May 15, 20253 min readNew Google AI Chatbot Tackles Complex Math and ScienceA Google DeepMind system improves chip designs and addresses unsolved math problems but has not been rolled out to researchers outside the companyBy Elizabeth Gibney & Nature magazine DeepMind says that AlphaEvolve has helped to improve the design of AI chips. MF3d/Getty ImagesGoogle DeepMind has used chatbot models to come up with solutions to major problems in mathematics and computer science.The system, called AlphaEvolve, combines the creativity of a large language model (LLM) with algorithms that can scrutinize the model’s suggestions to filter and improve solutions. It was described in a white paper released by the company on 14 May.“The paper is quite spectacular,” says Mario Krenn, who leads the Artificial Scientist Lab at the Max Planck Institute for the Science of Light in Erlangen, Germany. “I think AlphaEvolve is the first successful demonstration of new discoveries based on general-purpose LLMs.”On supporting science journalismIf you're enjoying this article, consider supporting our award-winning journalism by subscribing. By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today.As well as using the system to discover solutions to open maths problems, DeepMind has already applied the artificial intelligence (AI) technique to its own practical challenges, says Pushmeet Kohli, head of science at the firm in London.AlphaEvolve has helped to improve the design of the company’s next generation of tensor processing units — computing chips developed specially for AI — and has found a way to more efficiently exploit Google’s worldwide computing capacity, saving 0.7% of total resources. “It has had substantial impact,” says Kohli.General-purpose AIMost of the successful applications of AI in science so far — including the protein-designing tool AlphaFold — have involved a learning algorithm that was hand-crafted for its task, says Krenn. But AlphaEvolve is general-purpose, tapping the abilities of LLMs to generate code to solve problems in a wide range of domains.DeepMind describes AlphaEvolve as an ‘agent’, because it involves using interacting AI models. But it targets a different point in the scientific process from many other ‘agentic’ AI science systems, which have been used to review the literature and suggest hypotheses.AlphaEvolve is based on the firm’s Gemini family of LLMs. Each task starts with the user inputting a question, criteria for evaluation and a suggested solution, for which the LLM proposes hundreds or thousands of modifications. An ‘evaluator’ algorithm then assesses the modifications against the metrics for a good solution (for example, in the task of assigning Google’s computing jobs, researchers want to waste fewer resources).On the basis of which solutions are judged to be the best, the LLM suggests fresh ideas and over time the system evolves a population of stronger algorithms, says Matej Balog, an AI scientist at DeepMind who co-led the research. “We explore this diverse set of possibilities of how the problem can be solved,” he says.AlphaEvolve builds on the firm’s FunSearch system, which in 2023 was shown to use a similar evolutionary approach to outdo humans in unsolved problems in maths. Compared with FunSearch, AlphaEvolve can handle much larger pieces of code and tackle more complex algorithms across a wide range of scientific domains, says Balog.DeepMind says that AlphaEvolve has come up with a way to perform a calculation, known as matrix multiplication, that in some cases is faster than the fastest-known method, which was developed by German mathematician Volker Strassen in 1969. Such calculations involve multiplying numbers in grids and are used to train neural networks. Despite being general-purpose, AlphaEvolve outperformed AlphaTensor, an AI tool described by the firm in 2022 and designed specifically for matrix mechanics.The approach could be used to tackle optimization problems, says Krenn, or anywhere in science where there are concrete metrics, or simulations, to evaluate what makes a good solution. This could include designing new microscopes, telescope or even materials, he adds.Narrow applicationsIn mathematics, AlphaEvolve seems to allow significant speed-ups in tackling some problems, says Simon Frieder, a mathematician and AI researcher at the University of Oxford, UK. But it will probably be applied only to the “narrow slice” of tasks that can be presented as problems to be solved through code, he says.Other researchers are reserving judgement about the tool’s usefulness until has been trialled outside DeepMind. “Until the systems have been tested by a broader community, I would stay sceptical and take the reported results with a grain of salt,” says Huan Sun, an AI researcher at the Ohio State University in Columbus. Frieder says he will wait until an open-source version is recreated by researchers, rather than a rely on DeepMind’s proprietary system, which could be withdrawn or changed.Although AlphaEvolve requires less computing power to run than AlphaTensor, it is still too resource-intensive to be made freely available on DeepMind’s servers, says Kohli.But the company hopes that announcing the system will encourage researchers to suggest areas of science in which to apply AlphaEvolve. “We are definitely committed to make sure that the most people in the scientific community get access to it,” says Kohli.This article is reproduced with permission and was first published on May 14, 2025.

