Google AlphaEvolve halves training time

Algorithms are the real story here — not chatbots, not copilots, but the code that decides how efficiently huge systems run. That matters because modern AI is expensive in ways people rarely see: training time, chip utilization, data-center scheduling, routing, simulation, all of it. The gap has been that language models are good at suggesting code, but bad at proving a change actually helps. AlphaEvolve is Google’s answer to that, and this week Google used a one-year update to show it moving from flashy research into real infrastructure and customer-facing cloud product territory. ### What is AlphaEvolve, exactly? It’s a coding agent from Google DeepMind that uses Gemini models to generate possible algorithm changes, then tests those changes with automated evaluators. Basically, it doesn’t just write code and hope for the best. It runs candidate programs, scores them, keeps the winners, and uses that feedback loop to evolve better versions over time. Google first introduced it in May 2025 as a system for algorithm discovery and optimization. (blog.google) ### Why is that different from a normal coding agent? A normal coding agent mostly helps a human write software faster. AlphaEvolve is aimed at a narrower but deeper problem — finding better algorithms where “better” can be measured. That measurement piece is the trick. If you can define an objective evaluator, like lower latency or fewer errors, the system can search through lots of code variants and keep improving. That makes it more like an optimization lab than an autocomplete tool. (deepmind.google) ### Did it really halve training time? No — at least not in Google’s own published numbers. The strongest official training claim is that AlphaEvolve sped up a vital kernel in Gemini’s architecture by 23%, which translated into a 1% reduction in overall Gemini training time. That is useful at Google scale, but it is nowhere near “2× faster model training.” So the viral version of the story overstates what Google has actually said publicly. (cloud.google.com) ### Where has Google actually used it? Inside Google’s infrastructure first. The company says AlphaEvolve found a better task-scheduling method for data centers that continuously recovered an average of 0.7% of Google’s global compute resources. It also worked on Gemini training kernels and earlier helped with data-center, chip-design, and AI-training processes. Those are boring-sounding targets, but they are exactly where tiny gains compound into huge savings. (cloud.google.com) ### What changed this week? Google’s May 7, 2026 update widened the picture. Instead of framing AlphaEvolve mainly as a math-and-computer-science research system, Google showed it working across genomics, power-grid optimization, earth-science modeling, molecular simulation, and neuroscience. The message was simple: this is no longer just an internal curiosity for proving clever theorems. It is becoming a general-purpose optimizer for any problem with a scoreable outcome. (cloud.google.com) ### What are the biggest non-AI results? The genomics result is probably the easiest to grasp. Google says AlphaEvolve improved DeepConsensus, a DNA-sequencing error-correction model, cutting variant-detection errors by 30%. In grid optimization, it pushed a trained model’s feasible-solution rate from 14% to more than 88%. In disaster prediction, it improved aggregate accuracy by 5% across 20 natural-hazard categories. Those are not toy benchmarks — they are domain results with operational consequences. (blog.google) ### Is this a Google Cloud product now? Yes, but with a catch. Google Cloud put AlphaEvolve into private preview in December 2025, so this is not a brand-new launch. What’s new is the stronger proof story around it. Google is now tying the cloud offering to a year of internal deployment and broader scientific wins, which makes the pitch feel less speculative. (deepmind.google) ### What’s the real takeaway? The point is not that Google found a magic button that halves AI costs overnight. The point is that it built a system that can keep squeezing measurable gains out of expensive, messy, real-world code — and then turned that into a cloud product. If that scales, the biggest AI advantage may come from better algorithms, not just bigger models. (blog.google) (cloud.google.com)