134: Meta AI Talent Turmoil: Why $100M+ Couldn't Keep People | Taking Stock of AI Organizations with Pokee AI's 朱哲清
Summary
Meta acquired 49% of Scale AI for $14.3B and reportedly offered $300M over four years to poach talent, yet still failed to buy organizational stability. By September, some new hires had already left, including 2 who returned to OpenAI; 朱哲清 says quitting within 2-3 months is unusual even in Silicon Valley, and a larger package may not offset post-pandemic bureaucratic bloat, approval burdens and “extremely intense internal political infighting.”
A 5,000-person Superintelligence Labs is unlikely to beat OpenAI on talent density when the true core team behind a SOTA model is usually no larger than 50 people. 朱哲清 estimates 20-plus people set the model direction, 20-plus handle infra, with a small number of PMs; if the goal is simply to make Llama beat GPT-5, he says a focused team of 150-250 people would be more appropriate. “More people are useless,” especially in Meta’s bottom-up culture, where multiple teams race each other without clear model ownership or product landing points.
Top AI talent is not really chasing an uncapped salary; they want to be one of the 10 core authors of the historical moment when AGI arrives. 朱哲清 says that with a 90% chance of making that author list, many people—including himself—might accept the job “without being paid a penny”; so-called “working for love” still includes the reward of having one’s contribution recognized by history. That helps explain why OpenAI and Anthropic attract talent ahead of Google, and why Meta’s massive packages cannot cover its organizational and mission risk.
OpenAI, Anthropic and Google are betting on 3 different paths to capturing AI value. Anthropic is pushing text-based action spaces to the limit, with coding as its strongest application; OpenAI is prioritizing every consumer use case, with Sam Altman’s framing that “when I have 1B DAUs, AGI will definitely happen”; 朱哲清 speculates that Google is using its deep technical stack to serve developers and let applications evolve on their own, potentially producing “mass flowering” across language, video, image, world models and robotics.
The decisive variable in the chatbot race is not the model leaderboard, but whether ChatGPT can move from a desktop productivity tool to the go-to place on mobile. 程曼祺 says ChatGPT weekly active users rose from roughly 300M at the start of the year to 700M-800M; Gemini is also growing rapidly, but has not disclosed directly comparable activity data. Chrome gives Google an almost natural distribution moat on desktop; if ChatGPT can become the first app users tap after picking up their phone, mobile stickiness could finally cut off Google’s pursuit.
AI-native startups gain organizational leverage by hiring less, raising task abstraction and using AI to sustain networked collaboration. At Pokee AI, every new role starts with the question: “Can AI replace this within 2-3 months?” If yes, they do not hire; engineers no longer receive granular tickets, but own the architecture of a complete feature and hand execution to AI. The trade-off is 3-10x more output per person, which makes line-by-line review impossible; trust, complementary skills and avoiding duplicated work therefore matter more than traditional process.
Talent pricing will ultimately be set by capital’s judgment on AI-native revenue, and the most dangerous mismatch is labeling ad gains from legacy recommendation systems as generative-AI returns. 朱哲清 speculates that if year-end and post-Q1 assessments fall short of high expectations, capital and the talent market could cool together in 2H next year; 程曼祺 points out that public markets do reward Meta for “AI improving ads.” The longer-term risk is an education system in which students rush into LLMs, diffusion models and startups, producing highly homogeneous talent while neglecting the critical thinking that matters most in PhD training.
Deep dive
1. Meta bought the stars, but not yet a stable collaboration structure
程曼祺’s timeline is as follows: in mid-June, Meta acquired 49% of Scale AI for $14.3B and recruited founder Alexandr Wang as one of 2 heads of its new AI team; it then began making aggressive offers for talent, with the most extreme figure reported at $300M over four years.
The plot reversed by September: multiple recent hires were preparing to leave, including 2 who went straight back to OpenAI. 朱哲清 stresses that job-hopping is common in Silicon Valley, but “leaving within 2-3 months is extremely unusual,” especially when Meta’s package is usually larger.
朱哲清 explicitly frames the explanation as speculation: researchers from OpenAI and similar companies are used to working around a unified goal and do not want to spend their energy adapting to big-company politics; when 10-plus companies are competing for the same person, they also have enough options to walk away quickly.
2. Meta went from shipping in hours to layers of VP approval
朱哲清 recalls 2017-2019 as the most startup-like period he experienced at Meta: from writing code to approval to production, “the whole thing took a few hours,” and cross-team product execution was also extremely fast.
