The Price of Intelligence: Why AI Researchers Can Cost More Than NBA Stars

The Price of Intelligence: Why AI Researchers Can Cost More Than NBA Stars

Public work · eighteen months · a hundred people · the infrastructure behind a talent map

“We have entered an era of intellectual super-monopoly: the most intelligent token can be worth more than money.” — Sam Gao, on why a small group of AI researchers has become so expensive

Guest: Sam Gao (Gao Daiheng)｜Founder of DINQ; formerly an algorithm researcher at Alibaba’s DAMO Academy and Tongyi Lab

Key takeaways

• The best AI researchers are not being priced as interchangeable employees. Companies are buying an option on the next discovery they might make, during a very short technological window.
• A prestigious degree or a stack of papers is a weaker signal than it used to be. Public work, experience inside a real training system, and the ability to move into a new problem quickly matter more.
• The shortage is not simply “people who understand AI.” It is people who have worked at the scarce intersection of post-training, data, inference infrastructure, and large-scale compute.
• Money and GPUs do not automatically produce research. Benchmark-maximizing, rigid incentives, and overlong vesting can all push away the people a lab is trying to retain.
• Finding these people is becoming an infrastructure problem: connecting fragmented public traces, specialized databases, and natural-language search—without pretending that a mysterious ranking can replace judgment.

In 2025, public reporting around the AI talent market produced figures that sounded more like sports contracts than research compensation. One widely discussed offer package was said to approach $200 million. If it were spread over four years, the annualized number would sit in the same universe as elite professional athletes. Yet the comparison breaks down at the most important point. When a star changes teams, the old club can receive a transfer fee. When an AI researcher leaves, the former employer receives nothing—and some of the most valuable knowledge leaves inside the researcher’s head.

Sam Gao has spent years studying that gap between visible companies and the people who make the technical work happen. He began as a civil-engineering graduate student with no conventional AI pedigree, moved into research through open-source contributions, later worked at Alibaba’s DAMO Academy and Tongyi Lab, and now builds DINQ, a company that treats people search as an AI-native infrastructure layer.

Raymond Huang, investor and host of Mossfire, came to the conversation with a blunt question: if a small group of researchers can be priced at nine figures, what exactly did they do, what makes their experience scarce, and can anyone map the next group before money makes them famous? Their discussion moves from open-source pull requests to the limits of managing research by KPI; from the recruiting tactics of companies such as Cursor to the problem of turning scattered public traces into a usable talent map.

The following conversation has been edited and condensed.

1. “A public work is harder to fake than a résumé.”

Raymond: You started as a civil-engineering graduate student rather than a computer-science researcher. How did that turn into offers from companies such as Meta, Nvidia, and Alibaba?

Sam Gao: In late 2015 and early 2016, I was in the first year of a civil-engineering master’s program. I felt that property prices and the broader market were becoming overheated, but I was already on that track. So the practical question was: what else am I good at?

AI was starting to send signals. The first generation of China’s computer-vision startups was appearing, and I began with Andrew Ng’s courses online. I had no AI background and almost no resources. GPUs were expensive, but consumer cards such as the GTX 1080 were still available. Few people were using them seriously for research or applications.

The route I chose was what people now call building in public. I wrote on CSDN, a Chinese developer community, and submitted pull requests to open-source projects. My first PyTorch pull requests were literally typo fixes. I did not yet understand much of what the framework was doing. But when a maintainer merged a contribution, a civil-engineering student’s code was suddenly running on other people’s computers. That feedback loop mattered. First you fix a typo; then you want to understand enough to take on a harder problem.

Over time, the work itself became the proof. A public artifact is hard to fake. It lets someone evaluating you see a real capability without spending as much effort on credentials, and it reveals something else: whether you initiate work before somebody assigns it. The offers I later received were driven by that work, not by someone giving me a reference.

Raymond: Today, people talk about learning in public all the time. At the time, did friends, teachers, or classmates see this as a strange detour from the normal path?

