TEMPO.CO, Jakarta - Google's artificial intelligence (AI) division, DeepMind, has released AlphaEvolve, an AI claimed to address problems with solutions that can be evaluated by machines. "In experiments, the system called AlphaEvolve has been able to help optimize some of the infrastructure used by Google to train its AI models," DeepMind stated as quoted from Tech Crunch, Wednesday, May 14, 2025.
DeepMind has mentioned that they are currently developing specific interfaces to make it easy for people to interact with AlphaEvolve. Not only that, Google through DeepMind plans to open an early access program for selected academics before launching the AI system widely.
According to a report by Ars Technica, AlphaEvolve is an artificial intelligence (AI) system developed based on Google's large language model (LLM) Gemini. With the addition of an evolutionary approach, the latest AI from DeepMind is capable of evaluating and optimizing algorithms for various use cases.
In essence, AlphaEvolve is a specialized AI agent for coding that is much more advanced than the usual Gemini chatbot. Unlike Gemini, which sometimes produces inaccurate information or "hallucinates" due to its probabilistic nature, AlphaEvolve uses a unique method to increase speed and accuracy in solving complex algorithm problems.
DeepMind explains that AlphaEvolve works with an automatic evaluation system. Researchers input the problem along with the solution or exploration path possibilities, then this AI produces various solutions using two versions of Gemini: the fast and lightweight Gemini Flash, and the more detailed Gemini Pro. Each solution is then rigorously evaluated by the system, allowing AlphaEvolve to select and perfect the best solution through an evolutionary process.
Unlike other AIs that usually focus on specific fields, such as AlphaFold which is specific to protein folding, AlphaEvolve is flexible and versatile. DeepMind states that this AI can assist in research on various programming and algorithm problems. Google has already started implementing AlphaEvolve in various business lines with promising results.
To use AlphaEvolve, users must provide a description of the problem to be solved, and can add instructions, equations, code snippets, or related literature as references. Users also need to provide an automatic evaluation mechanism, usually in the form of mathematical formulae.
Since AlphaEvolve can only solve problems that can be automatically evaluated, this system is most effective for fields such as computer science and system optimization. However, this AI is less suitable for non-numeric problems because the solutions it produces are only algorithms.
In tests, DeepMind gave about 50 mathematics problems from various branches, such as geometry and combinatorics. AlphaEvolve successfully rediscovered 75 percent of the known answers and found better solutions in 20 percent of the cases.
In addition to mathematics problems, AlphaEvolve was also tested on practical problems, such as improving Google's data center efficiency and speeding up AI model training. DeepMind reports that the generated algorithms were able to save 0.7 percent of Google's global computing resources and reduce Gemini model training time by 1 percent.
Although not always resulting in revolutionary discoveries--for example, it only improved the design of the previously recognized AI TPU chip--DeepMind asserts that AlphaEvolve greatly saves time and enables experts to focus on more important tasks.
Furthermore, AlphaEvolve also enhances generative AI efficiency, which heavily relies on matrix multiplication operations. The most efficient 4x4 complex matrix multiplication algorithm known so far was created by mathematician Volker Strassen in 1969 and has lasted for decades.
However, DeepMind claims that AlphaEvolve has managed to find a more efficient new algorithm. Despite being a general AI, its results are even better than AlphaTensor, the previous specialized AI agent that handled similar problems.
Next-generation Tensor hardware from Google will also benefit from AlphaEvolve. This AI optimizes the Verilog hardware programming language on chips by eliminating unnecessary bits to increase efficiency. Google is still verifying these changes and hopes to apply them to future processors.
Currently, only Google has full access to AlphaEvolve. Although it uses fewer computing resources than AlphaTensor, the system's complexity is still too high to be publicly released. However, the automatic evaluation approach that makes AlphaEvolve so powerful has the potential to be adapted to smaller AI devices for future research.
Editor's Choice: Google's Gemini AI to Launch on Android Auto-Equipped Cars
Click here to get the latest news updates from Tempo on Google News
Microsoft Announces Major Layoffs, Affecting 6,000 Employees
3 hari lalu
Microsoft announced a mass termination of employment affecting over 6,000 employees, approximately three percent of its global workforce, on Tuesday.
Google's Gemini AI to Launch on Android Auto-Equipped Cars
4 hari lalu
Google's Gemini AI is set to be available on Android Auto-equipped cars, including the Renault R5 and the Honda Passport, in a couple of months
Google Cloud Expands AI Data Center in Jakarta; Predicted to Contribute US$88bn to Economy
5 hari lalu
Google Cloud has announced an expansion of the computing capacity of its Jakarta Cloud Region.
Google Rolls out AI-Powered Chrome Tools to Protect Users Against Scams
5 hari lalu
Google announced that it will use Artificial Intelligence (AI) to assist in protecting Chrome browser users from scams.
Indonesia Develops AI-Based Diagnosis System to Identify Malaria Parasites
6 hari lalu
Indonesia's National Research Agency BRIN develops AI-based diagnosis system to assist healthcare workers in identifying malaria parasites.
Google Maps Draining Your Battery? Try These 10 Power-Saving Tips
6 hari lalu
As reported by Tech Wiser, here are several ways to save battery when using the Google Maps application.
Pope Leo XIV Begins Papacy with Cardinal Meeting, Raises AI Concerns
8 hari lalu
Pope Leo XIV holds his first meeting with cardinals from all over the world. He touches on the topic of AI
How Indonesia's AI System Detects Malaria
10 hari lalu
For a long time, malaria diagnosis has been done manually through a microscope, making the process prone to errors due to fatigue or limitations of expert personnel.
Spotify Adds New Tools to Enhance Playlists and Music Discovery for Premium Users
10 hari lalu
Spotify's update for premium users includes a redesigned Now Playing view with improved Queue controls and easier access to Shuffle or Repeat.
Indonesia Launches AI-Powered ZeroStunting to Combat Child Malnutrition
12 hari lalu
ZeroStunting aims to use AI to track children's nutrition, especially egg consumption, and empower parents with tools to improve their kids' diets.