Google DeepMind’s AI systems have taken big scientific strides in recent years — from predicting the 3D structures of almost every known protein in the universe to forecasting weather more accurately than ever before.  The UK-based lab today unveiled its latest advancement: AlphaEvolve, an AI coding agent that makes large language models (LLMs) like Gemini better at solving complex computing and mathematical problems.  AlphaEvolve is powered by the same models that it’s trying to improve. Using Gemini, the agent proposes programs — written in code — that try to solve a given problem. It runs each code snippet through automated tests that evaluate how accurate, efficient, or novel it is. AlphaEvolve keeps the top-performing code snippets and uses them as the basis for the next round of generation. Over many cycles, this process “evolves” better and better solutions. In essence, it is a self-evolving AI.   DeepMind has already used AlphaEvolve to tackle data centre energy use, design better chips, and speed up AI training. Here are five of its top feats so far.  1. It discovered new solutions to some of the world’s toughest maths problems AlphaEvolve was put to the test on over 50 open problems in maths, from combinatorics to number theory. In 20% of cases, it improved on the best-known solutions to them. The 💜 of EU techThe latest rumblings from the EU tech scene, a story from our wise ol' founder Boris, and some questionable AI art. It's free, every week, in your inbox. Sign up now! One of those was the 300-year-old kissing number problem. In 11-dimensional space, AlphaEvolve discovered a new lower bound with a configuration of 593 spheres — progress that even expert mathematicians hadn’t reached.  2. It made Google’s data centres more efficient The AI agent devised a way to better manage power scheduling at Google’s data centres. That has allowed the tech giant to improve its data centre energy efficiency by 0.7% over the last year — a significant cost and energy saver given the size of its data centre operation.  3. It helped train Gemini faster AlphaEvolve improved the way matrix multiplications are split into subproblems, a core operation in training AI models like Gemini. That optimisation sped up the process by 23%, reducing Gemini’s total training time by 1%. In the world of generative AI, every percentage point can translate into cost and energy savings.  4. It co-designed part of Google’s next AI chip The agent is also using its code-writing skills to rewire things in the physical world. It rewrote a portion of an arithmetic circuit in Verilog — a language used for chip design — making it more efficient. That same logic is now being used to develop Google’s future TPU (Tensor Processing Unit), an advanced chip for machine learning.   5. It beat a legendary algorithm from 1969 For decades, Strassen’s algorithm was the gold standard for multiplying 4×4 complex matrices. AlphaEvolve found a more efficient solution — using fewer scalar multiplications. This could lead to more advanced LLMs, which rely heavily on matrix multiplication to function. According to DeepMind, these feats are just the tip of the iceberg for AlphaEvolve. The lab envisions the agent solving countless problems, from discovering new materials and drugs to streamlining business operations. AI’s evolution will be a hot topic at TNW Conference, which takes place on June 19-20 in Amsterdam. Tickets for the event are now on sale — use the code TNWXMEDIA2025 at the checkout to get 30% off. Story by Siôn Geschwindt Siôn is a freelance science and technology reporter, specialising in climate and energy. From nuclear fusion breakthroughs to electric vehic (show all) Siôn is a freelance science and technology reporter, specialising in climate and energy. From nuclear fusion breakthroughs to electric vehicles, he's happiest sourcing a scoop, investigating the impact of emerging technologies, and even putting them to the test. He has five years of journalism experience and holds a dual degree in media and environmental science from the University of Cape Town, South Africa. When he's not writing, you can probably find Siôn out hiking, surfing, playing the drums or catering to his moderate caffeine addiction. You can contact him at: sion.geschwindt [at] protonmail [dot] com Get the TNW newsletter Get the most important tech news in your inbox each week. Also tagged with

AI evolution Google DeepMind creates super-advanced AI that can invent new algorithms AlphaEvolve has already made Google's data centers more efficient and improved Tensor chips. Ryan Whitwam – May 14, 2025 5:01 pm | 17 Credit: Google DeepMind Credit: Google DeepMind Story text Size Small Standard Large Width * Standard Wide Links Standard Orange * Subscribers only   Learn more Google's DeepMind research division claims its newest AI agent marks a significant step toward using the technology to tackle big problems in math and science. The system, known as AlphaEvolve, is based on the company's Gemini large language models (LLMs), with the addition of an "evolutionary" approach that evaluates and improves algorithms across a range of use cases. AlphaEvolve is essentially an AI coding agent, but it goes deeper than a standard Gemini chatbot. When you talk to Gemini, there is always a risk of hallucination, where the AI makes up details due to the non-deterministic nature of the underlying technology. AlphaEvolve uses an interesting approach to increase its accuracy when handling