Sam Gao: Not really, because most people did not know what it was. The conventional route was clear: get a strong degree, join a strong professor, then try to enter a lab such as SenseTime, Megvii, OpenAI, or DeepMind. What I was doing was not mainstream, but it was not controversial either. A path often looks deliberate only in retrospect. When you are inside it, you are usually just doing the next small thing that is available.

Raymond: You left DAMO Academy in April 2024, after the generative-AI wave had begun. What changed your view of the work?

Sam Gao: In 2023, my team worked on image projects: human motion and virtual try-on. A merchant might have a flat image of clothing; the idea was to let a model put it onto a person without arranging a new photo shoot. At first we approached the work in the classic research way—spending time on algorithms, ordinary differential equations, stochastic partial differential equations, and different ways of improving the diffusion process.

I used to describe diffusion models as “calculus for the AI era.” A traditional model feels like a deterministic mapping: put in A, apply a function, get B. Diffusion has a sequence of time steps, an iterative path toward a result. At the time, I had a simple intuition that images and video, as higher-dimensional media, might eventually dominate language. That has not been proven. Diffusion-based language models have made progress, but they have not simply displaced autoregressive language models.

What did become clear on that project was more mundane and more important: clean, high-resolution, well-labeled data changed the outcome more than many of our algorithmic adjustments. We collected better images of models against simple backgrounds, and the project gained thousands of GitHub stars and ranked highly on Hugging Face Spaces.

At roughly the same time, GitHub Copilot was becoming useful for routine scripting. Tasks that once took an hour or two could be done faster by the tool. If the decisive advantage in a project is excellent data, and an agent can take over more of the implementation, you begin to ask where your own value is moving. I was not saying algorithms had become worthless. I was saying that some work previously treated as a moat was being commoditized faster than people expected.

2. “The people who make the work are still less visible than the companies.”

Raymond: That is why you ended up printing posters and waiting outside OpenAI’s office?

Sam Gao: The poster episode was in early 2025, but I had been writing about the authors behind important AI techniques since 2019 or 2020. I wrote about people behind work such as StyleGAN, chain-of-thought prompting, and the Adam optimizer. The products were reaching ordinary users, while many of the people who made the work possible remained distant from public view.

Durk Kingma was one of the inspirations. He was an OpenAI technical co-founder and a first author of the Adam optimizer paper. His work was extraordinarily influential inside machine learning, but most people outside the field would not recognize his name. After ChatGPT, attention naturally concentrated on public figures such as Sam Altman and Ilya Sutskever. I wanted to know who had actually done the technical work, what their paths looked like before anyone treated them as inevitable, and how their work reached the world.

The Renaissance analogy was deliberately imperfect. I was not claiming today’s AI researchers are Michelangelo or Raphael. I meant that when a work is newly made, its maker can look like one more craftsperson in a workshop. Five hundred years later, people remember the work differently. The investor who funded a church may matter to historians, but the artist’s work can travel farther through time.

Raymond: That is a generous way to look at it. I am a more utilitarian reader. Tell me one hundred people have good stories and I may not read them. Tell me that some of them are worth hundreds of millions of dollars, and suddenly I want to know what they did.

Sam Gao: That is fair. Money is often the first language that makes a broad audience ask a technical question. But the underlying point is not that every individual is a lone genius. Early breakthroughs could sometimes come from a person working in a narrow corner. Frontier-model development is now much more like coordinated military-scale work: research, data, systems engineering, product, and compute are tightly coupled.

That makes it more important, not less, to distinguish what people have actually done. Alec Radford, a first author on GPT-1 and GPT-2, came from Olin College of Engineering, a small undergraduate institution near Boston. Patrick Lewis, a first author on retrieval-augmented generation, took a path through time in Vietnam, a startup later acquired by Meta, and doctoral work at University College London. These are not neat prestige narratives. They are reminders that a research trajectory can be visible in the work before a credential tells a clean story.

Raymond: So the point is not to turn technical biographies into celebrity gossip. It is to look at the trail left by work before a market has attached a price to it.

