Why large language models still don't truly reason, argues AI researcher
A researcher who helped build a Go-playing program recalls a famous 2016 Seoul match, where an unexpected move stunned commentators watching the human-versus-machine game. The memory frames a bigger argument, reported by MIT Technology Review: today's large language models (LLMs), the AI systems behind chatbots, may look like they're reasoning, but they aren't doing it the way humans do. The piece questions whether producing confident, fluent answers is the same as genuine understanding. This matters because many products and decisions now lean on LLMs, assuming they "think." If that assumption is wrong, it could affect how much we trust these systems in high-stakes situations.
Words to know
- Large language models (LLMs)
- — AI systems trained on huge amounts of text to generate human-like responses.
- Reasoning
- — The ability to think through problems logically, step by step, like humans do.
- Go
- — An ancient strategy board game once considered extremely hard for computers to master.
Summary written from MIT Technology Review's headline and teaser