Jason Wei is an AI researcher whose papers shaped much of how the field thinks about prompting and reasoning in large language models. He was a research scientist at Google Brain, then worked at OpenAI from 2023 to 2025 on reasoning and agents, and has since been at Meta Superintelligence Labs. His own account credits his Google Brain period with helping popularize “chain-of-thought, instruction tuning, and emergence in large language models” - three ideas that turned out to be central.
He is the lead author of “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models” (arXiv 2201.11903, January 2022), which showed that simply asking a model to work through intermediate steps sharply improved its performance on multi-step problems. He co-authored “Finetuned Language Models Are Zero-Shot Learners” (FLAN), an early demonstration that instruction tuning makes models far better at following natural-language tasks they were not specifically trained on, and “Emergent Abilities of Large Language Models,” which argued that some capabilities appear abruptly at certain scales rather than improving smoothly - a claim that itself became a subject of debate.
At OpenAI, Wei was a co-creator of o1, the reasoning model that learns to think before answering, connecting his earlier work on eliciting reasoning to a model trained to do it natively.
For business readers, Wei’s name is a thread running through the techniques that made today’s models usable: get them to show their work, train them to follow instructions, and expect new abilities to emerge as they scale.