The formal process did not change after the pandemic, but the organization became increasingly bloated and added more VP layers. Every product iteration required finding multiple VPs to review it; they might not understand the project and had little time to spare, yet still issued different instructions, leaving execution teams feeling that their time had been wasted.
He partly attributes the mental-health problems described by former employees on social media to this political environment. The goal of a new AI team may be to restore a flatter, more entrepreneurial culture, but placing new hires inside the existing system does not automatically eliminate its old approval and interest structures.
3. Bottom-up was once Meta’s speed engine; it also diluted ownership
Meta has no conventional business-group boundaries. Its foundational infrastructure, products and codebases are highly open: even unrelated teams can inspect, challenge and modify one another’s code. That lowers technical boundaries, but also weakens the exclusive ownership of “this is my product.”
朱哲清 uses a project from his own experience to show the upside: after discovering a problem with the Facebook feed team’s model, he could intervene directly and overhaul the entire model within 6 months. “Can you imagine this happening at any other large company? Impossible.” Meta could do it then.
The real constraint was on resources. Even if a VP managed 300 people, they could not independently cut half the team and redirect the budget to buy compute; the organization could only expand scope and compete for resources within its existing allocation, making aggressive strategic reallocation much harder than at a startup.
4. Once people outnumber important tasks, collaboration becomes a fight over credit
朱哲清 gives the classic conflict: one team is responsible for deploying a model across the company, while another carries a 2% DAU growth target. If adopting the first team’s model delivers the metric, does the credit belong to the model team or the product team? To avoid losing credit, the latter may prefer to build another model itself.
程曼祺 distinguishes this politics from a genuine business trade-off. The latter can be analyzed by a data scientist using historical data and resolved by a VP; the hardest issues are those where contribution cannot be objectively separated. 朱哲清 summarizes them as “basically all being about an unfair split of the spoils.”
His threshold model is to cut 80% of low-value work first and allocate the 20% of critical work; as long as some critical tasks remain unassigned, the objective is easy to unify. Once “the number of people exceeds the number of jobs,” some people are pushed into the 80% of peripheral work and naturally start competing for the 20%.
A mature company cannot eliminate peripheral work entirely: at sufficient scale, even a 1%-2% improvement can cover a team’s cost and support revenue, users and the share price. So the company keeps hiring, new employees want to enter the main line, and politics becomes a structural product of slowing growth rather than an accident.
5. A 5,000-person AI organization may still lose to a 50-person core team
朱哲清 says Superintelligence Labs already has roughly 5,000 people, versus about 1,900 at Anthropic; it may even be larger than OpenAI’s entire headcount. If the goal is simply to make Llama beat GPT-5, “your headcount should not exceed OpenAI’s.”
The true core behind a SOTA model is usually no larger than 50 people: 20-plus researchers work out how to adjust the model, 20-plus handle training infra and engineering, and a small number of PMs decide which capabilities matter. If the objective is only to beat GPT-5, he considers a focused team of 150-250 people more reasonable.
Meta’s problem may also be internal horse-racing: with no clear ownership, any team can train on a similar idea, potentially producing multiple completed models with no clear destination. Amazon also runs races, but its research teams sit within business units, so their output naturally serves a product; even duplicated directions have somewhere to go.
Google is more top-down, with technically respected researchers setting direction; OpenAI and Anthropic align the company around benchmarks and mission. 朱哲清’s conclusion is not that Meta lacks people, but that 5,000 people in a bottom-up system are more likely to block one another.
6. Mission-driven organizations are powerful—and most vulnerable to a second mission
The advantage of OpenAI- and Anthropic-style organizations is that people do not have to repeatedly argue over who gets what share of the credit: “There’s a goal, I know what needs to be done, so I focus on that.” Even with thousands of employees, entrepreneurial alignment can persist as long as critical tasks still outnumber people.
But 朱哲清 warns that mission is the only adhesive holding such organizations together; once individual missions diverge, the organization can unravel quickly. “利益绑定是这个世界上最牢靠的关系” — economic alignment is the most reliable relationship in the world — while mission-driven differences cannot be reconciled over the long run by reallocating rewards.
From this, he speculates that the fundamental conflict between Sam Altman and Ilya may have been that the former wanted to expand AGI’s user coverage, while the latter prioritized safety, reliability and model technology; the 2 paths consumed comparable resources. “A mission-driven company cannot contain 2 missions.” It eventually has to choose a side.