Sam Gao: Exactly. Today, the scarce asset is not a job title. It is the capacity to make something new for the first time. If we only look at chief executives, funding rounds, and company logos, we will miss where the next technical map begins.

3. “A company is buying an option on the researcher’s next discovery.”

Raymond: Why did the market for researchers suddenly start to resemble a transfer market?

Sam Gao: The visible escalation accelerated in 2025. Meta was under pressure in frontier-model competition, and it began assembling a more aggressive research effort. In sports, moving a player from one club to another usually creates a payment to the old club. In AI, the relevant verb is poaching. A researcher moves, the old employer receives no transfer fee, and some of the company’s most important intellectual context may move with that person.

The normal engineering ladder is relatively legible: perhaps a few years of experience, then a progression through job levels and compensation bands. Frontier researchers can follow a different curve. They may hold the same title—member of technical staff—while their economic value is being assessed in a much less standardized way. Before this cycle, an experienced researcher might have a package in the low millions, often with a large equity component whose eventual value was uncertain. A publicly traded company can offer a clearer price, and in a hot moment that price can jump by an order of magnitude.

The eye-catching nine-figure numbers should be handled carefully. The conversation referenced public reporting and market discussion, not employment contracts. The useful point is not whether every rumored package has the same structure. It is that only a small number of people are being discussed at that level, and their common feature is not simply education. They have recognizable work that other experts can connect to a real capability.

Raymond: The most famous example is a senior researcher who was reportedly offered an extraordinary package, spent only months at Meta, and then returned to OpenAI. If compensation can reach that level and still not retain someone, what is the company actually buying?

Sam Gao: It is not buying obedient labor. It is buying an option on where the next important discovery may happen.

For a very senior researcher, the difference between one already enormous number and another may be less important than the next two or three years. The research window feels short. Scope, collaborators, the stability of the organization, and the freedom to make a serious attempt can matter more than another increment of money. A lab that is constantly reorganizing may be less attractive than one that offers a clearer environment, even if the headline compensation is lower.

That is why the sports analogy is useful but incomplete. A researcher is not a player who can be placed into a known system and expected to reproduce last season’s output. The thing a lab is trying to buy is a possibility: that this person, with this group and these resources, will help create the next thing that nobody knows how to specify in advance.

4. “The most expensive skill is often a recent scar.”

Raymond: Which skills are gaining value, and which are losing it? Video, robotics, and multimodality all look hot from the outside, yet you argue that some multimodal research talent is becoming less scarce.

Sam Gao: It is important to separate product demand from the marginal value of one more algorithm researcher. Video may have strong product pull. That does not automatically mean another person working on the algorithm layer is equally scarce. In many multimodal settings, technical routes are converging and the bottleneck shifts toward data: how it is acquired, labeled, filtered, and turned into a reliable training pipeline.

The same logic weakens old proxies for talent. In 2020 or 2021, several first-author top-conference papers could be a powerful hiring signal. Today, producing something that looks like a paper can be much faster, and the count alone is less informative. That does not make papers irrelevant. It means that the market needs evidence closer to the actual scarce work.

Two areas Sam highlights are post-training and AI infrastructure. One is on-policy distillation, or OPD: not merely copying answers from a stronger model, but using the student model’s own trajectory, sampling, rewards, and data to improve behavior. The other is training and inference infrastructure: making a fixed set of machines emit more useful tokens per second. Moving throughput from 1.0x to 1.2x or 1.5x can be economically meaningful when high-quality tokens and compute remain constrained.

Raymond: This is why a 23-year-old intern can become unusually valuable?

Sam Gao: Potentially, yes. Suppose someone spent the last year on a problem that barely existed eighteen months ago—say, a particular post-training setup or large-scale inference bottleneck. That person may have encountered concrete failures that a senior researcher who has never touched the system has not encountered. It does not mean the intern is superior at everything. It means their recent experience can be scarce.

