On May 7, 2026, Google DeepMind reported that AlphaEvolve had moved from pilot testing into the core of its computing infrastructure. The AI coding agent was being used regularly to improve the design of the next generation of Google’s AI chips.[7]
A year earlier, DeepMind had reported an AI system that could make a small part of AI training run faster. Now its developers were describing a tool that helped design the hardware for what comes next.[6][7]
I think we lose sight of the pace when we keep telling the first story after the next chapter has arrived.
The story that got our attention
In February 2025, José R. Penadés asked Google whether it had access to his computer. Its AI co-scientist had proposed, in 48 hours, the same explanation his team had spent years testing. They hadn’t published it. Google said it had no access to his computer.[5]
The scientists’ decade included experiments the AI hadn’t performed.[5] Their results let them recognise that the machine had offered a good explanation. The team was also investigating another idea the system had suggested.[5]
We had a striking glimpse of AI helping scientists choose what to investigate. We didn’t yet have evidence that whole research programmes would finish faster.
In May 2025, DeepMind gave us another glimpse. AlphaEvolve improved a calculation used in Gemini training, making that part run 23% faster and reducing overall training time by 1%.[6]
The saving caught my attention because a system powered by Gemini had helped improve the process used to train Gemini.[6] A gain had gone back into the machinery that makes the next AI.
We should have kept following that machinery.
A year of work behind the screen
DeepMind’s May 2026 update described work across computing, biology and other fields. It reported 30% fewer errors in detecting differences in DNA after AlphaEvolve helped improve a model used to correct DNA readings.[7]
In computing, DeepMind reported that AlphaEvolve had found better rules for deciding what data to keep in fast-access memory. It said the work took two days, compared with months of intensive human effort previously.[7]
Both results came from the system’s developer. They deserve independent scrutiny. They also give us specific changes to examine, rather than another general promise that AI will accelerate science.
The progress was spreading beyond Google’s own systems. A June 30, 2026 case study described work at Schrödinger, a company that builds software for scientists. It used AlphaEvolve to improve code for models of energy and forces in molecules. Google Cloud reported that both training those models and running them became four times as fast.[8]
The fourfold gain applies to those computer calculations. Scientists still have to find promising molecules, test them and learn whether they work safely. A faster calculation doesn’t settle those questions.
Faster computing changes what researchers can afford to explore. Given the same computing time, a faster model can let them examine more possibilities. Given the same task, it can let them reach the next decision sooner.
The 2025 training result and the 2026 molecular result concern different jobs. We cannot turn 1% and fourfold into a single growth rate. What we can see is the method being applied to more of the work behind research.
The checking is changing too
More ideas and faster code bring another problem: researchers need to know which results deserve their trust.
Schrödinger’s case study described tests that AI-generated programs had to pass. The code had to produce correct results, including on difficult molecular systems, and run faster than the baseline.[8] A persuasive explanation couldn’t substitute for those checks.
A September 14, 2026 research paper, HypoEvolve, examined how teams of AI agents could develop better research ideas. The team varied how agents chose, combined and revised ideas, then checked their proposals against biological data kept outside the search process.[9]
They reported higher average scores than six other methods on both outside checks.[9] The work was reported in a preprint. The study checked ideas against existing data; it didn’t establish new treatments for patients.
I find the design lesson important. Researchers were testing how to organise the AI’s work and how to check its output. They were asking whether a team could do better than a single answer, and whether the improvement survived a check beyond the team’s own judgement.
Better models matter. So do better ways of putting them to work.
Where the curve begins to bend
In our first essay, I used a penny that doubles each day. AI has no such fixed schedule. The connection is that an improvement can become part of the starting point for the next improvement.
A faster molecular model can help researchers test more ideas. Better research software can save time across many projects. An AI-designed change to a chip can go into hardware used to run later AI systems. Each successful change can reach further than the task that first produced it.
We haven’t measured a dependable exponential rate across science. We have dated reports of AI helping improve the tools used for research and for building AI, followed by broader applications a year later.[6][7][8]
We often judge progress by the chatbot on our screen. Its mistakes are easy to see. Changes inside a chip design or a research program are harder to notice, even when they help determine what the next system can do.
Which day are we on? Part of the answer lies in what AI can do today. Another part lies in how much it is helping people build tomorrow’s tools.
The first surprising result deserves its place in the story. We also need to notice when researchers stop treating the tool as a demonstration and start using it in the work that produces the next generation.
Sources
[5] BBC: Penadés and the AI co-scientist — February 2025
[6] Google DeepMind: AlphaEvolve — May 2025
[7] Google DeepMind: AlphaEvolve’s expanding impact — May 2026
[8] Google Cloud: Schrödinger’s fourfold speedup — June 2026