AI Trends

The father of chain-of-thought Jason Wei leaves Meta to join the $20 billion-valued Mercor as Chief Scientist

36Kr · AIUpdated Nov 18 17:41

AI summary

On November 18, Mercor co-founder and CEO Brendan Foody announced on the X platform that Jason Wei, who proposed the chain-of-thought prompting method, had joined Mercor as Chief Scientist, and Jason Wei subsequently confirmed the news. Jason Wei previously worked at Google Brain, OpenAI, and Meta Superintelligence Labs, and participated in o1, Deep Research, as well as evaluation benchmarks such as SimpleQA, BrowseComp, and HealthBench. Mercor was founded in 2023 and shifted from AI talent recruitment to providing expert data and evaluation services for model companies; on November 10 it announced the completion of a $750 million Series D funding round at a valuation of $20 billion, and its research team also includes Edward J. Hu, the first author of LoRA.

Why read this

AI score 57

From CoT to evaluation benchmarks such as SimpleQA, tracing this researcher's path toward a data and evaluation company offers a view of shifting AI talent flows.

Topics

Tags

Body · AI translation

Meta Superintelligence Labs has lost another heavyweight researcher.

This time it is Jason Wei, known as the "father of chain of thought."

On November 18, Brendan Foody, co-founder and CEO of AI startup Mercor, announced on X that Jason Wei has officially joined Mercor as Chief Scientist.

Jason Wei subsequently confirmed the news. He said that after more than a wonderful year at Meta TBD Lab, he was delighted to join Mercor. He also specifically noted that as more and more companies begin training AI models, data will shape the next frontier of AI development.

It is worth noting that Jason Wei's new employer is not Google DeepMind, OpenAI, or another frontier large-model lab, but Mercor, a startup that began with AI talent recruitment and expert data services.

Just a week ago, Mercor had just announced the completion of a $750 million Series D funding round, reaching a valuation of $20 billion.

From Google Brain to OpenAI, and then to Meta Superintelligence Labs, Jason Wei has now chosen Mercor. Why would this important researcher in the field of large-model reasoning place his next stop at an AI data company?

From Google Brain to OpenAI and Meta, the research path of the father of chain of thought

In the field of large-model research, Jason Wei is a name that cannot be avoided.

Jason Wei has worked at Google Brain, OpenAI, and Meta, with research focuses covering large language model reasoning, agents, and model evaluation.

What made him widely known was the 2022 paper "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models."

As the first author of this paper, Jason Wei, together with researchers at Google Brain, proposed the Chain-of-Thought Prompting (CoT) method.

At the time, large language models could already complete many natural language tasks, but they often performed poorly on math problems requiring multi-step reasoning.

The research team found that rather than having the model give an answer directly, it was better to include examples containing intermediate reasoning steps in the prompt, guiding the model to complete the calculation step by step.

For example, for a math word problem, a normal prompt might only ask the model to output the final number, while chain-of-thought prompting would show how to analyze the conditions, list the calculation steps, and then arrive at the answer.

This method does not require modifying model parameters; simply changing the prompt examples provided to the model can significantly improve the multi-step reasoning performance of sufficiently large language models.

Since then, CoT has become one of the widely influential methods in large-model reasoning research, and studies on chain-of-thought self-consistency and reasoning path search have also emerged.

In 2023, Jason Wei joined OpenAI, participating in projects such as OpenAI o1 and Deep Research, and also participated in the development of multiple model evaluation benchmarks.

SimpleQA, released in 2024, is mainly used to measure the accuracy of large language models in answering factual questions. It contains 4,326 questions with clear answers, used to examine whether models can answer accurately and whether they know that they cannot answer a certain question.

BrowseComp, released in 2025, mainly evaluates the complex web search capabilities of AI agents. Compared with ordinary factual question answering, questions in BrowseComp often require multiple rounds of search, information filtering, and cross-verification to find the answer.

HealthBench, meanwhile, is aimed at medical scenarios, using detailed scoring criteria developed by doctors to evaluate model performance in real medical conversations across multiple dimensions such as accuracy, communication quality, and instruction following.

From CoT to SimpleQA, BrowseComp, and HealthBench, Jason Wei's research experience spans the improvement of large-model reasoning capabilities and how to evaluate these capabilities.

In July 2025, he moved from OpenAI to Meta's newly formed Superintelligence Labs, becoming one of the well-known AI researchers Meta recruited from competitors at the time.

Now, after more than a year at Meta, he has once again chosen a new direction.

Having just raised $750 million, what is Mercor?

Compared with the several AI giants where Jason Wei previously worked, Mercor's name may not yet be so widely known. But in the AI training data and model evaluation market, this company has grown into a unicorn valued at $20 billion.

