Google DeepMind unveils AlphaEvolve

Background

A language model can write code, but producing code that improves on what experts have already achieved is a harder test. A coding agent is a system that does not just autocomplete a line of code: it plans an approach, writes a program, runs it, looks at the result, and tries again. Google DeepMind built its entry in this category, AlphaEvolve, on top of Gemini.^[1]

The target was algorithms — precise, step-by-step procedures for solving a problem. The cost of an algorithm is often what matters most in practice: how much computing a task consumes, or how many elementary operations it needs. Two of the areas Google DeepMind highlighted, scheduling work across computers and multiplying matrices, are exactly of this kind, where a small change in procedure can shift a large amount of work.

What was announced

On 14 May 2025, Google DeepMind announced AlphaEvolve, describing it as a Gemini-powered coding agent for designing advanced algorithms.^[1] The announcement presented it through concrete results rather than through a product launch, and reported what the system had produced in several different areas at once.

The results

In computing infrastructure, AlphaEvolve found a better scheduling heuristic and thereby recovered on average 0.7% of Google's worldwide compute.^[1] A scheduling heuristic is a rule of thumb that decides which job goes where and when; it does not change the hardware, only the order and placement of work. Because the same rule is applied across an entire fleet, even a modest improvement in how well the machines are kept busy adds up to a large amount of capacity in absolute terms.

In mathematics, the agent found a 48-multiplication algorithm for 4×4 complex matrices, improving on Strassen's 1969 result.^[1] The usual way to measure the cost of a matrix multiplication procedure is the number of scalar multiplications it performs, so a procedure that needs fewer of them is strictly cheaper, whatever the hardware. Strassen's 1969 method had stood as the reference point for this case for more than half a century.

More broadly, Google DeepMind tested the system on more than 50 open mathematical problems, and it improved the best known solutions on about 20% of them.^[1] That figure is worth reading carefully: in the large majority of cases the system did not beat the existing record. What the announcement shows is a method that succeeds often enough, across a wide set of problems, to be useful as a search tool for researchers.

Access

At the time of the announcement the system was not publicly released. Google DeepMind said it planned an Early Access Program for selected academic users.^[1] The team also said it was building a user interface for AlphaEvolve together with the People + AI Research team,^[1] and that interested users could register through a sign-up form linked from the announcement.^[1]

Why it matters

Two things stand out in the announcement. The first is the range: the same kind of system was applied to an operational problem inside a company's own infrastructure and to questions in pure mathematics, where progress is usually measured in single results published after long effort. The second is the size of the gains. A fraction of a percent of worldwide compute, or one fewer multiplication in a small matrix product, is not dramatic on its own; but these are the quantities that compound, because a scheduling rule runs continuously and a matrix multiplication sits inside countless larger computations.

The early access arrangement also sets the immediate expectation. The system was presented as a research result with a limited path for academic users, together with work on an interface, rather than as a generally available tool.^[1]