7. Top research leadership comes from historical contribution, not a new title
朱哲清 believes that in a high-end AI organization like OpenAI, CEO ability cannot substitute for long-term accumulation in research direction, technical quality and talent judgment. At OpenAI, Sam Altman and Mira do not directly manage research; it is usually led by people who understand research and can command the team’s confidence. Removing those layers and having the CEO manage research directly “definitely would not work.” Meta’s gap is not a shortage of people, but a failure to put the right technical leaders in the right positions.
Meta does not lack strong researchers; it needs to place people who are respected and aligned with the current direction in the right roles. Those who truly command the room generally did important work before ChatGPT and are now in their 30s or 40s. David Silver, whose early AlphaGo papers and work established his authority, is a typical example.
Pure researchers often lose interest in implementation once they have shown that a direction works, while large companies need to turn technology into products. 朱哲清’s ideal leadership layer would pair 5-6 pioneering scientists with 5-6 execution-oriented technical leaders, then staff reliable researchers and engineers below them.
8. Scientists set the next decade; engineers decide whether the company survives 3 years
On xAI, 朱哲清 sees Elon Musk as the archetypal engineering thinker: once he accepts an existing scientific path, he uses scale, compute and highly complex engineering to push capabilities to a level others cannot easily replicate. Musk even challenged Yann LeCun with the claim that there are “no research scientists, only engineers.”
Google represents the opposite model: first seek the next scientific leap, then let that leap drive a product revolution. xAI is more inclined to hire people who will deploy known technology “at any cost,” so it should not be judged by the same standard as a top fundamental-research organization.
程曼祺 tried to compare the difference to the division of labor between China and the US, which 朱哲清 explicitly rejected: both countries have scientist-type and engineer-type talent. The more accurate formula is: “Engineering talent generally determines your development over the next 2-3 years, while scientist-type talent determines your development over the next 10 years—but you may not survive the next 2-3.”
9. The scarcity of fundamental work is ultimately still a researcher’s choice
朱哲清 believes the industry currently lacks direction: many innovations remain large steps along the main LLM path, but no clearly defined new track has emerged that explains LLM bottlenecks and fundamentally replaces the existing approach.
PhD students lack the compute to train frontier models and can only look for gaps that large companies have not covered; once they enter industry, they continue doing incremental work along the LLM mainstream. He relays Percy Liang’s blunt judgment from a language-model course: “Research is almost dead.”
But he rejects the idea that this is an unavoidable technology cycle. A good researcher usually has 3-5 ideas worth challenging; the real choice is whether to take a high salary at a large company and work on a project that ships within 6 months, or stay in academia “begging for compute” while advising students who may not cooperate.
Breakthroughs may come from 20-30 teams challenging consensus, with 29 failing and 1 succeeding. If no one voluntarily chooses the high-risk path, no new paradigm will emerge. The US and Europe once allowed people to challenge consensus for 10 years without results, but as government funding declines and capital flows into LLMs, the room for failure is shrinking.
10. Billionaires can provide compute, but struggle to provide a career path for research teams
The foundations of Bill Gates, Mark Zuckerberg, Larry Page and Sergey Brin, along with some corporate research organizations, are already funding projects with no short-term target whose only mandate is to “push the frontier.” 朱哲清 believes the resources themselves are ample.
The difficult part is attracting the second layer of talent to do the work alongside top scientists. If the direction succeeds, the whole team takes off; if there is no result after 5 years, members’ resumes and promotions may stall, while 5 years at a large company can move them up an established ladder.
He uses the choice between becoming a Stanford professor and joining a foundation as a research fellow: even if the foundation offers more money and resources, the professorship is usually more attractive. The bottleneck for original research is not only funding for the principal scientist, but also how the supporting team prices the risk of failure.
11. Top talent wants to be on the 10-person author list for AGI
For the most elite researchers, the central question is whether they can be part of the moment AGI arrives—and a core part of it. If the model called AGI has only 10 authors, they want their names on that list.
朱哲清 calls this motivation “working for love,” but that does not mean no return is required. Like pursuing a Nobel Prize or Turing Award, researchers love the work and also want the world and history to see their contribution. “It is hard to encounter a historical moment in one lifetime.” Once the opportunity appears, they may pay almost any price.
His extreme test is this: if they receive no pay but have a 90% chance of making the 10-person AGI list, many in this group would join—and he might do the same. Beyond a certain threshold, compensation is just a number; the direction, the position on the core path and the people one works with determine where talent flows.