complex algorithmic problems. According to DeepMind, this AI uses an automatic evaluation system. When a researcher interacts with AlphaEvolve, they input a problem along with possible solutions and avenues to explore. The model generates multiple possible solutions, using the efficient Gemini Flash and the more detail-oriented Gemini Pro, and then each solution is analyzed by the evaluator. An evolutionary framework allows AlphaEvolve to focus on the best solution and improve upon it. Many of the company's past AI systems, for example, the protein-folding AlphaFold, were trained extensively on a single domain of knowledge. AlphaEvolve, however, is more dynamic. DeepMind says AlphaEvolve is a general-purpose AI that can aid research in any programming or algorithmic problem. And Google has already started to deploy it across its sprawling business with positive results. The team turned AlphaEvolve loose on Google's Borg cluster management system for its data centers. The AI suggested a change to the scheduling heuristics, which has been implemented to save Google 0.7 percent on its computing resources globally. For a company the size of Google, that's a significant financial benefit. AlphaEvolve may also be able to make generative AI more efficient, which is necessary if anyone is ever going to make money on the technology. The internal workings of generative systems are based on matrix multiplication operations. The most efficient way to multiply 4×4 complex-valued matrices was devised by mathematician Volker Strassen in 1969, and that held for decades, but DeepMind says AlphaEvolve has discovered a new algorithm that's even more efficient. DeepMind has worked on this problem before with narrowly trained AI agents like AlphaTensor. Despite being a general AI, AlphaEvolve came up with a better solution than AlphaTensor. Google's next-generation Tensor processing hardware will also benefit from AlphaEvolve. DeepMind reports that the AI created a change to the chip's Verilog hardware description language that dropped unnecessary bits to increase efficiency. Google is still working to verify the change but expects this to be part of the upcoming processor. So far, only Google has been able to tinker with AlphaEvolve. While it uses fewer computing resources than AlphaTensor did, it's still too complex to provide publicly. That may change in the future, but the evaluation approach that makes it so capable could also be integrated with smaller AI tools for research. Ryan Whitwam Senior Technology Reporter Ryan Whitwam Senior Technology Reporter Ryan Whitwam is a senior technology reporter at Ars Technica, covering the ways Google, AI, and mobile technology continue to change the world. Over his 20-year career, he's written for Android Police, ExtremeTech, Wirecutter, NY Times, and more. He has reviewed more phones than most people will ever own. You can follow him on Bluesky, where you will see photos of his dozens of mechanical keyboards. 17 Comments

Google DeepMind's AlphaEvolve AI system breaks a 56-year-old mathematical record by discovering a more efficient matrix multiplication algorithm that had eluded human mathematicians since Strassen's 1969 breakthrough.Read More

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Google DeepMind today pulled the curtain back on AlphaEvolve, an artificial-intelligence agent that can invent brand-new computer algorithms — then put them straight to work inside the company’s vast computing empire. AlphaEvolve pairs Google’s Gemini large language models with an evolutionary approach that tests, refines, and improves algorithms automatically. The system has already been deployed across Google’s data centers, chip designs, and AI training systems — boosting efficiency and solving mathematical problems that have stumped researchers for decades. “AlphaEvolve is a Gemini-powered AI coding agent that is able to make new discoveries in computing and mathematics,” explained Matej Balog, a researcher at Google DeepMind, in an interview with VentureBeat. “It can discover algorithms of remarkable complexity — spanning hundreds of lines of code with sophisticated logical structures that go far beyond simple functions.” The system dramatically expands upon Google’s previous work with FunSearch by evolving entire codebases rather than single functions. It represents a major leap in AI’s ability to develop sophisticated algorithms for both scientific challenges and everyday computing problems. Inside Google’s 0.7% efficiency boost: How AI-crafted algorithms run the company’s data centers AlphaEvolve has been quietly at work inside Google for over a year. The results are already significant. One algorithm it discovered has been powering Borg, Google’s massive cluster management system. This scheduling heuristic recovers an average of 0.7% of Google’s worldwide computing resources continuously — a staggering efficiency gain at Google’s scale. The discovery directly targets “stranded resources” — machines that have run out of one resource type (like memory) while still having others (like CPU) available. AlphaEvolve’s solution is especially valuable because it produces simple, human-readable code that engineers can easily interpret, debug, and deploy. The AI agent hasn’t stopped at data centers. It rewrote part of Google’s hardware design, finding a way to eliminate unnecessary bits in a crucial arithmetic circuit for Tensor Processing Units (TPUs). TPU designers validated the