Transcription
So, Google built an AI that evolves, not learns, it evolves. It copies itself thousands of times, kills off the weak copies, and breeds the strong ones. And it just made Google's own AI 23% faster to train.
Let me explain the actual mechanism. So, you hand it a piece of code and a way to grade it. So, speed, cost, accuracy, whatever you're optimizing for. It then generates a batch of mutated copies, runs everyone through your rubric, keeps the winners, and deletes the rest. Then mutates the winners again and again, thousands of generations in a matter of a few minutes.
Google's DeepMind calls it AlphaEvolve. You might have heard of it. And the first place they pointed it was at themselves. It went digging through Gemini's own training code and found one kernel nobody had touched in a while and sped that single piece up by 23%. Now, if you multiply it across the pipeline, it saved about 1% off training the entire model. It doesn't sound huge until you remember what training these models actually cost. A flat 1% is real money in real time.
Then it moved to Google's next generation TPU chips and helped design more efficient circuits for them. Somebody even pointed it at the software that corrects DNA sequencing errors, and it cut the mistake rate by 30%.
Now, all of this is over a year old. The key point here is that AlphaEvolve was mostly inside Google, only accessible internally. But, this month they put it on the Google Cloud, meaning any company can now point it at their own slow, expensive problem, run a few thousand generations, and get real improvements.
So, to summarize, Google built an AI to improve code, and the first thing it improved was the AI that built it. And now they've opened it up for anyone to try out. But, keep in mind, it's pretty expensive. Follow and I'll keep you posted.