AI talent has a short half-life, much like models do. Someone who was excellent at training GANs may find that the specific skill has little market value today. Moving from search, ads, or recommendation systems into frontier models is not simply learning a new library. In the old world, improving click-through rate by a few points translated relatively directly into revenue. A benchmark improvement does not necessarily produce a better user experience. The feedback is noisier, and the problem changes while you are learning it.

Raymond: Why can’t every well-funded company simply run more experiments until it catches up?

Sam Gao: Because the bottleneck is not merely willingness. Large-scale work consumes scarce resources. You need compute, and the hardware is often a mixed fleet rather than a perfect cluster. You need carefully designed reward rubrics for tasks that do not have a natural reward signal. You need to pay for sampling, labeling, filtering, and failed runs. Even if two people could attempt the same project, a company may not have enough capacity to let both teams burn through thousands of modern accelerators and see who wins.

That is also why distillation becomes attractive. Many teams know they will trail the best labs. Instead of reproducing an entire reinforcement-learning stack and its reward-design burden, they try to learn from high-quality outputs. The work rapidly concentrates among the people who have actually done it under real constraints.

5. “Research cannot be managed as a KPI factory.”

Raymond: A manager might hear all this and conclude: give researchers more money, more GPUs, harder targets, and better dashboards. Why is that not enough?

Sam Gao: Because research is not a production line with a known endpoint. You can raise a benchmark score without making a model more useful in the tasks people care about. If a system rewards only what is easy to measure, people can optimize toward the measurement rather than the underlying problem. In machine learning, that becomes a form of reward hacking.

Resources matter enormously. But important work also needs room for an attempt that does not yet have a business case or a stable metric. If the answer were simply to make everyone work longer and hit every benchmark, the strongest manager would always win. That is plainly not how discovery works. People leave, or they learn to produce the work that looks best in a review cycle while avoiding the work that might actually move the field.

This creates uncomfortable internal questions about fairness. A lab may pay newly recruited people more than long-tenured employees. It may use stock-based compensation because valuation grows faster than revenue. Some organizations can tolerate this because colleagues understand who can independently own a difficult part of training or a release. But compensation is only one part of retention. A decade-long vesting schedule can feel absurd when a researcher’s current specialty may be obsolete in eighteen months. Restricting someone from publishing, writing a technical blog, or sharing an open-source tool may remove another reason they create in the first place.

Raymond: Then is the market a bubble?

Sam Gao: I would not reduce it to that. There is still a long distance between current AI talent spending and the amount of white-collar work models might eventually affect. But the path matters. Models have to become genuinely useful for more work, and society has to accept the transition. If those things do not happen, the economics will be repriced. If they do, today’s packages may look less irrational in retrospect.

The deeper point is that compute and data remain the largest direct costs. Yet many companies can buy hardware and acquire data. The variance in outcomes comes from people and organizations: who frames the experiment, notices the failure, selects the data, and knows when the benchmark is lying. Frontier models are an infinite game. No company can declare a final version and assume everyone else will stop.

6. “Do not wait for the market to hand you the people who remain.”

Raymond: What does a better recruiting process look like when the relevant group is so small?

Sam Gao: Cursor offers a useful example. The company reportedly identifies perhaps fifty to one hundred people it wants to understand over time, rather than opening an enormous funnel and waiting for applications. The well-known story about sending a candidate a guitar is not really about the gift. It signals that the company has paid attention to a person rather than treating them as another résumé in a queue.

This is high-touch work. The question is not only whether someone has one narrow skill today. Technology changes too quickly for that. You want people who can change shape when the problem changes. A work trial can reveal more than a sequence of interviews or a paper list: can this person work through an ambiguous problem with the team?

Raymond: It is the opposite of what you call the “desperation funnel.” In investing, you can look at the deals that reach you after every other round and mistake that residue for a real selection set. Recruiting seems similar.

Sam Gao: Yes. The alternative is to make a judgment early, accept that it may be wrong, and invest in a relationship before the market has fully repriced the person. That takes resources, and not every company can operate this way. But for genuinely scarce people, waiting for inbound applications is often just a way of choosing from whoever was left over.

