Real Brains Mix Neuron Types. AI Uses One — And It's Costing Us

Your brain doesn't use one kind of neuron. It has steady ones, bursting ones, fast-firing ones — all mixed together in the same circuit. Artificial intelligence does the opposite: every 'neuron' in a typical AI network uses the exact same rule for deciding what signal to pass along. That rule is called an activation function, and sticking to just one of them puts a ceiling on what the network can figure out.

A team of researchers (Romain Claret, Michael O'Neill, Paul Cotofrei and Kilian Stoffel), publishing at the ALIFE 2026 conference, let evolution pick the rule for each neuron individually, from a menu of 18 options. They ran over 4,500 experiments covering simple logic puzzles, number prediction and image-style sorting. The results were stark. When they tested a classic puzzle where a network must count whether the number of switches turned on is odd or even, every 'always-grows-upward' rule scored zero. Wavy, oscillating rules scored 100%. Middle-of-the-road rules landed somewhere in between, from about 7% to 80%.

Here's the twist that makes it interesting: the wall isn't about what's mathematically possible. Those always-growing rules can solve the puzzle — they just need the right starting numbers, and the standard training method (called gradient descent, which nudges numbers in tiny steps) finds them fine. The barrier is specific to evolution-based search in bare-bones networks. Add feedback loops, and the gap disappears entirely.

The bigger takeaway is about discovery. When the researchers let evolution choose freely, it built weird, lopsided mixtures of rules that a human engineer would never pick by hand — and those mixtures worked. That suggests today's uniform approach to designing AI may be leaving capability on the table. Practical caveat: this is fundamental research about how learning works, not a tool you can use. Any payoff is years out, and most such findings stay in the lab.