154: From Qwen’s Shake-Up to “AI Heroes” | A Conversation with DINQ’s 高岱恒 on Legendary AI Researchers
Summary
Qwen’s personnel shake-up has turned the latent mobility of a 100-plus-person team into a visible target in the global AI talent market. Related searches on DINQ tripled, with roughly 2,000–3,000 queries concentrated on large language models, reinforcement learning and Agent RL. Searchers were mainly HR professionals and headhunters, including Meta’s Executive Search team. When users previously asked which Chinese team drew the most searches, the answer was no longer DeepSeek but Qwen; after the shake-up, the market is more directly betting that “some people may want to look at other opportunities.”
Qwen’s hardest-to-replicate moat is not a single leaderboard result, but an open-source ecosystem built on model sizes, modalities, developer mindshare and distribution. Sam says Qwen’s cumulative downloads on Hugging Face and ModelScope exceed those of several other Chinese open-source model teams combined, with a full family spanning 0.6B and 1B models through image, video, reasoning and embedding models, making it the de facto standard in papers and embodied-AI projects. ModelScope is like “the GitHub of the AI era”: usage, likes and comments can be monetized into hosting, inference and storage revenue, but more importantly, they tell Alibaba what frontier developers are paying attention to.
This shift could accelerate the industry’s move from competing on open-source buzz to competing in the trenches on post-training, RL and Agent tool use. Sam expects Qwen-linked talent to be fought over aggressively by Meta, xAI and OpenAI, while model companies that believe scaling can continue to unlock more intelligence will direct more resources toward real-world task performance rather than simply chasing benchmarks; open-source investment may shrink in relative terms. He is particularly bullish on third-party data and reinforcement-learning environment vendors, expecting more frontier labs in 2026 to outsource work they would previously have kept in-house.
Top research organizations are recreating the Renaissance-era relationship between workshops and patrons, with researchers increasingly priced on signature work rather than employer pedigree. Sam’s analogy is that the Medici family once “gave Michelangelo a seat at the table”; today, Zuckerberg’s packages for star researchers can exceed those of NBA or Premier League stars. As compute projects become more expensive, technical judgment that avoids dead ends and improves compute output becomes more valuable. Organizational forms may expand from internal teams at large companies to specialist studios handling a single link such as Agent RL, and eventually to enterprise-facing “capillaries” like Cohere.
AI talent supply is expanding rapidly, but the shelf life of degrees and company names is shortening; public work is becoming a more reliable screening signal. Sam recalls ICLR submissions rising from roughly 1,000–2,000 in 2020 to more than 30,000 five or six years later. Many key techniques were developed by people under 30, and interns often carry critical work at Microsoft, Qwen, DeepSeek and MiniMax. “The algorithm era rewards results”: big-tech experience with GANs and VAEs yesterday does not guarantee continued relevance in diffusion models or the next Agent paradigm.
Sky-high pay may win researchers over not by inspiring a sense of mission, but by compensating for pressure and buying several years of their attention. DINQ Roast initially set its highest band at annual salaries above $10M, while Meta’s recruiting bids pushed the market’s imagination to $100M; Manqi also mentioned hearing of people with $500K salaries but $20M in options vesting over 4 years, which could make monthly equity income exceed their wages. People who receive such offers often first “freeze,” then choose to “work quietly, make as much money as possible for a few years, then retire,” because a single experiment using 4,000 GPUs carries enormous technical and psychological pressure.
As models move into applications, the talent premium is spreading from the people who build the F1 cars to the drivers who can take models into the real world. Sam observes that research-paper authors’ social posts often draw only a few hundred likes and two or three generic congratulations, while Peter, the creator of 小龙虾, Andrej Karpathy and Boris Cherny, the creator of Claude Code, can see comment-to-like ratios of 1:4 or 1:5. The reason is direct: users may not understand foundational research, but they can immediately feel the benefits of an Agent organizing email, executing workflows or completing tasks automatically.
