New AI Learns Like a Human, Mixing Language and Code to Make Sense of the World

For decades, scientists have wondered how people learn so much from so little. A baby sees a few dogs and knows what a dog is. Adults pick up new ideas from a single conversation. Computers, on the other hand, usually need millions of examples. A new paper from researchers at MIT and other institutions offers a fresh answer: the brain may act like a hybrid system, mixing everyday language with the logic of code.

The team created a model that represents knowledge as "mental programs" — not pure words, not pure equations, but a combination of both. Think of it as your brain writing a quick recipe for itself, where the ingredients are described in language (like "round fruit") and the steps follow the rules of code (like "if red, it may be ripe"). The model updates its ideas step by step, much like a person revising their beliefs after new evidence.

What makes this special is how it was tested. The model didn't just solve puzzles better than existing AI. It also mimicked real human quirks. For example, people tend to "anchor" on the first piece of information they hear, and they sometimes head down the wrong path before catching themselves — a "garden path." The model did the same, suggesting it captures something deep about human reasoning. Pure large language models, like the ones behind ChatGPT, either failed the tests or needed absurd amounts of computing power. Classic Bayesian models, which many scientists used before, couldn't match human behavior either.

Why should you care? Because machines that learn like us could eventually mean AI tutors that figure out exactly what you're confused about, personal assistants that ask the right questions, and robots that adapt to your home without needing thousands of instructions. This research is still early, but it points toward a future where AI doesn't just spout answers — it learns the way you do, one curious question at a time.