# Meta's $100M Poach List Is a Sourcing Map. 7 of 11 Point to Tsinghua.

_Meta Superintelligence Labs published its hires. 7 of 11 trace to Chinese universities. Here is how to build a frontier AI talent pool from that map._

- Canonical URL: https://www.refolk.ai/blog/meta-superintelligence-labs-tsinghua-sourcing-map
- Published: 2026-09-23
- Reading time: 9 min
- Keywords: Meta Superintelligence Labs hires, sourcing AI researchers, frontier AI talent pool, Tsinghua Yao Class alumni, FAIR layoffs 2025

Meta Superintelligence Labs published a hiring list. Seven of the eleven named technical hires graduated from Tsinghua, Peking University, Zhejiang, or the University of Science and Technology of China. If you are a sourcer, that is not a headline. That is a map.

Stop trying to counter-bid Zuckerberg's $100M signing bonuses. Work backward from his target list instead.

## What Meta's public list actually reveals

MSL's roster is the cleanest frontier AI target set ever leaked, because Meta leaked it themselves. Eleven named technical hires (excluding Alexandr Wang and Nat Friedman), seven from Chinese universities, four from Tsinghua alone.

The named hires from Chinese universities:

- **Bi Shuchao**
- **Chang Huiwen** - Tsinghua Yao Class, Princeton PhD, Google, then OpenAI where she worked on GPT-4o imaging.
- **Lin Ji** - Tsinghua 2018, MIT PhD 2023, internships at Google, Adobe and Nvidia, then OpenAI on multimodal reasoning and synthetic data.
- **Ren Hongyu**
- **Sun Pei**
- **Yu Jiahui**
- **Zhao Shengjia** - MSL chief scientist, Stanford PhD, co-author on the original ChatGPT paper and a key researcher on OpenAI's o1.

The pattern repeats: Chinese undergrad at a top-4 school, US PhD at MIT, Stanford, Princeton or Berkeley, one lab tour at Google or OpenAI, then Meta. If you index on those universities and those programs, you cover most of the frontier pool. If you filter LinkedIn by "AI researcher" and boolean the rest, you do not.

> **7 of 11** MSL technical hires from Chinese universities
>
> Four from Tsinghua alone. The pipeline is narrower than the comp war suggests.

## The two-school concentration nobody is pricing in

Two universities produce most of the researchers Meta is paying nine figures to hire. Tsinghua and Peking University sit on top of China's Gaokao funnel, which concentrates the country's top math and CS scores into a handful of programs.

Tsinghua's **Yao Class**, run by Turing laureate Andrew Yao, admits roughly 30 to 50 undergraduates per year. Chang Huiwen came out of it. So did a disproportionate share of frontier lab researchers now working in the US. The class is finite, named, and searchable. You can enumerate it.

Here is what that looks like in Refolk's index of professional profiles:

| Slice | Count | Source |
|---|---|---|
| US ML professionals mentioning "Tsinghua" in profile | 122 | Refolk's index |
| US ML professionals mentioning "Peking University" | 32 | Refolk's index |
| Ratio Tsinghua : Peking in the US ML diaspora | 3.8x | Derived |
| Top employer of US "Tsinghua" ML alumni | Google (3) | Refolk's index |
| Top employer of US "Peking University" ML alumni (ex-academia) | TikTok (3) | Refolk's index |
| Top metro for US "Tsinghua" ML alumni | Bellevue, WA (4) | Refolk's index |
| Named MSL hires from Chinese universities | 7 of 11 | SCMP |
| MSL hires from Tsinghua specifically | 4 of 11 | Cryptopolitan / SCMP |
| MSL post-launch defections in first ~2 months | 8 | AI Tool Report |
| MSL offers made to OpenAI researchers | 45+ | Spyglass / NYT |

The 3.8x gap between Tsinghua and Peking in the US ML diaspora is not random. Tsinghua's CS department is larger, its US grad-program placement is denser, and the Yao Class specifically funnels into Princeton and MIT. If you have one afternoon to build a frontier AI talent pool, spend it on those 122 profiles.

