Brain-Inspired AI Could Slash the Power Bill of Every Chatbot
Less energy per answer means cheaper, more private AI that fits in your pocket.
The AI that powers ChatGPT and most chatbots is called a Transformer. It's remarkably capable, but also remarkably hungry — training and running these systems consumes enormous amounts of electricity. Two researchers, Matei-Ioan Stan and Oliver Rhodes, have proposed a different design called ADPTNet, inspired by how the brain's hearing centre processes sound. Their goal is an AI that is cheaper to run but still smart enough to compete.
The brain angle matters. Your brain runs on roughly 20 watts, about a dim light bulb, yet it understands speech and remembers conversations. One reason is that the auditory cortex works on fixed internal rhythms rather than constantly readjusting. ADPTNet copies that idea, using some advanced maths to keep its internal clock steady and predictable. The payoff: it tracks long-range patterns better than Mamba, a well-known efficient model, and matches Transformers on a sequence-of-images task while using fewer parts.
The team also built a version called SpikingADPTNet, which mimics neurons that fire in short pulses rather than a constant stream — more like Morse code than a steady hum. That version set a new best score of 83.56% on the Spiking Speech Commands dataset, a standard test for recognising spoken words from event-based audio. The authors say their method also makes a common parallel-processing technique work for this class of models for the first time.
The catch: this is an 80-page academic paper, not a product. The wins come from narrow, controlled tests, not real chatbots answering your questions. Its rigid internal timing could also limit flexibility on messier real-world tasks. Still, the direction matters — lower-power AI could ease pressure on data centres, cut costs, and eventually put capable assistants on tiny devices like earbuds and hearing aids.
- Transformers, the tech behind ChatGPT, are powerful but hugely power-hungry — this new design aims to do similar work for far less energy.
- It borrows the brain's hearing system, which runs on roughly 20 watts, by keeping a steady internal rhythm instead of constantly readjusting.
- One version hit 83.56% accuracy on a standard speech test, beating older spiking models — but it's still early research, not a product you can try.
Why It Matters
Cheaper, lower-power AI could mean faster, more private tools running on your phone instead of distant data centres.