Recruiting relationships may ultimately shift from matching people to companies to matching people to tasks, with observable AI-native behavior repricing talent. Sam projects that cognitive projects once requiring 6–9 months of repeated client-vendor calibration could be compressed into a few weeks; three moves in five years, organizational loyalty and conventional titles may weaken accordingly. Claude Code token consumption, personal task-automation rates and the number of AI products used could become “one-leaf-reveals-autumn” proxies for ability. His advice is not to chase temporary first place, but to treat the career as an “infinite game”: protect health and emotional stability, and move value toward judgment, communication, coordination and the ability to create new work.
Deep dive
1. Qwen’s shake-up made talent mobility visible first
Following the personnel changes, queries on DINQ about Qwen members and paper authors tripled to roughly 2,000–3,000. Large language models, reinforcement learning and Agent RL were the hottest topics, while image and voice fell noticeably behind.
Searchers were mainly HR professionals and headhunters, with Meta Executive Search also appearing. Manqi asked whether overseas firms were using the opening to poach talent. Sam’s restrained explanation was: “People just think that some of them may want to look at other opportunities.”
Qwen’s core team has only a little over 100 people, but it has already become the primary talent pool surfaced by overseas demand. When users previously searched broadly for a type of capability, the system often returned Qwen and Kimi members; the shake-up merely turned latent demand into targeted searches.
2. Qwen won open-source mindshare with a complete model matrix
Sam assesses Qwen by influence rather than commercial revenue. On Hugging Face and ModelScope, he says, its total model downloads exceed those of several other Chinese open-source teams combined and are well ahead of Mistral, which he used as a benchmark.
The advantage comes from the breadth of the family: 0.6B and 1B models at the small end, alongside image, video, reasoning and embedding models. Schools and developers with limited resources can find a suitable version directly instead of bearing the cost of training one themselves.
From the second half of 2023 onward, more top-conference papers began citing Qwen technical reports or using it as an Agent’s “brain” and a key module in their research stacks. Sam calls this mindshare “the de facto standard”; once established, it makes marketing and switching costs high for later entrants.
3. ModelScope’s real asset is developer intent
When Sam was at Damo Academy, ModelScope was not yet popular, and colleagues were at one point required internally to contribute models and datasets. Its traffic and downloads later continued to grow, which he saw as another indication that China’s and the Chinese-speaking world’s AI developer base was expanding.
He views ModelScope and Hugging Face as “the GitHub of the AI era”: the richer the models, datasets and Spaces applications, the more usage, contributions, comments and likes concentrate there, allowing the platform to identify what frontier practitioners actually care about.
Inference hosting, model deployment and storage for large datasets all generate direct revenue. But Sam believes their strategic value to Alibaba is far greater than any isolated revenue line; the behavioral data also serves as an environmental signal for allocating R&D resources in reverse.
4. Personnel shock will push competition toward post-training and RL
Sam expects Qwen-linked members and talent to be fought over aggressively by Meta, xAI and OpenAI. He says that after Musk fired a large number of early members, he was ravenous for this group; Sam also speculated that the wealth effect of a SpaceX–xAI merger could strengthen the pull on researchers. Talk of a 2026 IPO was his speculation, not a firm view.
The larger industry shift is toward convergence. As long as companies believe continued scaling can produce more intelligence, they will keep doubling down on model capability while shifting the focus toward post-training, RL and Agent tool use rather than merely competing on open-source buzz.
Sam believes benchmarks matter less to most users now. What companies actually care about is whether a model performs better on real tasks, so people with overseas post-training and RL experience will enter domestic teams through full-time and advisory roles.
He expects more third-party data and reinforcement-learning environment vendors to emerge in 2026, supplying frontier labs with work they would previously have tended to keep in-house. “From the second half of 2025 onward,” he observed, some of the most profitable companies in AI have already included vendors selling reinforcement-learning environments.
5. Research organizations are recreating Renaissance workshops
Sam compares the relationship between frontier labs and researchers to workshops, artists and patrons. Apprentices enter a studio not to prove “I spent time here,” but to complete work they can carry with them and use to establish themselves in the world.
The Renaissance raised the social standing of artisans, with the Medici family “giving Michelangelo a seat at the table.” Today, Zuckerberg’s packages for star researchers exceed those of sports superstars. Sam sees the same logic at work: whoever can make expensive compute perform best commands an enormous premium.