change for correctness, and it’s now headed into an upcoming chip design. Perhaps most impressively, AlphaEvolve improved the very systems that power itself. It optimized a matrix multiplication kernel used to train Gemini models, achieving a 23% speedup for that operation and cutting overall training time by 1%. For AI systems that train on massive computational grids, this efficiency gain translates to substantial energy and resource savings. “We try to identify critical pieces that can be accelerated and have as much impact as possible,” said Alexander Novikov, another DeepMind researcher, in an interview with VentureBeat. “We were able to optimize the practical running time of [a vital kernel] by 23%, which translated into 1% end-to-end savings on the entire Gemini training card.” Breaking Strassen’s 56-year-old matrix multiplication record: AI solves what humans couldn’t AlphaEvolve solves mathematical problems that stumped human experts for decades while advancing existing systems. The system designed a novel gradient-based optimization procedure that discovered multiple new matrix multiplication algorithms. One discovery toppled a mathematical record that had stood for 56 years. “What we found, to our surprise, to be honest, is that AlphaEvolve, despite being a more general technology, obtained even better results than AlphaTensor,” said Balog, referring to DeepMind’s previous specialized matrix multiplication system. “For these four by four matrices, AlphaEvolve found an algorithm that surpasses Strassen’s algorithm from 1969 for the first time in that setting.” The breakthrough allows two 4×4 complex-valued matrices to be multiplied using 48 scalar multiplications instead of 49 — a discovery that had eluded mathematicians since Volker Strassen’s landmark work. According to the research paper, AlphaEvolve “improves the state of the art for 14 matrix multiplication algorithms.” The system’s mathematical reach extends far beyond matrix multiplication. When tested against over 50 open problems in mathematical analysis, geometry, combinatorics, and number theory, AlphaEvolve matched state-of-the-art solutions in about 75% of cases. In approximately 20% of cases, it improved upon the best known solutions. One victory came in the “kissing number problem” — a centuries-old geometric challenge to determine how many non-overlapping unit spheres can simultaneously touch a central sphere. In 11 dimensions, AlphaEvolve found a configuration with 593 spheres, breaking the previous record of 592. How it works: Gemini language models plus evolution create a digital algorithm factory What makes AlphaEvolve different from other AI coding systems is its evolutionary approach. The system deploys both Gemini Flash (for speed) and Gemini Pro (for depth) to propose changes to existing code. These changes get tested by automated evaluators that score each variation. The most successful algorithms then guide the next round of evolution. AlphaEvolve doesn’t just generate code from its training data. It actively explores the solution space, discovers novel approaches, and refines them through an automated evaluation process — creating solutions humans might never have conceived. “One critical idea in our approach is that we focus on problems with clear evaluators. For any proposed solution or piece of code, we can automatically verify its validity and measure its quality,” Novikov explained. “This allows us to establish fast and reliable feedback loops to improve the system.” This approach is particularly valuable because the system can work on any problem with a clear evaluation metric — whether it’s energy efficiency in a data center or the elegance of a mathematical proof. From cloud computing to drug discovery: Where Google’s algorithm-inventing AI goes next While currently deployed within Google’s infrastructure and mathematical research, AlphaEvolve’s potential reaches much further. Google DeepMind envisions applications in material sciences, drug discovery, and other fields requiring complex algorithmic solutions. “The best human-AI collaboration can help solve open scientific challenges and also apply them at Google scale,” said Novikov, highlighting the system’s collaborative potential. Google DeepMind is now developing a user interface with its People + AI Research team and plans to launch an Early Access Program for selected academic researchers. The company is also exploring broader availability. The system’s flexibility marks a significant advantage. Balog noted that “at least previously, when I worked in machine learning research, it wasn’t my experience that you could build a scientific tool and immediately see real-world impact at this scale. This is quite unusual.” As large language models advance, AlphaEvolve’s capabilities will grow alongside them. The system demonstrates an intriguing evolution in AI itself — starting within the digital confines of Google’s servers, optimizing the very hardware and software that gives it life, and now reaching outward to solve problems that have challenged human intellect for decades or centuries. Daily insights on business use cases with VB Daily If you want to impress your boss, VB Daily has you covered. We give you the inside scoop on what companies are doing with generative AI, from regulatory shifts to practical deployments, so you can share insights for maximum ROI. Read our Privacy Policy Thanks for subscribing. Check out more VB newsletters here. An error occured.

