideas by hand, they could describe the problem and let AlphaEvolve do the heavy lifting.

Why AlphaEvolve matters

AlphaEvolve is more than just a coding tool. It could be a new kind of teammate. One that can try hundreds or thousands of options quickly, learn from its mistakes, and deliver tested, working code.

For businesses, this means faster innovation. For researchers, it could unlock new knowledge. And for engineers, it is a powerful new partner that can take on time-consuming coding and testing tasks.

The key takeaway is that AlphaEvolve works best when humans and AI collaborate. The human sets the goal and defines what “good” looks like. The AI explores and delivers smart, working ideas.

This is a clear example of how AI is evolving from a tool into a true collaborator.

Search across your past Websets

Articles: Exa AI content

Search Criteria

Enrichments

Find more results