Mercor was founded in 2023 by Brendan Foody, Adarsh Hiremath, and Surya Midha.

The company initially entered the market through an AI-driven talent recruitment business, helping companies find and match suitable talent. As generative AI developed rapidly, Mercor gradually shifted its business focus toward providing specialized talent, training data, and evaluation services to frontier model companies.

For example, if an AI lab wants to improve a model's financial analysis capabilities, general internet text alone may not be enough; it also needs professionals who understand financial workflows to help design real tasks, evaluate model outputs, and provide high-quality training samples.

By organizing professionals across different fields, Mercor provides data and feedback for AI model training and evaluation. Currently, its expert network covers more than 5 million people, and it has become an important service provider for frontier model labs.

On November 10, Mercor announced the completion of a $750 million Series D funding round at a valuation of $20 billion, with participating investors including Benchmark, General Catalyst, as well as Nvidia, Salesforce, and others.

However, Mercor's goals are no longer limited to providing expert data to AI labs.

At the same time, Mercor is also expanding capabilities originally aimed at frontier labs to enterprise customers.

Different enterprises have their own business systems, internal rules, and quality standards. Even if a general-purpose model achieves good results on public benchmarks, it may not be able to directly complete work for a specific enterprise.

Therefore, enterprises need to build evaluation systems tailored to their own tasks, and then use expert feedback and specialized data to improve model performance.

For Mercor, this means its business is extending from organizing experts to participate in model training further toward helping customers evaluate, improve, and adapt AI models.

And Jason Wei's joining happens to coincide with a stage in which the company is increasing its research investment.

The father of chain-of-thought and the author of LoRA have converged at Mercor

In fact, before Jason Wei joined, Mercor had already begun building its own AI research team.

On October 8, Mercor published an article titled "Why Mercor Is Building a Research Team," formally introducing the company's research plans.

Its core approach is: first, use evaluations to identify real work tasks that models cannot complete; then use domain experts to understand the reasons for failure and build training data; improve model capabilities through post-training; and finally re-evaluate the improvement results.

To this end, Mercor is researching issues such as evaluation of real work tasks, training data quality, model post-training, and how human professional expertise can drive improvements in AI capabilities.

In terms of real work evaluation, Mercor has already launched the APEX series of benchmarks, covering professional fields such as management consulting, investment banking, corporate law, accounting, and software engineering. These evaluations differ from traditional knowledge Q&A; the focus is on testing whether AI can complete tasks with real work requirements.

In terms of model training, Mercor has also published specific experimental results. In post-training research, Mercor conducted experiments using the Qwen3.5-397B model, raising its Pass@1 score on the APEX-Agents benchmark from 16% to 27%.

More interestingly, Mercor has already gathered a group of researchers with representative research achievements in the field of large models.

Among them is Edward J. Hu.

Edward J. Hu is the first author of the famous paper "LoRA: Low-Rank Adaptation of Large Language Models" and previously worked at OpenAI.

LoRA performs parameter-efficient fine-tuning of large models through low-rank matrices, greatly reducing the number of parameters that need to be trained when adapting a model, and has become one of the widely adopted techniques in the field of large-model fine-tuning.

In addition to Edward J. Hu, Mercor's research team also includes Minyang Tian, first author of SciCode and CritPt, Victor Barres, first author of tau²-bench, and other researchers from institutions such as Anthropic and Ai2.

Now, with Jason Wei joining, this company that started with AI talent and data services has welcomed another researcher with significant influence in the fields of reasoning and model evaluation.

It is worth mentioning that Mercor founder Brendan Foody revealed in the official announcement that Jason Wei has not only been a Mercor customer over the past few years but also one of the company's important exchange partners.

In the official announcement, Brendan Foody stated that Jason Wei will help Mercor build its research team, with a focus on advancing research in data and evaluation.

This direction has considerable overlap with Jason Wei's previous work at OpenAI. He participated in developing evaluation benchmarks such as SimpleQA, BrowseComp, and HealthBench, while Mercor is expanding model evaluation to real work scenarios such as finance, law, and software engineering, and attempting to use evaluation results to improve models.

In his personal announcement, Jason Wei also talked about the importance of data. He believes that as more and more companies begin training AI models, data will shape the next frontier of AI development.

From chain-of-thought to model evaluation, and now to real work tasks and data research, Jason Wei's next stop is worth continuing to watch.

Reference links:

This article comes from the WeChat public account "Machine Heart" (ID: almosthuman2014), editor: Neo, published by 36Kr with authorization.

Source

36Kr · AI