Chain-of-Thought First Author Jason Wei Leaves Meta to Join Data Startup Mercor as Chief Scientist
Jason Wei announced on X that he is leaving Meta's TBD Lab to join data startup Mercor as Chief Scientist, where he will build a research team and oversee data and evaluation work; Edward Hu, first author of the LoRA paper, is joining at the same time to lead model training. Mercor focuses on AI training data and evaluation services, and on November 10 it just closed a $750 million Series D funding round at a $20 billion valuation. Jason Wei is the first author of Chain-of-Thought (CoT) and FLAN, has worked on o1 and Deep Research, and has over 160,000 Google Scholar citations.
Why read this
AI score 55
The path from Chain-of-Thought author to a data startup reveals shifts in where top researchers are heading and in the landscape of training data supply.
Topics
Body · AI translation
Damn! Last year Zuckerberg shelled out a sky-high annual package worth hundreds of millions of dollars to pry away the "chain-of-thought" author Jason Wei from OpenAI—
and just over a year after joining Meta, he up and walked out!
And his destination wasn't some giant, but a data startup that has only just gotten off the ground...
Just moments ago, Jason Wei posted on X to make it official:
he will leave Meta's TBD lab and join the startup Mercor as chief scientist.
In a welcome post, Mercor co-founder and CEO Brendan Foody further revealed that Jason will help build the research team and oversee all data and evaluation work.
Worth noting: fighting alongside Jason is Edward Hu, first author of the LoRA paper, who will be in charge of model training.
Though it came out of nowhere—so much so that Jason hadn't even had time to update his X bio when he made the announcement—this job change wasn't entirely without warning signs.
Reportedly, Meta has been a Mercor customer for the past few years, and Jason's TBD Lab has also used Mercor's data annotation services.
Not only that, the Mercor CEO said he has stayed in close contact with Jason over the past few years.
You could say the big tech company paid a startup for data, and the startup turned around and poached the guy who was buying it.
Jason Wei's move draws so much attention not only because he authored a series of classic works such as chain-of-thought, but also because he is the kind of "little genius" who skipped a PhD and went straight from undergrad into the core of the AI world.
(Though there's no official word, based on Jason Wei's earlier posts, he is only 27 or 28 this year.)
After graduating from Dartmouth College with a bachelor's in computer science in 2020, Jason didn't follow the conventional path into a PhD and instead chose to join Google Brain directly.
In 2022, at just 23, he earned the title of Google "Research Scientist."
(Possibly one of the youngest research scientists in Google's history.)
That same year, as first author, he published the foundational chain-of-thought (CoT) paper.
At the same time, he was also first author of FLAN, a classic work on instruction fine-tuning.
You could say the reason today's large models "think step by step" traces back to him.
He then joined OpenAI in 2023, playing a deep role in the core R&D of the first-generation reasoning model o1 and Deep Research, and also personally built SimpleQA, BrowseComp, and HealthBench, a health evaluation benchmark developed with 262 doctors.
(At the 2024 o1 "fan meetup," the person in the front-row center seat was Jason Wei; next to him was another key o1 figure, Shengjia Zhao, who is still at Meta.)
To date, Jason Wei's Google Scholar citations have exceeded 160,000.
After being poached by Zuckerberg last July for a sky-high price, he joined TBD Lab under the superintelligence lab, where he led development of the Muse Spark series of models and pushed large models into frontier medical research.
Just one month before leaving, he shared an experiment in which his team used Muse for medical research. Medical researchers could call on a coding tool powered by Muse Spark 1.3 and Muse Wonder to analyze about 5 million de-identified electronic medical records.
Doctors asked the questions; the model wrote the code and ran the analyses. In just four weeks, the team produced ten research manuscripts.
Jason had a key observation about this work: although the model lets doctors verify a hypothesis faster and try a few more directions, medical record data is full of noise—which variables are worth looking at and what a set of results actually shows still ultimately depend on professionals to decide.
And Mercor, the company he is joining, happens to be a firm that goes all in on "high-barrier professional data."
What kind of company is Mercor
Mercor was founded in January 2023 and is headquartered in San Francisco.
The three founders, Brendan Foody, Adarsh Hiremath and Surya Midha, are old acquaintances from their high school debate team, and in the early days the three worked out of dorm rooms at Harvard and Georgetown.
Today, this young company has not only recruited top researchers such as Jason Wei and Edward Hu, but also assembled a large network of industry experts behind it, including doctors, lawyers, investment banking analysts and seasoned engineers.
According to its website, Mercor's main business is AI training data and evaluation services. It organizes experts in various fields to break real work down into highly difficult training tasks and sets strict scoring rubrics, helping AI labs and enterprises train models and test their ability to solve complex professional problems.
In addition, Mercor also helps enterprises monetize their own business data and deploy AI in their daily workflows.
After Felicis led both its Series B and Series C, just eight days before Jason announced his move (November 10), Mercor had just announced the completion of a $750 million Series D at a valuation of $20 billion.
Investors include not only top venture firms such as Benchmark, General Catalyst and Sapphire Ventures, but also industry giants such as Nvidia, Salesforce and Zoom.
Reportedly, the core logic behind capital's bet on Mercor is that general internet corpora are nearly exhausted, and only the "tacit judgment" of seasoned experts (such as the contract risk points a senior lawyer can spot at a glance) can truly teach AI to do the work.
Currently, Mercor's customers range from foundation model labs to vertical-sector giants such as legal AI company Harvey and finance software firm Ramp.
As Jason lamented earlier on X, writing a high-quality question for a top benchmark today may already carry as much influence as publishing a paper in Nature or Science.
As more and more enterprises begin training their own models, companies like Mercor are trying to define the next frontier of AI evolution by controlling "high-quality data and evaluation standards."
One more thing
Then again, with Jason Wei's departure, "work at Meta for at most a year" seems to have become a curse.
Earlier, former Apple foundation models head Ruoming Pang gave up a pay package worth more than $200 million and, after only seven months on the job, decisively jumped ship to OpenAI;
On this, many voices in the industry joke that big tech's sky-high poaching deals usually come with a "1-year cliff."
The first large signing bonus and stock often unlock and vest all at once at the one-year mark. Once that time node passes, it's as if a "player option" activates automatically, and whether the person stays depends entirely on the lab's real fighting strength and technical vision.
Meta's chief AI officer Alexandr Wang, born in 1997, remains steadfast and still highly valued by Zuckerberg.
Solid.
AI Trends