Terence Tao is stunned: Google unveils the "AI Edison", breaking three 18 - year - old math records. Is scientific research no longer reliant on inspiration?

In mid-May, Google released AlphaEvolve. In just 30 days, it solved a problem that had remained unsolved for 18 years. It may even have initiated a scientific revolution that doesn't rely on "inspiration": In the future, scientists will no longer depend on intuition but rely on AI to solve difficult problems!

In mid-May, Google dropped a bombshell in the fields of science and computing: AlphaEvolve.

Using the Gemini model, it discovers brand - new algorithms.

For example, in just 30 days, AlphaEvolve, in collaboration with humans, solved a mathematical problem that had been untouched for 18 years three times! The well - known Chinese mathematician Terence Tao was slightly surprised by this.

Not only has it made significant progress in computer science and mathematics, but AI may also impact a wider range of scientific fields.

It is not just a text - generating tool, nor is it a simple template generator. It symbolizes the infinite possibilities of AI, just like the "divine move" of AlphaGo, which demonstrated breakthroughs that humans had never achieved.

This could even be a step towards AI self - improvement.

In an in - depth conversation, Chinese investor Sarah Guo interviewed Pushmeet Kohli (left in the picture below), the vice - president of science and strategy at Google DeepMind, and research scientist Matej Balog (right in the picture below).

They shared the story behind AlphaEvolve. Beyond mathematics and computer science, they also speculated further: Can the concept behind AlphaEvolve disrupt more basic scientific fields?

AlphaEvolve proves that replacing "luck" with intelligence can also disrupt science. AlphaEvolve may be initiating a scientific revolution that doesn't rely on "inspiration".

AlphaEvolve: Progress That Surprised Terence Tao

DeepMind's mission is to build artificial intelligence responsibly for the benefit of humanity. Over the years, DeepMind has been searching for new algorithms in the scientific field.

What makes AlphaEvolve different?

Pushmeet Kohli believes that the difference can be seen from a historical perspective.

It all starts with AlphaGo.

AlphaGo was not only able to efficiently explore all possible situations in Go but also propose the best move at that time. In the decades - long history of Go, humans had never discovered such a playing style.

In a sense, AlphaGo is an AI agent. In a vast search space, it can efficiently explore and propose the optimal solution. This ability surprised people because Go is very complex, and scientists thought it would take a long time for AI to make a breakthrough in this field.

From the work of AlphaGo, DeepMind got inspired:

If AI can search all possible situations in Go so efficiently, can a similar idea be used to search the algorithm space?

This is the basis for starting the development of AlphaTensor.

For decades, people thought that the complexity of matrix multiplication was cubic. That is, if you have two matrices with a dimension of n, the time complexity of the calculation is n³.

More than 50 years ago, German mathematician Strassen proposed a very counter - intuitive method, proving that in fact, the complexity of matrix multiplication is lower than previously expected.

Through searching, AlphaTensor discovered a more efficient solution than previously known algorithms. It not only outperformed traditional algorithms in terms of efficiency but also proved that AI can achieve super - human - level breakthroughs.

However, the problem is that AlphaTensor is specifically designed for matrix multiplication. So, can this method be extended to more general problems? This led to further exploration of AlphaEvolve.

AlphaEvolve can not only handle specific tasks but is more general and can handle a wider range of problems.

AlphaEvolve uses an evolutionary algorithm similar to AlphaTensor. But it is no longer limited to the specific problem of matrix multiplication and can search in a wider programming space to propose solutions.

Continuous Evolution and Self - Improvement

It sounds like AlphaEvolve is similar to evolutionary selection, right? How does it improve in each generation?

In each generation, AlphaEvolve continuously improves, and each generation is optimized based on the strong solutions of the previous generation.

Through the gene pool and the evaluation function, it ensures that the improvement in each generation can enhance the overall quality of the solutions while maintaining diversity to discover the best solution in the vast search space.

What is the scale of this evolutionary process? How to control the number of model iterations?

Regarding this issue, AlphaEvolve has a great feature, which is that it can adapt to the difficulty of the problem.

If AlphaEvolve is asked to solve a relatively simple problem, it can get the answer almost immediately. But if it is a very complex problem, the solution may take longer and more generations to continuously improve.

Fortunately, AlphaEvolve can continuously improve. Even when facing extremely difficult problems, it can still keep getting better.

This is very valuable because when continuously optimizing, many traditional systems often encounter bottlenecks in the early stage and cannot continue to improve.

As for predicting how many generations are needed to reach the optimal solution, this problem is quite complex. The difficulty of the problem is unpredictable. Especially in the scientific field, some seemingly simple problems may actually be very difficult, and vice versa. But fortunately, as long as AlphaEvolve keeps running, it will get better results over time.

Significance for Coding Agents

How is AlphaEvolve different from general coding agents?

Compared with general coding agents, AlphaEvolve's advantage lies in its ability to handle more complex tasks with higher efficiency and creativity.

When facing complex or ambiguous tasks, most general - purpose coding agents are prone to getting stuck or making mistakes because they usually rely on direct task descriptions, which are often not precise enough, or they don't have strong judgment abilities.

AlphaEvolve, on the other hand, relies on a strict evaluation function. It can distinguish between valid and invalid solutions.

Its "creativity" is not only reflected in proposing new algorithms but also in its ability to effectively evaluate and optimize solutions.

Whenever a new solution is proposed, the evaluation function helps determine whether it is valid.

For example, when optimizing data center scheduling, the evaluation function may be a simulator that can judge how a given scheduling algorithm performs in reality.

This evaluation process helps AlphaEvolve search the solution space more accurately.

For developers, designing a good evaluation function is indeed very challenging. You need to clearly define what kind of result is a good solution. In some cases, developers can use existing simulators for evaluation, while in other more complex cases, they may need to develop customized evaluation tools.

The evaluation function not only needs to be able to judge the quality of the solution but also needs to be flexibly applied in different tasks. For example, in the data center scheduling optimization problem, the complexity of the evaluation function may be much higher than that of some simpler tasks.

This is why the importance of the evaluation function in the AI system is emphasized. Only with an accurate evaluation function can AI innovate effectively.

Left: The heuristic function customized by AlphaEvolve for Google's workload and capacity; Right: Visualization of the heuristic scoring function

Scientists Changing Roles

Both Matej Balog and Pushmeet Kohli believe that in the future, the role of scientists will change to some extent.

It is imaginable that in the future, scientists will focus more on how to define problems, design evaluation functions, and how to interpret the results generated by AI.

AI will become a powerful tool for scientists, helping them solve complex problems more quickly.

AI not only gives answers but also provides algorithms. Scientists can study the algorithms to understand the underlying principles, which is very important for in - depth understanding of problems and solutions.

This is exactly why AlphaEvolve dominates multiple fields.

Mathematicians and scientists can not only see the final solution but also understand the path to reach this solution. This new perspective is crucial for promoting scientific development.

In addition, AlphaEvolve not only promotes technological innovation but also helps scientists discover new ways of thinking and challenge existing cognitive frameworks.

The emergence of AlphaEvolve marks a new era in scientific research. It has not only created miracles in the field of algorithms but also paved the way for future scientific revolutions.

Driven by AlphaEvolve, perhaps we are about to witness: Science no longer relies on "inspiration" but on "intelligence".

References:

https://www.youtube.com/watch?v=2Fs6VZpsiMQ

This article is from the WeChat public account "New Intelligence Yuan", author: KingHZ. Republished by 36Kr with permission.