The same logic applies after hiring. Long-term equity can help, but it cannot substitute for meaningful work, trusted collaborators, and room to leave a mark. An engineer who values public technical exchange may care deeply about whether they can still publish a blog post or contribute a tool. The company is not retaining a salary line. It is trying to retain a person’s willingness to do difficult work there.

7. “A talent map is built from traces, not from a magic list.”

Raymond: If labs are becoming less open, how can anyone identify the people who have actually worked at scale?

Sam Gao: You have to distinguish the kinds of experience first. One kind is physical and operational: data centers, connecting machines, scheduling them, replacing failures. Another is closer to model training and inference infrastructure. Contributors to projects such as Fully Sharded Data Parallel, Megatron, vLLM, SGLang, or TileLang can reveal useful experience. If you have debugged precision alignment, parallel training, or failures in a real large-scale environment, you have scars that do not show up in a classroom exercise.

Companies can restrict papers, but a person rarely goes from having no visible technical history to becoming a core contributor on an important system with no preceding traces at all. Papers, GitHub commits, Hugging Face projects, talks, interviews, former collaborators, and old social posts can be assembled into a trajectory. This is not a LinkedIn résumé and it should not become a claim that someone “built” a confidential model. It is a way to ask better questions.

Personal traces can sometimes provide context too. Sam mentions Ben Mildenhall, a co-creator of Neural Radiance Fields, whose public presence included piano performances. That does not prove technical ability. It is simply a reminder that people are not keywords. Their public work can show how they allocate attention and relate to a community. At the frontier, finding a person is often a process of matching many partial signals rather than filtering a perfect database.

8. “The product is not a new LinkedIn page. It is people search as an atomic capability.”

Raymond: This brings us to DINQ. At first I assumed you were building a specialist recruiting firm for the next $200 million researcher. That is too narrow, isn’t it?

Sam Gao: Much too narrow. My shorthand is “OpenRouter for people search.” In China, recruiting teams may need to pay separately for Maimai, a Chinese professional network; BOSS Zhipin, a Chinese hiring marketplace; and Liepin, a Chinese recruitment platform. Elsewhere, they move among LinkedIn, Indeed, JobStreet in Southeast Asia, Xing in Germany, and many other sources. The information is fragmented, the formats differ, and the search logic differs.

DINQ aims to integrate these sources through APIs so that people search can be used inside a product or called by an agent such as Codex, Claude Code, or WorkBuddy. A user should be able to ask for “an Anthropic forward-deployed engineer in Singapore” without first deciding which databases to buy and how to search each one.

Keyword search misses the point. Someone working on world models may never write “world model” on a profile; the public trace may say only that they worked on a project such as Sora. A system that can interpret the relationship between work, technical artifacts, location, and intent can make a better first pass. It does not eliminate human judgment. It reduces the number of disconnected doors a person must open before judgment can begin.

Raymond: That makes the use case broader than research hiring. An HR team mapping a market for a chief executive may need a reference set. A company expanding abroad may need salespeople in Latin America. An investor may want to notice a researcher becoming unusually active before a move or a startup.

Sam Gao: Exactly. The core result is finding the person. Opening LinkedIn, BOSS Zhipin, or Indeed is a process, like choosing a plane or a robotaxi to get from A to B. The user cares about reaching B safely.

That does not mean the system can produce a mystical list of future founders. Investor interest in talent movement is real, but the data are incomplete. Many people do not update LinkedIn. Public activity before a move is a clue, not a conclusion. A useful recommendation system has to reason about role, technical work, time, geography, and team relationships. More data does not mean the answer is already sitting inside the database.

The question that began this conversation therefore returns in a different form. A market may place a dramatic number on an AI researcher, but the number is not the explanation. It is a crude public acknowledgment that a person’s next choice—where they work, what they try, and who they work with—may still change a company or an entire technical direction.

This interview has been translated from Chinese, edited and condensed for clarity.

If you're working on this too — or you think we've got it wrong — write to us at [email protected]; if you'd rather not write, just leave your email below.