Manqi’s counterquestion is worth preserving: giants such as Alibaba, ByteDance, Meta and Google will still tend to keep core technology in-house. Sam does not deny that; he describes an evolution through three stages—internal teams, external specialist studios and independent model-service companies.
Agent RL could become the archetypal assignment for an external workshop: the client supplies an industry scenario and a specialist team handles the training. Manqi cited Cohere and Tinker, launched by Thinking Machines Lab; as Agent orchestration and toolchains mature, small teams could become the industry’s “capillaries.”
6. Employers are way stations; work is the long-term passport
Sam says interns may carry critical work at Microsoft, Qwen, DeepSeek and MiniMax. Years spent as a full-time employee do not automatically translate into an advantage; this talent regime is fundamentally different from traditional engineering.
He has seen paradigms eliminate prior advantages. Someone working on GANs and VAEs in 2021 who failed to embrace diffusion models in 2022–2023 would see their previous big-tech experience lose value quickly. Joining Qwen or DeepSeek can only be an intermediate stage, not the endpoint of a life.
Papers, training frameworks, data processing, infrastructure and open-source projects can all serve as signature work. Frontier institutions may still look at degrees and experience, but the market is increasingly pricing people by what they have made, while the value of past credentials depreciates very quickly.
7. Sky-high pay buys attention—and compensates for pressure
Asked what money means to these people, Sam gave the honest answer: “I honestly don’t know.” Human history has seen very few pure intellectual workers receive compensation on this scale, and many young people who receive huge offers initially do nothing more than “freeze.”
The typical mindset is not immediate consumption, but “get to work quietly, make as much money as possible for a few years, then retire,” and later pursue free research without KPI pressure. High pay therefore functions like a purchase of several years of concentrated attention.
Manqi also mentioned hearing of people with $500K salaries but $20M in options vesting over 4 years, which could make monthly equity income exceed their wages. Sam sees high pay as partly compensation for research pressure: a single experiment may consume 4,000 GPUs, leaving researchers worried about both the model and the scheduling—and afraid that the cluster will fail midway.
8. Signature work creates more direct positive feedback than abstract mission
Manqi guessed that some people want to leave a contribution to scientific history, while others are driven by a mission to achieve a more powerful state of intelligence. Sam thinks such motives may exist in the subconscious, but the conscious explanation is often simpler: “I need some work to prove that I’m capable.”
Stars, feedback and contributions on an open-source project, or seeing one’s name near the top of a Qwen technical report or contributor list, all create the positive feedback of “I can make this happen.” That feedback drives the next job, research project or startup.
This is not a one-time clearance event but an “infinite game.” Some people can keep playing until 100, while others in their 30s exhaust their energy and mental bandwidth and turn to easier pursuits. Prestige and money matter, but work remains the passport into the next round.
9. DINQ Roast first took the temperature of the talent bubble
Sam came up with the idea in March 2025 and built DINQ Roast in April–May. Users entered a Google Scholar, GitHub or LinkedIn profile; the system analyzed the résumé and produced a “roast” plus an estimated market value, hitting the market mood around “What am I actually worth?”
The initial top band was an annual salary above $10M. Other bands included below $1M, $1M–$3M and $3M–$5M; a few months later, Meta offers above $100M made the original, seemingly outrageous ceiling look conservative.
The initial buzz centered on OpenAI figures including Ilya, John Schulman, 欧阳龙 and 翁嘉仪. Later, users became more interested in testing themselves and their colleagues. The PK function compared total citations, first-author citations and top-conference counts, though Sam acknowledged that this was unfair to research with low citations but significant theoretical impact.
林俊旸 and 周畅 each had just over 20,000 citations at the time, and the PK result was “too close to call”; when they saw it, they “lost it.” Professors and star researchers using the product themselves confirmed Sam’s view that talent not only needs to be discovered, but also has a strong need for self-understanding.
10. People who close the loop are priced as craftsmen, not job titles
Sam defines creative talent not by a degree or position, but by the ability to close the loop: identify a problem, formulate it, design an experiment, solve it and ultimately deliver a complete result.
Such people resemble Renaissance artists because they are not completing one isolated step on an assembly line; they are “making a complete work.” A paper has been the most typical form of that work for the past decade, but frameworks, data and foundational tools qualify as well.
