Jason Wei
Last reviewed
Sources
14 citations
Review status
Source-backed
Revision
v1 · 1,734 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
14 citations
Review status
Source-backed
Revision
v1 · 1,734 words
Add missing citations, update stale details, or suggest a clearer explanation.
Jason Wei is an American artificial intelligence researcher known for foundational work on the reasoning and scaling behavior of large language models. While a research scientist at Google Brain, he was the lead author of three influential papers that shaped how the field understands modern language models: chain-of-thought prompting, the emergent abilities of large models, and instruction tuning through the FLAN project. He later joined OpenAI, where he helped build the company's o-series of reasoning models and its Deep Research agent, and in July 2025 he moved to Meta Superintelligence Labs. [1][2][10]
Wei grew up in Fairfax County, Virginia. [3] He attended Dartmouth College, where he majored in computer science and graduated with the class of 2020. He has recounted that he first encountered machine learning during the summer after his first year, while interning at the AI startup Protago Labs. [3]
As an undergraduate, Wei concentrated on applying machine learning to medical imaging and computer vision, working in computational pathology under Dartmouth faculty including Saeed Hassanpour and Lorenzo Torresani. [3][4] One project produced an algorithm for analyzing lung-cancer histology images that reportedly matched or exceeded pathologists in testing, and by his senior year he had co-authored several peer-reviewed papers. [3] In 2019 he received a Goldwater Scholarship, one of the most competitive United States undergraduate awards in the natural sciences, mathematics, and engineering. [3] His 2020 senior thesis, "Label Noise Reduction Without Assumptions," advised by Lorenzo Torresani, proposed a method for training neural networks on noisy labels that up-weights per-example gradients similar to others in the same minibatch, without requiring assumptions about the noise or a clean validation set. [4]
Wei's most cited undergraduate contribution was a 2019 paper written with Kai Zou, "EDA: Easy Data Augmentation Techniques for Boosting Performance on Text Classification Tasks," presented at EMNLP 2019. It introduced four simple text-editing operations, synonym replacement, random insertion, random swap, and random deletion, and became a widely used baseline for data augmentation in natural language processing. [5]
Contrary to some secondary profiles that describe Wei as holding a doctorate, his Dartmouth degree is an undergraduate one. He entered industry research directly through Google's AI Residency rather than by completing a PhD. [2][3]
Wei joined Google Brain in October 2020 as an AI Resident, a program intended to bring early-career researchers into the lab. He was promoted to research engineer in December 2021, research scientist in June 2022, and senior research scientist in October 2022. [2] He left Google in February 2023 to join OpenAI, where his work centered on reasoning and AI agents, and he remained there until mid-2025 before moving to Meta. [1][2][10]
| Period | Organization | Role |
|---|---|---|
| Oct 2020 to Dec 2021 | Google Brain | AI Resident |
| Dec 2021 to Jun 2022 | Google Brain | Research Engineer |
| Jun 2022 to Oct 2022 | Google Brain | Research Scientist |
| Oct 2022 to Feb 2023 | Google Brain | Senior Research Scientist |
| Feb 2023 to Jul 2025 | OpenAI | Research Scientist |
| Jul 2025 to present | Meta Superintelligence Labs | Researcher |
At OpenAI, Wei was part of the team behind the company's o-series of reasoning models. When OpenAI o1 was announced in September 2024, he publicly summarized its central ideas: instead of eliciting a chain of thought purely through prompting, the model is trained to reason with reinforcement learning, and its chain of thought acts as a form of adaptive compute that can be scaled at inference time. [12] According to reporting on his later departure, his main focus at OpenAI subsequently shifted to the o3 model and the company's Deep Research agent. [10]
During this period Wei also created several widely adopted evaluation benchmarks. SimpleQA, released in 2024 with Karina Nguyen as joint first author, measures the short-form factuality of language models by testing them on fact-seeking questions with single, verifiable answers. [14] BrowseComp, released in 2025, is a benchmark for web-browsing agents built around questions that are hard to answer but easy to verify. [2]
Meta Superintelligence Labs (MSL) was announced by Mark Zuckerberg on June 30, 2025, as a consolidation of Meta's foundation-model, product, and research teams, including the Llama effort. The unit is led by former Scale AI chief executive Alexandr Wang as Meta's Chief AI Officer, with former GitHub chief executive Nat Friedman leading AI products. [11]
On July 16, 2025, TechCrunch and other outlets reported that Wei would join MSL together with his longtime collaborator Hyung Won Chung, with whom he had worked closely at both Google and OpenAI. [10] Wei's own profile lists Meta Superintelligence Labs as his current employer. [1] The hires came amid an aggressive recruiting drive in which, according to multiple reports, Meta offered leading AI researchers compensation packages worth as much as 300 million US dollars over four years, although the specific terms of Wei's agreement were not disclosed. [10][11]
Wei is most closely associated with three ideas, all developed during his time at Google Brain, that became central to how researchers describe and build large language models.