Google DeepMind has once again used large language models to discover new solutions to long-standing problems in math and computer science. This time the firm has shown that its approach can not only tackle unsolved theoretical puzzles, but improve a range of important real-world processes as well. Google DeepMind's new tool, called AlphaEvolve, uses the Gemini 2.0 family of large language models (LLMs) to produce code for a wide range of different tasks. LLMs are known to be hit and miss at coding. The twist here is that AlphaEvolve scores each of Gemini’s suggestions, throwing out the bad and tweaking the good, in an iterative process, until it has produced the best algorithm it can. In many cases, the results are more efficient or more accurate than the best existing (human-written) solutions. “You can see it as a sort of super coding agent,” says Pushmeet Kohli, a vice president at Google DeepMind who leads its AI for Science teams. “It doesn’t just propose a piece of code or an edit, it actually produces a result that maybe nobody was aware of.” In particular, AlphaEvolve came up with a way to improve the software Google uses to allocate jobs to its many millions of servers around the world. Google DeepMind claims the company has been using this new software across all of its data centers for more than a year, freeing up 0.7% of Google's total computing resources. That might not sound like much, but at Google’s scale it’s huge. Jakob Moosbauer, a mathematician at the University of Warwick in the UK, is impressed. He says the way AlphaEvolve searches for algorithms that produce specific solutions—rather than searching for the solutions themselves—makes it especially powerful. “It makes the approach applicable to such a wide range of problems,” he says. “AI is becoming a tool that will be essential in mathematics and computer science.” AlphaEvolve continues a line of work that Google DeepMind has been pursuing for years. Its vision is that AI can help to advance human knowledge across math and science. In 2022, it developed AlphaTensor, a model that found a faster way to solve matrix multiplications—a fundamental problem in computer science—beating a record that had stood for more than 50 years. In 2023, it revealed AlphaDev, which discovered faster ways to perform a number of basic calculations performed by computers trillions of times a day. AlphaTensor and AlphaDev both turn math problems into a kind of game, then search for a winning series of moves. FunSearch, which arrived in late 2023, swapped out game-playing AI and replaced it with LLMs that can generate code. Because LLMs can carry out a range of tasks, FunSearch can take on a wider variety of problems than its predecessors, which were trained to play just one type of game. The tool was used to crack a famous unsolved problem in pure mathematics. AlphaEvolve is the next generation of FunSearch. Instead of coming up with short snippets of code to solve a specific problem, as FunSearch did, it can produce programs that are hundreds of lines long. This makes it applicable to a much wider variety of problems.     In theory, AlphaEvolve could be applied to any problem that can be described in code and that has solutions that can be evaluated by a computer. “Algorithms run the world around us, so the impact of that is huge,” says Matej Balog, a researcher at Google DeepMind who leads the algorithm discovery team. Survival of the fittest Here’s how it works: AlphaEvolve can be prompted like any LLM. Give it a description of the problem and any extra hints you want, such as previous solutions, and AlphaEvolve will get Gemini 2.0 Flash (the smallest, fastest version of Google DeepMind’s flagship LLM) to generate multiple blocks of code to solve the problem. It then takes these candidate solutions, runs them to see how accurate or efficient they are, and scores them according to a range of relevant metrics. Does this code produce the correct result? Does it run faster than previous solutions? And so on. AlphaEvolve then takes the best of the current batch of solutions and asks Gemini to improve them. Sometimes AlphaEvolve will throw a previous solution back into the mix to prevent Gemini from hitting a dead end. When it gets stuck, AlphaEvolve can also call on Gemini 2.0 Pro, the most powerful of Google DeepMind’s LLMs. The idea is to generate many solutions with the faster Flash but add solutions from the slower Pro when needed. These rounds of generation, scoring, and regeneration continue until Gemini fails to come up with anything better than what it already has. Number games The team tested AlphaEvolve on a range of different problems. For example, they looked at matrix multiplication again to see how a general-purpose tool like AlphaEvolve compared to the specialized AlphaTensor. Matrices are grids of numbers. Matrix multiplication is a basic computation that underpins many applications, from AI to computer graphics, yet nobody knows the fastest way to do it. “It’s kind of unbelievable that it’s still an open question,” says Balog. The team gave AlphaEvolve a description of the problem and an example of a standard algorithm for solving it. The tool not only produced new algorithms that could calculate 14 different sizes of matrix faster than any existing approach, it also improved on AlphaTensor’s record-beating result for multipying two four-by-four matrices. AlphaEvolve scored 16,000 candidates suggested by Gemini to find the winning solution, but that’s still more efficient than AlphaTensor, says Balog. AlphaTensor’s solution also only worked when a matrix was filled with 0s and 1s. AlphaEvolve solves the problem with other numbers too. “The result on matrix multiplication is very impressive,” says Moosbauer. “This new algorithm has the potential to speed up computations in practice.” Manuel Kauers, a mathematician at Johannes Kepler University in Linz, Austria, agrees: “The improvement for matrices is likely to have practical relevance.” By coincidence, Kauers and a colleague have just used a different computational technique to find some of the speedups AlphaEvolve came up with. The pair posted a paper online reporting their results last week. “It is great to see that we are moving forward with the understanding of matrix multiplication,” says Kauers. “Every technique that helps is a welcome contribution to this effort.” Real-world problems Matrix multiplication was just one breakthrough. In total, Google DeepMind tested AlphaEvolve on more than 50 different types of well-known math puzzles, including problems in Fourier analysis (the math behind data compression, essential to applications such as video streaming), the minimum overlap problem (an open problem in number theory proposed by mathematician Paul Erdős in 1955), and kissing numbers (a problem introduced by Isaac Newton that has applications in materials science, chemistry, and cryptography). AlphaEvolve matched the best existing solutions in 75% of cases and found better solutions in 20% of cases.   Google DeepMind then applied AlphaEvolve to a handful of real-world problems. As well as coming up with a more efficient algorithm for managing computational resources across data centers, the tool found a way to reduce the power consumption of Google’s specialized tensor processing unit chips. AlphaEvolve even found a way to speed up the training of Gemini itself, by producing a more efficient algorithm for managing a certain type of computation used in the training process. Google DeepMind plans to continue exploring potential applications of its tool. One limitation is that AlphaEvolve can’t be used for problems with solutions that need to be scored by a person, such as lab experiments that are subject to interpretation.    Moosbauer also points out that while AlphaEvolve may produce impressive new results across a wide range of problems, it gives little theoretical insight into how it arrived at those solutions. That’s a drawback when it comes to advancing human understanding.   Even so, tools like AlphaEvolve are set to change the way researchers work. “I don't think we are finished,” says Kohli. “There is much further that we can go in terms of how powerful this type of approach is.”

