Jason Wei

wiki
, , , ,
1

American AI researcher who helped define the reasoning era of large language models through chain-of-thought prompting, instruction tuning, and the study of emergent abilities.

Profile
Born c. 1998, United States
Nationality American
Current Institution(s) Meta Superintelligence Labs
Research Areas Large Language Models, Chain-of-Thought Prompting, Instruction Tuning, Emergent Abilities, Reasoning, Reinforcement Learning
Education BA, Computer Science, Dartmouth College (2020)
Website jasonwei.net
X / Twitter @_jasonwei
GitHub jasonwei20
Google Scholar Jason Wei

Overview

Jason Wei is an American AI researcher whose work on chain-of-thought (CoT) prompting, instruction tuning, and emergent abilities has shaped how the field understands and trains large language models. He rose to prominence at Google Brain between 2020 and 2023, co-authoring a set of papers that became foundational references in the LLM literature. He then joined OpenAI, where he was a core contributor to the reasoning model series culminating in o1. In July 2025, Wei and longtime collaborator Hyung Won Chung moved together to Meta Superintelligence Labs, continuing their focus on reasoning and reinforcement learning. Despite holding no doctoral degree, Wei is among the most-cited researchers of his generation in NLP, a trajectory built almost entirely from industry research positions.

Early Life & Education

Wei attended Thomas Jefferson High School for Science and Technology in Alexandria, Virginia, a magnet school known for producing competitive STEM talent. He enrolled at Dartmouth College in 2016, studying computer science under professors Lorenzo Torresani and Soroush Vosoughi. While still an undergraduate he published his first conference paper — an NLP data augmentation technique — at EMNLP 2019, an unusually early research debut. He graduated in 2020 with a bachelor’s degree and moved directly into industry research.

Career

Google Brain (2020–2023)

Wei joined Google Brain in October 2020 as an AI Resident, a competitive one-to-two-year fellowship for early-career researchers. He was promoted to Research Engineer in December 2021, to Research Scientist in June 2022, and to Senior Research Scientist in October 2022, an unusually fast progression.

During this period he was the lead or a central contributor on three papers that collectively reoriented LLM research. The FLAN work (2021) demonstrated that instruction fine-tuning dramatically improves zero-shot generalization. The chain-of-thought paper (NeurIPS 2022) showed that prompting models to produce intermediate reasoning steps enables reliable multi-step problem solving. The emergent abilities paper (TMLR 2022) characterized the phenomenon by which qualitatively new capabilities appear at discrete scale thresholds — a finding that influenced how practitioners thought about model scaling and evaluation. He also contributed to the PaLM and Med-PaLM papers as a supporting author.

OpenAI (2023–2025)

Wei announced his move to OpenAI’s ChatGPT team in February 2023. At OpenAI he worked on reasoning and agentic systems, eventually becoming a co-creator of the o1 model series, released in preview in September 2024. The o1 models are trained via reinforcement learning on chain-of-thought traces, enabling substantially higher performance on mathematics, science, and programming benchmarks compared to prompt-based CoT alone. Wei also contributed to the deep research product and co-authored SimpleQA (a short-form factuality benchmark) and BrowseComp (a benchmark for browsing agents). During this period he became a vocal proponent of reinforcement learning as both a technical paradigm and a personal philosophy, describing himself publicly as an “RL diehard.”

Meta Superintelligence Labs (2025–present)

In July 2025, Wei and Hyung Won Chung left OpenAI together and joined Meta’s newly formed Superintelligence Labs. Their departure was part of a broader wave of senior OpenAI researchers moving to Meta as the company ramped up its AGI ambitions with substantial compensation packages. Wei’s expertise in reasoning and reinforcement learning was cited by multiple sources as a key motivation for Meta’s recruitment effort.

Key Contributions

  • Chain-of-Thought Prompting — The 2022 NeurIPS paper “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models” (with Xuezhi Wang, Denny Zhou, and others) demonstrated that prompting a model to articulate intermediate reasoning steps before giving a final answer dramatically improves performance on arithmetic, commonsense, and symbolic reasoning tasks. The technique became standard practice across both research and deployment and remains one of the most-cited papers in NLP.