## Why the Bellevue cluster matters more than the Bay Area

The largest US metro cluster of Tsinghua-linked ML professionals in Refolk's index is Bellevue, Washington, with 4 people, ahead of any single Bay Area city individually. Bay-Area-only sourcing misses this.

The mechanism is boring and obvious once you see it. Microsoft AI is in Redmond. Meta has a large AI office in the Bellevue and Redmond corridor. Google Kirkland sits across the lake. Chinese-origin ML PhDs land in the Seattle area and do not move south for a small comp bump. If your saved search has "San Francisco Bay Area" as its geo filter, you are systematically undercounting the exact people MSL is paying to hire.

This is where plain-English sourcing beats boolean. Describing "Tsinghua-affiliated ML researcher in the greater Seattle area with FAANG or frontier lab experience" is the kind of query [Refolk](/) is built to resolve in one pass across GitHub, LinkedIn, and the open web, rather than four saved searches you forget to run.

## The public paper is the target list

The best-vetted target list for frontier AI hiring is not on LinkedIn. It is the author line of the model papers. Shengjia Zhao was one of 20-plus "foundational researchers" listed on OpenAI's o1 paper. He is now MSL chief scientist. The rest of that author list is a free, better-vetted target set than any recruiter database.

The same is true for:

1. The **GPT-4o** system card author list.
2. The **Gemini** technical report contributors.
3. The **Claude 3** model card acknowledgements.
4. The **Llama 3 and Llama 4** author lists (many of whom are now on the FAIR layoff market, more on that below).

These lists have a property that no ATS pipeline shares: everyone on them has already shipped a frontier model. The signal-to-noise ratio is close to one. The work is name resolution, not filtering, and that is a much easier problem than "find me an AI researcher".

> **Example search:** "Find US-based ML researchers who did undergrad at Tsinghua or Peking University, PhD at a top-10 US CS program, and have shipped at OpenAI, Anthropic, Google DeepMind, or Meta FAIR." - [run it on Refolk](https://www.refolk.ai/s/pd45kx3syy)
>
> You get a ranked shortlist with paper authorship, current employer, and metro, not a keyword-matched dump.

## The retention gap is the real signal

Meta is trailing rival labs with a **64% retention rate**, per SignalFire's 2025 State of Talent Report, and eight MSL researchers quit inside two months of launch. At least two of them, Avi Verma and Ethan Knight, went back to OpenAI. That tells you two things.

First, the headline packages are back-weighted. Fortune reports Matt Deitke's $250M package puts up to $100M in reach in year one, but the rest sits behind multi-year vests and performance gates. Which means a researcher who took MSL money in mid-2025 hits a resentment window around mid-2026 to late-2026, right when the pay-gap conversations with old colleagues at OpenAI start biting.

Second, the *marginal* researcher available for hire on the open market has already been rejected once. OpenAI countered by pushing top researchers past $10M annually, with $2M-plus retention bonuses and $20M-plus equity packages. The ones who left were, in OpenAI's own model, replaceable. That does not make them bad hires. It does mean the contrarian move is to source from stayers via warm intros, not from the visible switchers via LinkedIn InMail.

> The marginal researcher available on the open market has already been rejected once. Price accordingly.

## The FAIR layoff cohort is your near-term pool

Roughly 600 people were cut from FAIR and Meta's AI infrastructure units in October 2025 after Llama 4's lukewarm reception and the dissolution of the AGI Foundations team. That is the largest single addressable pool of ex-Meta AI researchers on the market right now, and it is time-boxed.

The right move is not to wait for them to update LinkedIn. It is to:

- Pull the Llama 3 and Llama 4 paper author lists.
- Cross-reference with public Meta org charts from mid-2025.
- Filter for people whose last commit or last public post predates the October cuts.
- Reach out with a specific role, not a "coffee chat".