Google DeepMind has once again used large language models to discover new solutions to long-standing problems in math and computer science. This time the firm has shown that its approach can not only tackle unsolved theoretical puzzles, but improve a range of important real-world processes as well. Google DeepMind's new tool, called AlphaEvolve, uses the Gemini 2.0 family of large language models (LLMs) to produce code for a wide range of different tasks. LLMs are known to be hit and miss at coding. The twist here is that AlphaEvolve scores each of Gemini’s suggestions, throwing out the bad and tweaking the good, in an iterative process, until it has produced the best algorithm it can. In many cases, the results are more efficient or more accurate than the best existing (human-written) solutions. “You can see it as a sort of super coding agent,” says Pushmeet Kohli, a vice president at Google DeepMind who leads its AI for Science teams. “It doesn’t just propose a piece of code or an edit, it actually produces a result that maybe nobody was aware of.” In particular, AlphaEvolve came up with a way to improve the software Google uses to allocate jobs to its many millions of servers around the world. Google DeepMind claims the company has been using this new software across all of its data centers for more than a year, freeing up 0.7% of Google's total computing resources. That might not sound like much, but at Google’s scale it’s huge. Jakob Moosbauer, a mathematician at the University of Warwick in the UK, is impressed. He says the way AlphaEvolve searches for algorithms that produce specific solutions—rather than searching for the solutions themselves—makes it especially powerful. “It makes the approach applicable to such a wide range of problems,” he says. “AI is becoming a tool that will be essential in mathematics and computer science.” AlphaEvolve continues a line of work that Google DeepMind has been pursuing for years. Its vision is that AI can help to advance human knowledge across math and science. In 2022, it developed AlphaTensor, a model that found a faster way to solve matrix multiplications—a fundamental problem in computer science—beating a record that had stood for more than 50 years. In 2023, it revealed AlphaDev, which discovered faster ways to perform a number of basic calculations performed by computers trillions of times a day. AlphaTensor and AlphaDev both turn math problems into a kind of game, then search for a winning series of moves. FunSearch, which arrived in late 2023, swapped out game-playing AI and replaced it with LLMs that can generate code. Because LLMs can carry out a range of tasks, FunSearch can take on a wider variety of problems than its predecessors, which were trained to play just one type of game. The tool was used to crack a famous unsolved problem in pure mathematics. AlphaEvolve is the next generation of FunSearch. Instead of coming up with short snippets of code to solve a specific problem, as FunSearch did, it can produce programs that are hundreds of lines long. This makes it applicable to a much wider variety of problems.     In theory, AlphaEvolve could be applied to any problem that can be described in code and that has solutions that can be evaluated by a computer. “Algorithms run the world around us, so the impact of that is huge,” says Matej Balog, a researcher at Google DeepMind who leads the algorithm discovery team. Survival of the fittest Here’s how it works: AlphaEvolve can be prompted like any LLM. Give it a description of the problem and any extra hints you want, such as previous solutions, and AlphaEvolve will get Gemini 2.0 Flash (the smallest, fastest version of Google DeepMind’s flagship LLM) to generate multiple blocks of code to solve the problem. It then takes these candidate solutions, runs them to see how accurate or efficient they are, and scores them according to a range of relevant metrics. Does this code produce the correct result? Does it run faster than previous solutions? And so on. AlphaEvolve then takes the best of the current batch of solutions and asks Gemini to improve them. Sometimes AlphaEvolve will throw a previous solution back into the mix to prevent Gemini from hitting a dead end. When it gets stuck, AlphaEvolve can also call on Gemini 2.0 Pro, the most powerful of Google DeepMind’s LLMs. The idea is to generate many solutions with the faster Flash but add solutions from the slower Pro when needed. These rounds of generation, scoring, and regeneration continue until Gemini fails to come up with anything better than what it already has. Number games The team tested AlphaEvolve on a range of different problems. For example, they looked at matrix multiplication again to see how a general-purpose tool like AlphaEvolve compared to the specialized AlphaTensor. Matrices are grids of numbers. Matrix multiplication is a basic computation that underpins many applications, from AI to computer graphics, yet nobody knows the fastest way to do it. “It’s kind of unbelievable that it’s still an open question,” says Balog. The team gave AlphaEvolve a description of the problem and an example of a standard algorithm for solving it. The tool not only produced new algorithms that could calculate 14 different sizes of matrix faster than any existing approach, it also improved on AlphaTensor’s record-beating result for multipying two four-by-four matrices. AlphaEvolve scored 16,000 candidates suggested by Gemini to find the winning solution, but that’s still more efficient than AlphaTensor, says Balog. AlphaTensor’s