The market rewards work directly. Once a model, algorithm or open-source project proves the result, employers can rely less on degrees, years at a company or generic interview questions to infer ability. “The algorithm era rewards results.”
11. Age and provenance are breaking old-school screening
Sam cited the brothers Eric Luman and Troy Luman, who began conducting diffusion-model research around 2020 and consistently posted their work on arXiv rather than at peer-reviewed conferences. OpenAI eventually discovered them.
Chinese companies likewise search GitHub, top-conference papers and open-source projects for clues. “Heroes are judged by what they do, not where they come from” does not mean there is no screening; it means the target of screening is shifting from school names to public results and sustained action.
Around 2020, Jonathan Ho’s DDPM, 宋飏’s related diffusion research and subsequent Transformer advances were gradually falling into place. ICLR received roughly 1,000–2,000 submissions a year at the time; five or six years later, Sam recalls, the number had surpassed 30,000.
Alec Radford’s CLIP, 苏剑林’s rotary positional embeddings, MoE work involving Albert Q. Jiang and Robin Rombach’s team’s Stable Diffusion all reinforce Sam’s age observation: many important results appeared before their authors turned 30.
12. Application builders are becoming the more compelling “drivers”
Sam compares researchers to people who build F1 cars or Gundam, and people who use models to create new products to drivers. As foundation models mature, the latter group’s influence and talent premium are both rising.
He observes that research-paper promotion often draws only a few hundred likes and two or three “congratulations”-style comments, while posts by Peter, the creator of 小龙虾, Andrej Karpathy and Boris Cherny, the creator of Claude Code, can reach comment-to-like ratios of roughly 1:4 or 1:5.
The reason is not that research is unimportant, but that application feedback is more direct. Ordinary people may not understand a model improvement, but they can feel an Agent organizing email, executing workflows or handling everyday tasks. Manqi described this as the shift from chasing model capability to spreading applications.
13. Sam entered open source on a single act of credulity—and a few typos
In 2017, Sam, who had a civil-engineering background, worried that his original field would offer poor job prospects and switched into AI. In an Andrew Ng course, he heard that completing it would put students ahead of 95% or 99% of Silicon Valley engineers. Many people might not have believed it, but he said: “I really believed it then.”
In 2018, he decided that PyTorch’s dynamic graph was better for debugging than TensorFlow’s static graph, but he could not understand the underlying code. He mocked himself with the saying, “Nine of my ten apertures were open—the tenth was completely blocked,” and began searching file by file for spelling and syntax typos.
A few changes with almost no technical content were merged into the main repository, giving him powerful positive feedback. He then began real image and video open-source projects and connected with early core contributors including Soumith Chintala and Edward Yang.
14. DeepFaceLab proved work can bypass the degree barrier
In 2019, Sam worked on the algorithmic side of DeepFaceLab, aiming to let people who could not code generate film-quality face swaps with a mouse click. Before AIGC had reached the mainstream, deepfakes were one of the few applications with real public impact.
When Facebook held a deepfake challenge at ICCV in Seoul, it proactively sought out the project’s contributors. Others later suggested that Sam join companies such as ByteDance. What fascinated him most was: “Once you have work, no one will interview you on whether you can code, because that no longer matters.”
The positive externality of work is that it eliminates repeated self-verification. A technical lead can see the result and understand what someone can do; school, major and formal training no longer need to be the center of the conversation.
15. Following authors turned tech history from paper titles into people
Sam ported StyleGAN from TensorFlow to PyTorch and earned several hundred to roughly 1,000 GitHub stars. What truly triggered his interest was the quality of the code: “Who on earth made something this well written?”
The answer broke his assumptions. The main author, Tero Karras, came from NVIDIA’s Helsinki research institute and did not have a conventional university degree; one of the largest contributors to StyleGAN, Adam Paszke, was still a third-year undergraduate at the University of Warsaw. “Damn, these guys are incredible.”
Compared with the standard route of competition medals, elite schools and academic stardom, the contrast and drama were easier to remember. From 2019–2020 onward, Sam began collecting details about researchers’ development over the long term.
16. An illustrated windmill congratulatory note yielded an Agent roadmap
To reach Diederik Kingma, Sam drew an avatar with a Dutch windmill when Kingma received the Test of Time Award, then emailed his congratulations and said he hoped to meet him. Kingma thought it was “so much fun,” and the two began corresponding.
