AI Used AI to Design a Brain-Like Model That Sips Power

A team of researchers has found a new way to build AI that runs on far less electricity — by copying how brain cells actually work. Their approach, called spiking neural networks (AI that fires in short bursts, like neurons, instead of running nonstop), uses energy only when there's something to react to. Think of the difference between leaving every light in the house on versus only lighting the room you're standing in. Existing AI models mostly burn power continuously, which is why running them is so expensive.

The clever part is how they found these designs. Instead of hand-building them, the team let large language models (the AI behind chatbots) write and rewrite the model's code over many rounds, keeping whatever worked and improving on it — like a very fast, tireless engineer iterating on a blueprint. The search cost the equivalent of 132 days of work on specialized graphics chips, and it produced several original designs, one of which beat a comparable standard model on a common text task.

The headline number: up to 50.6 times less arithmetic energy at the architecture level compared with a typical dense Transformer, the design behind most chatbots today. If results like that hold up in real products, it could mean smaller electricity bills for AI companies, less strain on power grids, and the ability to run capable AI on phones, watches, or sensors without draining the battery.

But be careful with the hype. This is a research paper, not a product. The energy savings are estimated, not measured on real hardware, and no code or models have been released yet — the authors say they're coming soon. Also, spiking systems are still hard to train and rarely used in commercial AI. Think of this as a promising signpost, not a finished road. Still, it shows AI helping design better, cheaper AI — a loop that could speed everything up.