The AI Giants Are Buying Up Your Screen Recordings

“A résumé is an advertisement and an interview is a performance — companies don’t actually know what a person’s real work looks like.” — Michael, on why he went from helping people win interviews to selling digital workers

In the summer of 2023, Michael set out to build a real-life Jarvis and incidentally produced an interview copilot — over 20 million users in two years. “You wouldn’t believe how many people in this world are looking for a job.” In late 2025 things took a turn: several frontier AI labs came knocking, wanting to buy the desktop screen recordings of knowledge workers on his platform to train digital workers. Qwen trained on Excel tutorial videos from YouTube and beat frontier models by 3x on the task. After Musk closed his $20 billion round, the internal focus shifted off AGI and onto deploying digital workers. And Zuckerberg pushing company-wide screen recording at Meta got him roasted publicly — while helping Michael, who had been doing the same thing six months earlier, close a round.

In this episode, Raymond talks with Michael, CEO and co-founder of Hellyeah (formerly Final Round, the company that invented the interview copilot category): what separates a digital worker from a skill? Why is an old hand at Ogilvy worth distilling over and over? Why does Stripe run 14 rounds for one PM role? And — when every large model is distilling you, shouldn’t they be the ones paying?

What follows is the full conversation, edited and condensed.

1. “You wouldn’t believe how many people in this world are looking for a job”

Raymond: The interview copilot category was in fact invented by Final Round, and over the past 12 months there have been a great many copycats. The company has now changed, with a new name, Hellyeah. Today we have Hellyeah’s CEO Michael to talk about that change.

Michael: Hello everyone, I’m Michael, CEO and co-founder of Hellyeah. We’ve always built products around the post-AI work context. In the summer of 2023, all of Silicon Valley was still in the GPT-3.5 era and every interaction was chat. We wanted to build an AI like Jarvis: one that hears what you hear, sees what you see, and helps you before you know you need help. So we first built a meeting copilot that could join online meetings and feed you lines in real time. We quickly discovered users were using it for sales and for online dating —

Raymond: How does one flirt in real time? I’m extremely interested.

Michael: I certainly never tried it (laughs). Later we found the interview use case was especially explosive. One user emailed asking whether he could pay — at that point you had to invite friends to earn free sessions, and there was no monetization at all. So I made up a big number, $99 a month, sent it over, and he actually paid. So we turned it into an interview copilot and created a new category: over 20 million users in two years — you wouldn’t believe how many people in this world are looking for a job — and revenue passed a run rate of over $10 million. Then it became an end-to-end career platform: AI writes your résumé, applies for you, prepares you for interviews, and there are even post-onboarding features. That was 2024.

In 2025, companies started coming to us, urgently wanting an entirely new hiring approach — because the interview copilot works too well: read off the teleprompter and someone who doesn’t understand computer science can, at least from an interview standpoint, answer fluently. Our judgement was that hiring in the AI era has to look at the substance of the work and can no longer be confined to fancy degrees, résumés and interviews — you can now generate a flashy résumé in ten seconds, so everyone is a perfect-score candidate. So we built a trustless hiring platform, which is essentially job simulation: AI generates tasks in an environment, invites the candidate to genuinely do a stretch of work, and AI watches and scores throughout.

Raymond: A bit like the Chinese reality show Offer, where interns are followed by a camera while six people in a studio commentate in real time. Do you use this work trial for your own hiring?

Michael: We do. It’s a bit like living together before marriage: you move in and work out the friction first, rather than getting married straight after the matchmaking. Our earliest hiring was straight off GitHub profiles — the CTO likes hackers, so we talked to a bunch from all over the world. Later, at an offsite, we found in-person collaboration far more efficient, so we consolidated people into four offices — San Francisco, Shanghai, Bangalore and Seoul — twenty-odd people in total, and every colleague came in through a work trial. We genuinely couldn’t assemble enough in Europe, so there’s no European office. People are still social animals.

Raymond: And then after work trial, you took another big leap forward.