solution also only worked when a matrix was filled with 0s and 1s. AlphaEvolve solves the problem with other numbers too. “The result on matrix multiplication is very impressive,” says Moosbauer. “This new algorithm has the potential to speed up computations in practice.” Manuel Kauers, a mathematician at Johannes Kepler University in Linz, Austria, agrees: “The improvement for matrices is likely to have practical relevance.” By coincidence, Kauers and a colleague have just used a different computational technique to find some of the speedups AlphaEvolve came up with. The pair posted a paper online reporting their results last week. “It is great to see that we are moving forward with the understanding of matrix multiplication,” says Kauers. “Every technique that helps is a welcome contribution to this effort.” Real-world problems Matrix multiplication was just one breakthrough. In total, Google DeepMind tested AlphaEvolve on more than 50 different types of well-known math puzzles, including problems in Fourier analysis (the math behind data compression, essential to applications such as video streaming), the minimum overlap problem (an open problem in number theory proposed by mathematician Paul Erdős in 1955), and kissing numbers (a problem introduced by Isaac Newton that has applications in materials science, chemistry, and cryptography). AlphaEvolve matched the best existing solutions in 75% of cases and found better solutions in 20% of cases.   Google DeepMind then applied AlphaEvolve to a handful of real-world problems. As well as coming up with a more efficient algorithm for managing computational resources across data centers, the tool found a way to reduce the power consumption of Google’s specialized tensor processing unit chips. AlphaEvolve even found a way to speed up the training of Gemini itself, by producing a more efficient algorithm for managing a certain type of computation used in the training process. Google DeepMind plans to continue exploring potential applications of its tool. One limitation is that AlphaEvolve can’t be used for problems with solutions that need to be scored by a person, such as lab experiments that are subject to interpretation.    Moosbauer also points out that while AlphaEvolve may produce impressive new results across a wide range of problems, it gives little theoretical insight into how it arrived at those solutions. That’s a drawback when it comes to advancing human understanding.   Even so, tools like AlphaEvolve are set to change the way researchers work. “I don't think we are finished,” says Kohli. “There is much further that we can go in terms of how powerful this type of approach is.”

Google DeepMind's AlphaEvolve AI system breaks a 56-year-old mathematical record by discovering a more efficient matrix multiplication algorithm that had eluded human mathematicians since Strassen's 1969 breakthrough.Read More

Google DeepMind has once again used large language models to discover new solutions to long-standing problems in math and computer science. This time the firm has shown that its approach can not only tackle unsolved theoretical puzzles, but improve a range of important real-world processes as well. Google DeepMind’s new tool, called AlphaEvolve, uses the Gemini 2.0 family of large language models (LLMs) to produce code for a wide range of different tasks. LLMs are known to be hit and miss at coding. The twist here is that AlphaEvolve scores each of Gemini’s suggestions, throwing out the bad and tweaking the good, in an iterative process, until it has produced the best algorithm it can. In many cases, the results are more efficient or more accurate than the best existing (human-written) solutions. “You can see it as a sort of super coding agent,” says Pushmeet Kohli, a vice president at Google DeepMind who leads its AI for Science teams. “It doesn’t just propose a piece of code or an edit, it actually produces a result that maybe nobody was aware of.” In particular, AlphaEvolve came up with a way to improve the software Google uses to allocate jobs to its many millions of servers around the world. Google DeepMind claims the company has been using this new software across all of its data centers for more than a year, freeing up 0.7% of Google’s total computing resources. That might not sound like much, but at Google’s scale it’s huge. Jakob Moosbauer, a mathematician at the University of Warwick in the UK, is impressed. He says the way AlphaEvolve searches for algorithms that produce specific solutions—rather than searching for the solutions themselves—makes it especially powerful. “It makes the approach applicable to such a wide range of problems,” he says. “AI is becoming a tool that will be essential in mathematics and computer science.” AlphaEvolve continues a line of work that Google DeepMind has been pursuing for years. Its vision is that AI can help to advance human knowledge across math and science. In 2022, it developed AlphaTensor, a model that found a faster way to solve matrix multiplications—a fundamental problem in computer science—beating a record that had stood for more than 50 years. In 2023, it revealed AlphaDev, which discovered faster ways to perform a number of basic calculations performed by computers trillions of times a day. AlphaTensor and AlphaDev both turn math problems into a kind of game, then search for a winning series of moves. FunSearch, which arrived in late 2023, swapped out game-playing AI and replaced it with LLMs that can generate code. Because LLMs can carry out a range of tasks, FunSearch can take on a wider variety of problems than its predecessors, which were trained to play just one type of game. The tool was used to crack a famous unsolved problem in pure mathematics. AlphaEvolve is the next generation of FunSearch. Instead of coming up with short snippets of code to solve a specific problem, as FunSearch did, it can produce programs that are hundreds of lines long. This makes it applicable to a much wider variety of problems.     