12. OpenAI and Anthropic look like startups, but capital has made them too big to fail
朱哲清 ranks OpenAI and Anthropic in the first tier for talent appeal, with Google in the second. Google’s research capabilities and reputation could put it in the first tier, but large-company headcount limits individual upside; Meta should have been in the same tier as Google, yet recent disorder makes candidates hesitate even in the face of massive packages.
He says OpenAI and Anthropic cannot be treated as ordinary startups: they effectively have “unlimited capital,” because either one failing would hit the entire US AI investment cycle, giving them too-big-to-fail characteristics. AWS, Azure and Google Cloud also help sell their models.
程曼祺 contrasted this with Chinese talent data, where AI researchers remain concentrated in large companies such as ByteDance. 朱哲清 attributes the difference to the lack of a domestic equivalent that combines startup organization with giant-company resources, not to American researchers having an inherent preference for small companies.
13. The 3 companies are betting on 3 different value-capture paths
朱哲清 describes Anthropic’s main line as pushing a model whose action space is primarily text to the limit, making coding its strongest and most suitable application. OpenAI is trying to do every consumer use case as well as possible, using user coverage to push the capability frontier.
Google has taken a step back: it gives models and capabilities to developers and lets consumer and enterprise companies explore user habits that have not yet been determined. He compares this to “raising Gu,” giving 2,000 companies the technology and letting natural evolution eliminate failures and leave effective products.
“Buy the winner, or build a similar product and replace the winner” is a further inference from 朱哲清’s outside observation, not a public statement from Google. What he is more certain about is that Google is actively building the developer environment and ecosystem rather than trying to pick every end application itself.
14. Google’s slowness is the cost of letting 20-30 technical lines bloom at once
朱哲清 speculates that many Google projects started early but took longer to launch because large companies have longer release cycles. If the real push began at the end of 2023 and runs through the end of 2025, 2 years is a normal life cycle for a large product, while OpenAI and Anthropic can release more frequently around a single point.
Google may have 20-30 groups developing their strongest capabilities in parallel; once those projects clear the organizational process, language, video, image, world models and future robotics could appear in batches. Gemini and the stronger performance of Nano Banana mentioned on the show look to him like deep technical accumulation arriving “like spring wind overnight.”
程曼祺 relayed another view from US equity investors: Meta may be better positioned than Google because Facebook, Instagram and WhatsApp have application moats that are difficult to break, while Google faces search cannibalization. 朱哲清’s reservation is that Google’s technical accumulation may exceed that of OpenAI and Anthropic; business vulnerability cannot be assessed separately from research strength.
15. In OpenAI’s roadmap, 1B DAUs and AGI are the same problem
朱哲清 fully agrees with Sam Altman’s choice to put 1B DAUs ahead of AGI: “When I have 1B DAUs, AGI will definitely happen.” 1B users means almost every profession, problem and corresponding solution-provider is inside the system; if AI can cover their tasks, its capability definition approaches AGI.
程曼祺 says ChatGPT weekly active users rose from roughly 300M at the start of the year to 700M-800M. 朱哲清 says Gemini’s growth is also astonishing and may already have reached 300M or even 400M users. 程曼祺 notes that Google has not disclosed directly comparable active-user figures, mentioning instead that weekly-active growth was around 42% as of roughly April.
Chrome is Google’s “unbeatable natural moat” on desktop: when users open a browser, they still search Google first rather than necessarily entering ChatGPT. After Chrome was ruled not to require a sale, 朱哲清 also said he was more bullish on Google’s long-term position.
The missing data point is ChatGPT’s mobile-app activity relative to desktop. Chatbots are still primarily productivity tools, while mobile time belongs more to communication, social media, entertainment and daily-life services; only by becoming the first go-to place when users pick up their phones can OpenAI build stickiness that Google cannot easily catch.
16. AI-native organizations ask “why hire?” before deciding how to divide departments
Pokee AI had 8 full-time employees at the time and had completed a $12M seed round; its Agent had also been listed by a16z as a benchmark company for AI-agent workflows. 朱哲清 treats the small team as an organizational experiment, not a temporary state before expansion.
Whenever a new role appears, the team first asks whether “this thing could be directly replaced by AI within 2-3 months.” If the answer is yes, they do not hire. “Give it to AI if AI can do it,” because fewer people means less coordination and less duplicated work.