In 2024, Sam asked Kingma what direction he should pursue. The answer was clear: “You should work on AI Agents.” Foundation models require infrastructure at the scale of 1,000 GPUs, which individual algorithmic ability cannot solve; Agents rely more on creativity, and a strong individual can still build something good independently.
17. Alec Radford shows the cold bench can be worth more than consensus
Reinforcement learning was the early mainstream at OpenAI: robots solving Rubik’s Cubes, Dota and other tasks with clear feedback. Alec Radford was the only person persistently studying next-token prediction; his results came slowly and unevenly, while he faced the pressure of colleagues constantly producing wins.
In Sam’s recollection, GPT-1 was rejected by ICLR, while GPT-2 was published instead as an OpenAI blog post, as if to avoid the pain of another rejection. GPT-3 was not formally published until NeurIPS.
Radford was then in his 20s and an undergraduate graduate of Olin College. The story supports the ability Sam values most: judging a direction correctly and staying on the cold bench when there is no short-term positive feedback.
18. Camping outside OpenAI was itself a reckless experiment
In February 2025, because he was living in downtown San Francisco, Sam went to camp outside OpenAI and try to intercept people. He especially wanted to ask Jason Wei why the idea for chain-of-thought had appeared during meditation, and hoped to run into a researcher from the GPT-4o project.
He did not meet either target, but persuaded two layers of security to let him put a flyer on the bulletin board. He first compared himself to Vasari, the Renaissance chronicler of artists, emphasizing the historical value of recording technology figures, and then explained the cost of traveling from far away.
Manqi summarized the “social engineering” as “creating a sense of history” plus “telling the story of how hard it was for yourself.” Sam admitted the method was reckless, but in his view, even leaving behind a flyer counted as an extra reward.
19. Innovation sometimes grows from peripheral paths and relaxed states
RAG author Patrick Lewis attracted Sam not because his career could be planned smoothly. After completing his PhD, Lewis lived in Vietnam and Europe, joined a relatively obscure small company, and later entered a major platform after the company was acquired by Meta. He worked on RAG around 2020.
Sam’s explanation remains deliberately uncertain: casual travel and experience at a small company “may” make it easier for someone to notice new problems, but that does not mean such experiences necessarily produce innovation.
Manqi added that innovation often requires first spending a long time inside the context of a problem, then making the connection while walking, meditating or otherwise relaxed. She also warned that the Newton-and-the-apple story may be a legend embellished after the fact; narrative cannot substitute for evidence.
20. Good stories turn a technical list into “linked fate”
Llama author Hugo Touvron produced important work at 29. Sam placed that alongside the fact that Hugo wrote The Hunchback of Notre-Dame at 29: “The fates link up—and zap, the electricity starts.” He acknowledged that the power of the association depends on how it is told.
Alec Radford and Luke Metz were classmates at Olin College. After graduation they started a company in Boston and collaborated with Soumith Chintala on DCGAN. At GTC in April 2016, Jensen Huang credited Meta with the technology while overlooking Alec and Luke’s contributions.
Sam speculates that the two may have concluded that Boston’s contributions were difficult to get recognized and that they needed to move to the Bay Area. Alec went to OpenAI, while Luke Metz went to Google. For Sam, such details are not academic conclusions but a memory technique: conflicts, migrations and chance relationships help explain why the same people later crossed paths again at OpenAI, Google and Thinking Machines.
21. Elite researchers respond more to specific questions than grand praise
Sam first looks for resilience. When opening a new direction, results are slow and failures are frequent, while peers may be collecting rapid feedback in other fields. Anyone unable to tolerate sitting on the cold bench will ultimately produce only follower work.
Generic praise such as “you influenced a generation” usually has limited effect in relationship-building. Researchers would rather discuss a specific technology, a past direction or a future judgment. For recruiters and CEOs, an effective opening is a conversation around a concrete problem, not broad flattery.
The first generation of researchers was more like a group of observers on social media: they rarely posted opinions themselves, but liked, commented and joined discussions. As competition intensified, the new generation became more proactive, and public expression became a way to let opportunities discover them.