In theory, AlphaEvolve could be applied to any problem that can be described in code and that has solutions that can be evaluated by a computer. “Algorithms run the world around us, so the impact of that is huge,” says Matej Balog, a researcher at Google DeepMind who leads the algorithm discovery team. Survival of the fittest Here’s how it works: AlphaEvolve can be prompted like any LLM. Give it a description of the problem and any extra hints you want, such as previous solutions, and AlphaEvolve will get Gemini 2.0 Flash (the smallest, fastest version of Google DeepMind’s flagship LLM) to generate multiple blocks of code to solve the problem. It then takes these candidate solutions, runs them to see how accurate or efficient they are, and scores them according to a range of relevant metrics. Does this code produce the correct result? Does it run faster than previous solutions? And so on. AlphaEvolve then takes the best of the current batch of solutions and asks Gemini to improve them. Sometimes AlphaEvolve will throw a previous solution back into the mix to prevent Gemini from hitting a dead end. When it gets stuck, AlphaEvolve can also call on Gemini 2.0 Pro, the most powerful of Google DeepMind’s LLMs. The idea is to generate many solutions with the faster Flash but add solutions from the slower Pro when needed. These rounds of generation, scoring, and regeneration continue until Gemini fails to come up with anything better than what it already has. Number games The team tested AlphaEvolve on a range of different problems. For example, they looked at matrix multiplication again to see how a general-purpose tool like AlphaEvolve compared to the specialized AlphaTensor. Matrices are grids of numbers. Matrix multiplication is a basic computation that underpins many applications, from AI to computer graphics, yet nobody knows the fastest way to do it. “It’s kind of unbelievable that it’s still an open question,” says Balog. The team gave AlphaEvolve a description of the problem and an example of a standard algorithm for solving it. The tool not only produced new algorithms that could calculate 14 different sizes of matrix faster than any existing approach, it also improved on AlphaTensor’s record-beating result for multipying two four-by-four matrices. AlphaEvolve scored 16,000 candidates suggested by Gemini to find the winning solution, but that’s still more efficient than AlphaTensor, says Balog. AlphaTensor’s solution also only worked when a matrix was filled with 0s and 1s. AlphaEvolve solves the problem with other numbers too. “The result on matrix multiplication is very impressive,” says Moosbauer. “This new algorithm has the potential to speed up computations in practice.” Manuel Kauers, a mathematician at Johannes Kepler University in Linz, Austria, agrees: “The improvement for matrices is likely to have practical relevance.” By coincidence, Kauers and a colleague have just used a different computational technique to find some of the speedups AlphaEvolve came up with. The pair posted a paper online reporting their results last week. “It is great to see that we are moving forward with the understanding of matrix multiplication,” says Kauers. “Every technique that helps is a welcome contribution to this effort.” Real-world problems Matrix multiplication was just one breakthrough. In total, Google DeepMind tested AlphaEvolve on more than 50 different types of well-known math puzzles, including problems in Fourier analysis (the math behind data compression, essential to applications such as video streaming), the minimum overlap problem (an open problem in number theory proposed by mathematician Paul Erdős in 1955), and kissing numbers (a problem introduced by Isaac Newton that has applications in materials science, chemistry, and cryptography). AlphaEvolve matched the best existing solutions in 75% of cases and found better solutions in 20% of cases.   Google DeepMind then applied AlphaEvolve to a handful of real-world problems. As well as coming up with a more efficient algorithm for managing computational resources across data centers, the tool found a way to reduce the power consumption of Google’s specialized tensor processing unit chips. AlphaEvolve even found a way to speed up the training of Gemini itself, by producing a more efficient algorithm for managing a certain type of computation used in the training process. Google DeepMind plans to continue exploring potential applications of its tool. One limitation is that AlphaEvolve can’t be used for problems with solutions that need to be scored by a person, such as lab experiments that are subject to interpretation.    Moosbauer also points out that while AlphaEvolve may produce impressive new results across a wide range of problems, it gives little theoretical insight into how it arrived at those solutions. That’s a drawback when it comes to advancing human understanding.   Even so, tools like AlphaEvolve are set to change the way researchers work. “I don’t think we are finished,” says Kohli. “There is much further that we can go in terms of how powerful this type of approach is.”