| Year | Paper | Venue | Theme |
|---|---|---|---|
| 2019 | EDA: Easy Data Augmentation Techniques | EMNLP | Text data augmentation |
| 2022 | Finetuned Language Models Are Zero-Shot Learners (FLAN) | ICLR | Instruction tuning |
| 2022 | Chain-of-Thought Prompting Elicits Reasoning in LLMs | NeurIPS | Reasoning and prompting |
| 2022 | Emergent Abilities of Large Language Models | TMLR | Scaling behavior |
| 2024 | Scaling Instruction-Finetuned Language Models | JMLR | Instruction tuning |
| 2024 | Measuring Short-Form Factuality (SimpleQA) | preprint | Evaluation |
In "Finetuned Language Models Are Zero-Shot Learners," presented at ICLR 2022, Wei and colleagues introduced instruction tuning, which finetunes a pretrained language model on a large collection of tasks phrased as natural-language instructions. The resulting model, called FLAN (Finetuned Language Net), substantially improved zero-shot performance on tasks it had not been trained on, demonstrating that instruction-style supervision helps models generalize. [6] The approach was scaled in the 2024 follow-up "Scaling Instruction-Finetuned Language Models," led by Hyung Won Chung with Wei among the authors, which produced the Flan-T5 and Flan-PaLM model families. [13] Instruction tuning has since become a standard ingredient in aligning chat-oriented language models.
"Chain-of-Thought Prompting Elicits Reasoning in Large Language Models," published at NeurIPS 2022 with Wei as lead author, showed that prompting a model with a few worked examples containing intermediate reasoning steps markedly improves its performance on arithmetic, commonsense, and symbolic reasoning tasks. In one widely cited result, prompting a 540-billion-parameter PaLM model with eight chain-of-thought exemplars achieved state-of-the-art accuracy on the GSM8K grade-school math benchmark. [7] Co-authored with Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou, the paper became one of the most heavily cited works in modern AI and helped lay the conceptual groundwork for later reasoning models that learn to "think" before answering. [7][12]
In "Emergent Abilities of Large Language Models," published in Transactions on Machine Learning Research in 2022, Wei and co-authors, including Yi Tay, Rishi Bommasani, Percy Liang, Jeff Dean, and William Fedus, defined emergent abilities as capabilities that are absent in smaller models but appear, often sharply and unpredictably, once models reach a sufficient scale. [8] The framing was influential but also drew scrutiny: a 2023 paper by Rylan Schaeffer and colleagues, "Are Emergent Abilities of Large Language Models a Mirage?," argued that some apparent emergence is an artifact of discontinuous evaluation metrics rather than a genuine property of the models, and won an outstanding paper award at NeurIPS 2023. [9] The exchange became a reference point in debates over how the capabilities of large models scale.
Beyond these signature works, Wei contributed to several large Google research efforts, including PaLM (JMLR 2023) and the medical model Med-PaLM, whose results were reported in the 2023 Nature paper "Large Language Models Encode Clinical Knowledge." [2]
Wei's research has been recognized primarily through its broad adoption: chain-of-thought prompting, instruction tuning, and the language of "emergent abilities" all became standard vocabulary and technique in the study of large language models. His 2019 Goldwater Scholarship marked early recognition during his undergraduate years. [3] He has been a frequent invited speaker, delivering keynotes and guest lectures at institutions including Stanford, MIT, Harvard, Princeton, Columbia, the University of California, Berkeley, the University of Pennsylvania, and the National University of Singapore between 2021 and 2025, as well as at OpenAI's DevDay. [2]