type
Post
status
Published
date
Nov 18, 2026
slug
summary
tags
category
行业动态
icon
password
Jason Wei Leaves Meta for Mercor, Betting on Data-Driven AI Frontiers
Jason Wei, the researcher known for co-authoring the influential "Chain of Thought" paper, has announced his departure from Meta. After spending just over a year at the company, Wei has joined the startup Mercor as its Chief Scientist. This move marks a significant shift in the AI landscape, highlighting the growing importance of specialized data in training next-generation models.
A New Role in Data-Centric AI
Wei announced the transition on X, stating that his move to Mercor is driven by the belief that data will shape the next frontier of AI. At Mercor, he will oversee all data and evaluation operations, helping to build the research team. He will work alongside Edward Hu, the first author of the LoRA paper, who is responsible for model training.
The transition appears well-planned rather than impulsive. Meta has been a client of Mercor for several years, and Wei’s team at Meta’s TBD Lab had previously utilized Mercor’s data annotation services. Mercor’s CEO, Brendan Foody, noted that the company had been in close communication with Wei for years. This relationship underscores a broader trend where large tech firms and startups are increasingly intertwined, with data becoming a critical asset for AI development.
Jason Wei’s Impactful Career
Wei’s career has been defined by rapid ascent and foundational contributions to AI. After graduating from Dartmouth College in 2020, he joined Google Brain directly without pursuing a PhD. At the age of 23 in 2022, he became a Google Research Scientist, a position held by few at such a young age.
During his time at Google, Wei published the seminal Chain of Thought paper as the first author. He also led the FLAN project, a key instruction tuning work. These achievements laid the groundwork for the reasoning capabilities that modern large language models exhibit today.
In 2023, Wei moved to OpenAI, where he played a central role in the development of the reasoning model o1 and the Deep Research initiative. He also created benchmark suites such as SimpleQA, BrowseComp, and HealthBench, the latter of which was developed in collaboration with 262 medical professionals. His impact is evident in his academic profile, with over 160,000 citations on Google Scholar.
Last year, Wei joined Meta’s TBD Lab, where he led the development of the Muse Spark model series. Shortly before his departure, his team conducted an experiment using Muse Spark 1.3 to analyze approximately 5 million de-identified electronic medical records. While the model accelerated hypothesis testing, Wei emphasized that the data contained significant noise, requiring professional expertise to interpret the results accurately.
Mercor’s Focus on High-Value Data
Mercor, founded in January 2023 in San Francisco, is positioning itself as a leader in AI training data and evaluation. The company was founded by Brendan Foody, Adarsh Hiremath, and Surya Midha. Mercor’s strategy involves engaging subject matter experts, including doctors, lawyers, and engineers, to break down complex real-world tasks into high-difficulty training problems with rigorous scoring criteria.
The company recently secured significant financial backing. In November 2025, Mercor announced a $750 million Series D funding round, valuing the company at $20 billion. Investors included Benchmark, General Catalyst, and Sapphire Ventures, alongside industry giants like Nvidia, Salesforce, and Zoom. The investment logic suggests that as general internet data becomes exhausted, the "implicit judgment" of senior experts is essential for teaching AI to perform complex professional tasks.
Industry Implications of High-Profile Moves
Wei’s departure is part of a larger pattern of talent mobility in the AI sector. Other researchers have also left Meta within their first year of employment, often due to contractual "cliff periods" where equity vests. This trend highlights the intense competition for top-tier AI talent.
As companies increasingly train proprietary models, organizations like Mercor are playing a crucial role in defining the next phase of AI evolution. By controlling high-quality data and evaluation standards, they are shaping how AI systems learn to handle specialized, high-stakes domains. Wei’s move signals a strategic pivot towards data-centric approaches, recognizing that raw model scale is no longer the only determinant of AI capability.
原文出处:量子位,https://www.qbitai.com/2026/11/510811.html
- 作者:数字Baby
- 链接:myaigc.top/article/41cb7c2d-0e98-81ea-b4d1-f14debbcc13d
- 声明:本文采用 CC BY-NC-SA 4.0 许可协议,转载请注明出处。