Asked about these people’s weaknesses, Sam did not force an answer: “Honestly, I don’t know. I haven’t thought about it.” The only balance he could confirm was that they must genuinely like the work and be willing to persist, without becoming fixated on the short-term outcome of every experiment.
22. The talent-search bottleneck is fragmented identity, not lack of information
An AI researcher may be active simultaneously on Hugging Face, GitHub, Google Scholar, OpenReview, Zhihu, Bilibili, YouTube, X, Substack and Medium. DINQ is addressing “not a content-richness problem, but a distribution problem.”
Overseas recruiters often start with a specific paper and search for its author and similar talent. Domestic demand is more often a broad filter—someone in a given field, with a top-conference paper, Chinese, born after 1995 and active on social media. The former has already defined the problem; the latter requires an Agent to refine the profile over multiple rounds.
When searching broadly for capabilities, Qwen and Kimi members are more likely to be returned. DeepSeek appears less often; Manqi and Sam speculate that one reason may be that its members disclose less publicly and are less active on social media, which could also relate to the team leader’s style.
When users search directly for institutions or models, Meta is searched frequently. They also search for the team behind Veo 3, Runway or Black Forest Labs, and care about a model’s contributors. Talent gets low-cost exposure, while the recruiting side finds value in information aggregation, similarity search and contact details, creating DINQ’s two-sided value.
23. Degrees remain probability anchors, but no longer provide a safety floor
Domestic recruiters often ask for “a Stanford degree, 3 top-conference papers, born in the 2000s and active on social media.” Startup pitch decks also like to say that their teams come from Stanford or CMU. Sam sees this as both inherited thinking and an attempt to anticipate the preferences of investors and resource holders.
For people with extreme agency, the route can be building in public: keep producing public work, because “if you’re really making good things, someone will find you.” The upside is high, but so is the downside; most people still use education to obtain mentors, peers and a degree of safety.
Sam did not claim that degrees are useless. Statistically, good schools still provide better research resources. MIT and Stanford already have AI majors, and Sam said that from 2025, some public schools in California also began offering AI programs, with undergraduate enrollment continuing to grow.
Manqi’s rebuttal was that technological change may move faster than individual development. AI’s ability to write code already exceeds that of many people, while the traditional programmer path and a job at a large company can no longer guarantee safety. Even the creative work promised by technological optimists will not be easy, because “people will do anything to avoid thinking.”
24. Recruiting will split hiring a person into buying an outcome
Sam projects that a need once requiring the hiring and six-to-nine-month ramp-up of fine-tuning or RL full-training staff could be completed by one expert in a week or a few weeks. Models will also sharply reduce the cost of aligning requirements between clients and vendors.
The result is that intellectual labor becomes faster and more fragmented, and there may be “no such thing as responsibility or loyalty” between people and organizations. Three moves in five years, once questioned, are already common among top researchers and may no longer be a core evaluation metric.
New screening signals will be more AI-native: Claude Code token consumption, the number of products used, personal task-automation rates and the specifics of what someone has built are closer to ability than an abstract title. Cursor recruiting from heavy users is an example of “one leaf revealing the whole autumn.”
DINQ’s matching engine can find people from a requirement and opportunities from a person. Sam expected it to add short-term project matching in April–May 2026. If the projection holds, the platform will connect not merely “people and companies,” but people with a specific piece of work.
25. An infinite game requires shifting human value toward judgment and state of mind
Sam believes most researchers have not systematically considered their social impact, with a few exceptions such as Ilya. More are simply pursuing “higher, faster, stronger”: longer, real-time and more consistent video, or lower inference costs and model capabilities approaching Anthropic’s Claude 4.5.
As title ladders weaken at top institutions and many people are uniformly called MTS, personal development is no longer centered on promotion inside a company. AI can already outperform many people at coding, so the harder-to-replace human contribution will shift further toward communication, organizational coordination and judgment; in specific AI-native work, people may be responsible for architecture and security testing.
These capabilities do not grow linearly by working 16–17 hours a day. Sam’s final advice is to protect health, emotions and curiosity, and “enjoy the daily collisions with AI.” Since temporary first place can be rewritten at any time, the only sustainable strategy is to treat the career as an “infinite game.”