He argues against building a pyramid and instead favors a founder-centered network. Traditional hierarchy compresses the information problem created by leaders not knowing frontline capabilities; if a company has only 20-30 people and AI can retrieve everyone’s skills, tasks and capacity, leadership can set global priorities directly across the organization.
程曼祺 asked whether this degree of visibility would make employees uncomfortable and whether a network structure must hit a scale limit. 朱哲清 limits the model to startups and speculates that AI could extend it to hundreds of people; he acknowledges, however, that information feedback, employees’ willingness to accept AI-assigned work and the CEO’s ability to build the system will all become constraints.
17. AI compresses coding time, but amplifies task design and data engineering
朱哲清 asks for more abstract tasks, not finer decomposition: a feature that once took 5-10 days can now be completed in 2 hours with a prompt to AI. Engineers should design the full architecture across steps 1, 2, 3, 4 and 5, then hand the execution to AI.
The team has only about a half-hour stand-up each day: reshuffle priorities as they change, check owners’ progress and cross-task impact, then write the next day’s abstract tasks. Pokee’s AI also handles mapping the roadmap to people, GitHub comments and update summaries.
Because 朱哲清 has not led pretraining work at a large company, he cannot comment specifically on its organizational structure. Pretraining requires annotation, data processing, storage, training clusters and throughput engineering; if annotation shifts from contractors to in-house, that layer alone could require hundreds of people, making it far heavier than fine-tuning and application development.
On researchers’ complaints, mentioned by 程曼祺, that Scale AI’s data quality is poor, 程曼祺 says Scale AI cannot simply be blamed: for specialized data in chemistry or physics Olympiads, mathematical proofs and complex code, even an ordinary PhD may need to look things up. As models approach the frontier of expertise, usable human answers become scarce in their own right.
18. Once output is amplified, trust and complementarity become new management infrastructure
AI can complete in 1 day what used to take 5-10 days, and can also push an individual’s delivery to 3-5x the previous level. If all 20 direct reports are close to “ten-x engineers,” the leader is facing roughly 200 people’s worth of output and cannot review every item.
The most important invariant is therefore trust between people: what can go unchecked and what requires deep review depends on long-term bonding and alignment around goals. AI raises delivery speed but does not proportionally increase the CEO’s or C-suite’s review bandwidth.
Teams should also maximize complementarity. Once members break down abstract tasks themselves, 2 people with highly overlapping skills can easily implement the same sub-feature at the same time; even without a fight over credit, one person’s time is wasted, and duplicated work becomes more difficult to see than in a traditional ticket system.
朱哲清 has not yet worked out the boundary for scaling a technical team: AI coding lets everyone “do everything,” expanding each person’s reach and making departmental collisions inevitable. He dislikes using horse-racing to solve the problem because he has seen multiple companies damaged by internal races; Meta, with its extensive overlap, is the reference case.
19. Meta’s restructuring depends on Zuckerberg; the talent bubble depends on revenue delivery
朱哲清 believes Zuckerberg has near-absolute power and could change Meta if he truly decided to; the hard part is not recruiting Alexandr Wang and 10 scientists in the first step, but seeing the second, third and fourth steps of “the backyard catching fire” in advance.
The core capability of a professional manager is predicting 2-3 steps of personnel and organizational change, while founders are better at opening new territory. Sheryl Sandberg once played that role; after her departure, 程曼祺 speculates that the current leadership is made up largely of early members who have followed Zuckerberg for years, and that the new AI organization also reflects a founder-and-investor profile. 朱哲清 says he does not know why Meta no longer brings in a strong professional manager, only that “if they wanted one, they could definitely find one.”
On the market, he thinks it may cool somewhat in “2H next year”: if the assessment formed around March, after year-end and Q1 earnings, shows that AI-native product revenue is below high expectations, capital will contract and talent prices will fall with it. 程曼祺 points out that public markets reward AI improvements in advertising; 朱哲清 warns against combining gains from legacy models with generative-AI returns, calling that disconnect “the most frightening” risk.
The longer-term supply risk comes from education. Professors increasingly encounter students who immediately ask, “Can I work on an LLM?” or insist on bringing in a diffusion model, and some even drop out to start companies. He relays a question from a Stanford classroom about whether completing one course would qualify someone to lead a team building a SOTA LLM; the instructor laughed. The result could be highly homogeneous talent, while the independent thinking and critical thinking that matter most in PhD training are weakened.