  • FLAN / Instruction Tuning — “Finetuned Language Models are Zero-Shot Learners” (ICLR 2022) introduced FLAN (Finetuned Language Net), showing that fine-tuning on a diverse collection of NLP tasks formatted as natural-language instructions yields strong zero-shot transfer. The follow-up “Scaling Instruction-Finetuned Language Models” (JMLR 2024) scaled this approach to PaLM and T5, producing FLAN-T5 and FLAN-PaLM, model families widely used in both research and production.

  • Emergent Abilities of Large Language Models — “Emergent Abilities of Large Language Models” (TMLR 2022, with Yi Tay, Rishi Bommasani, Barret Zoph, Percy Liang, Jeff Dean, and others) provided a systematic characterization of capabilities that appear abruptly as model scale increases, framing what had previously been anecdotal observations as a measurable and debated phenomenon in AI scaling research.

  • OpenAI o1 — As a co-creator of the o1 model series (released September 2024), Wei helped advance the paradigm of training LLMs to reason via reinforcement learning on chain-of-thought traces, extending CoT from a prompting trick to a core training objective. The o1 family achieved state-of-the-art results on competition-level mathematics and coding benchmarks at release.

  • SimpleQA — “Measuring Short-Form Factuality in Large Language Models” (2024) introduced SimpleQA, a benchmark of unambiguous, verifiable factual questions designed to calibrate factuality evaluation in LLMs, filling a gap left by longer-form or harder-to-score fact-checking datasets.

  • BrowseComp — “A Simple Yet Challenging Benchmark for Browsing Agents” (2025) proposed a benchmark for evaluating the ability of AI agents to retrieve information through multi-step web browsing, contributing to the emerging evaluation infrastructure for agentic systems.

  • EDA (Easy Data Augmentation) — The 2019 EMNLP paper introduced simple text augmentation operations — synonym replacement, random insertion, swap, and deletion — that consistently improved text classification performance. It became a widely used practical baseline in NLP.

Awards & Recognition

  • NeurIPS 2022 Spotlight / Oral — The chain-of-thought paper was presented at NeurIPS 2022 and has accumulated tens of thousands of citations, ranking among the most influential NLP papers of the 2020s.
  • ICLR 2022 — The original FLAN paper was accepted at ICLR 2022 and is among the foundational references for instruction-tuned language models.
  • Invited keynotes — Wei has delivered keynotes at KDD LLM Day (2023), WebConf LLM Day (2024), OpenAI DevDay San Francisco (2024), and Columbia University DAPLab (2025), among many other invited talks at Stanford, MIT, Princeton, Berkeley, and major industry conferences.

Key Relationships

  • Hyung Won Chung — Closest research collaborator; overlapped at Google Brain and OpenAI, co-authored the Flan-T5 scaling paper, and moved together to Meta Superintelligence Labs in July 2025.
  • Denny Zhou — Co-author and collaborator on the chain-of-thought prompting work at Google Brain; Google DeepMind principal scientist known for reasoning research.
  • Yi Tay — Co-author on the emergent abilities paper and several scaling papers at Google Brain; later moved to Reka AI.
  • Quoc Le — Senior co-author on multiple Google Brain papers including FLAN and chain-of-thought; Google Brain research director and a key figure in Wei’s early career.
  • Lorenzo Torresani — Dartmouth undergraduate research advisor; professor of computer science whose mentorship preceded Wei’s industry career.
  • Soroush Vosoughi — Dartmouth computer science professor who worked with Wei on undergraduate research projects and later invited him to guest lecture.
  • Zhiqing Sun — Research scientist collaborator at OpenAI; co-first author on BrowseComp; also joined Meta Superintelligence Labs in July 2025.

Personal Style

Wei’s research practice is characterized by a preference for clean, empirically grounded ideas that travel well — techniques that can be explained in a sentence and that generalize broadly across models and tasks. His public writing, collected on his personal site under a “Thoughts” section, tends toward short, direct observations about research methodology, the mechanics of reinforcement learning, and career development, with a candid quality unusual among researchers at his profile level. He has spoken openly about how immersion in reinforcement learning concepts has shaped his personal philosophy, particularly the value of exploring independently rather than imitating. His trajectory from undergraduate researcher to senior scientist at three of the world’s most competitive AI labs without a graduate degree is frequently noted as distinctive in a field dominated by PhD holders.

References