The same logic applies to the 20-plus foundational researchers on the o1 paper who did not follow Shengjia Zhao to MSL. Some are still at OpenAI. Some are at Thinking Machines Lab. Some are quietly at a stealth. The paper tells you they exist. The work is figuring out where they are now.

## The archetype hire: Trapit Bansal, and what to source next

Trapit Bansal is the archetype. At OpenAI since 2022, foundational contributor to o1, a key player on reinforcement learning alongside Ilya Sutskever. He jumped to MSL. He is what an eight-figure reasoning-model IC looks like: not a manager, not a research director, an individual contributor with a paper trail on the specific capability the buyer wants next.

If you are building a frontier AI talent pool in late 2026, the capabilities and the signals to source on are:

| Capability | Sourcing signal | Public artifact |
|---|---|---|
| Reasoning / RL | o1 author list, Bansal-adjacent ICs | Papers, model cards |
| Multimodal | GPT-4o, Gemini authors | System cards |
| Post-training | Llama 3/4 alignment leads | FAIR org, papers |
| Synthetic data | Lin Ji and co-authors | Paper acknowledgements |

Most of these people are not open to recruiters. They are open to a specific technical question from a specific person. Your job is not to convince them. Your job is to find them accurately, once, and hand the intro to the founder or research lead who can actually have that conversation.

## What to stop doing

Three habits are burning your time. Stop them.

1. **Stop counter-bidding on comp you cannot match.** If your best offer is $600k TC, you are not in the $100M market. You are in the market for people who have decided they do not want to work at MSL, and that is a different sourcing problem.
2. **Stop keyword-searching "AI researcher" on LinkedIn.** The frontier pool is small enough to enumerate. Every one of them is on a paper. Source the paper.
3. **Stop treating the diaspora as a monolith.** Tsinghua's US ML alumni cluster in Bellevue. Peking's cluster at TikTok. A geo-agnostic query returns geo-agnostic noise.

The gap between "I know this pool exists" and "here are 40 named people with current employers and papers" is where most sourcing work dies. It is also the exact gap Refolk closes: you describe the person in plain English, across GitHub, LinkedIn, and the open web, and you get a ranked shortlist back.

## FAQ

### How many people are actually in the frontier AI talent pool?

Small enough to enumerate. Refolk's index shows 122 US-based ML professionals mentioning Tsinghua and 32 mentioning Peking University, which are proxies for the two dominant undergrad pipelines. Zuckerberg reportedly made more than 45 offers to researchers at OpenAI alone, and the total poached across labs is above 50. The addressable, currently-employed-at-a-frontier-lab pool is in the low hundreds, not the thousands. That is why $100M offers exist.

### Is targeting by university legal in the US?

Sourcing on educational background is a normal recruiting signal and is not itself a protected characteristic under US employment law. What matters is that you evaluate candidates on the job-relevant criteria (papers shipped, systems built, roles held) rather than proxy-discriminating on national origin. Using "Tsinghua alumni" as a pipeline signal is the same category of move as "Stanford CS PhD" or "ex-Palantir engineer". Confirm with your own counsel for your jurisdiction and role.

### Should I bother sourcing OpenAI stayers if MSL keeps getting rejected?

Yes, but change the channel. The people OpenAI is retaining with $10M-plus packages are not answering cold InMail. They will answer a specific technical question from a founder they respect, or a warm intro from a former colleague. Your job as a sourcer is to build the intro map (who worked with whom, on what paper, in what year), not to send the outreach yourself.

### What about the FAIR layoffs? Is that pool worth working?

It is the highest-yield near-term pool in AI, and the window is short. Roughly 600 people were cut from FAIR and AI infrastructure in October 2025. Many are on H-1B clocks. Pull the Llama 3 and Llama 4 author lists, cross-reference with LinkedIn "past company: Meta" filters, and reach out with a real role. In six months this pool is placed and gone.

---

_From the Refolk blog. More posts: https://www.refolk.ai/blog/index.md_