Michael: After the product launched in late 2025, something unexpected happened: some frontier AI labs approached us saying the product was an excellent reinforcement learning environment, and they wanted to buy desktop screen recording data from knowledge workers — there is nowhere on the market you can simply buy work recordings, and they wanted to do large-scale digital worker training. At the same time the technical path had become very clear. The first signal was Qwen: in a few papers published in September and October 2025 they took Excel tutorial videos off YouTube to train digital workers, without much processing, and beat frontier models by 3x on precisely that task. The second was xAI — we have good friends inside, and Elon also said it on a podcast in early 2026: internally they’ve always believed the best route to AGI is a vision-based computer-use agent, the same as his cars, seeing like a human and operating like a human. After closing that $20 billion round they stopped working on AGI internally and moved to deploying digital workers. And the first deployment scenario wasn’t fancy stuff like legal or sales, it was very down-to-earth customer support. Sierra and Decagon help companies build AI-native chatbot systems — “migrate your old platform over to me.” Elon’s approach is that no legacy system needs replacing at all, no MCP, no API, the computer-use agent takes over directly — I’ll just send an employee in to work, except this employee is extremely fast. The first customers ran from Tesla to JP Morgan. And Anthropic paid a high price to acquire a company founded in early 2024 that trained digital workers at scale on video, and shipped Cowork right after closing.

All these signals say the same thing: this direction has been technically validated. And it has an enormous multiplier — suppose we were only a data supplier; that’s Scale AI ten years ago. He didn’t know then that the transformer worked. If he had known the transformer worked, he wouldn’t have done data labeling — he’d have finished labeling and gone off to train GPT himself. So we decided to go all in and build digital workers ourselves.

The 60-second version: our company didn’t start from a grand AI thesis, it started from a real problem — the hiring signal for most white-collar work is fake, a résumé is an advertisement and an interview is a performance. Stage one, help the candidate win the interview. Then we realized what’s valuable isn’t the interview process, it’s the substance of the work: how a person understands a problem, decomposes a task, expresses judgement, delivers results. Stage two turned hiring from a conversation into a genuine work sample. And since work can be simulated, evaluated and turned into data, the next step is to let AI genuinely participate, upgrading these datafied people into digital labor.

Raymond: In other words, you first helped people find jobs, then helped people keep jobs, and now you’re letting AI replace people.

Michael: Not replace — turn everyone into a manager. It used to be one leader with ten workhorses; in future it’ll be one leader with ten workhorses, each workhorse with ten digital workers. It’s still a story of a large productivity increase, and everyone gets a better future.

2. A digital worker is not a skill: “the self you perceive and the self others perceive are completely different”

Raymond: Since the second half of 2025 lots of people can do vibe coding and can burn their own skills, and there are workflow tools like n8n — freezing some capability of yours into software is already fairly common. How is that different from your digital workers?

Michael: Three things. First, a digital worker makes its own decisions, like a real employee. Second, it’s proactive — you also want your best employees to pick up work on their own rather than waiting to be assigned. Third, it can grow itself; it has a self-learning path. A skill is an important component of a digital worker, but it also needs memory and a recursive-growth iteration scheme. Garry Tan and Andrej Karpathy are both very publicly promoting “distill yourself,” and have distilled a lot and open-sourced it on GitHub. But not everyone is capable of distilling themselves — that requires very clear, objective self-knowledge. After observing large numbers of knowledge workers, we found that much of the time you have no idea what you actually did: the self you perceive and the self others perceive are completely different. So distillation isn’t just full-time tracking like that reality show; you also have to ask causal questions: why did you click here? Was there context off-screen? We use a VLM to observe work, generate questions where the AI can’t understand what happened, and do a daily reflect at the end of the day for the employee to answer — he might say, “I was just having a coffee and a flash of inspiration made me press that key.” My co-founder invented a framework he calls “unity of knowledge and action,” after the Chinese philosopher Wang Yangming: knowledge is the reasoning, how he reasoned through the task and why he made those decisions; action is the actual behavior. Only when cognition and behavior match do you have a digital worker — a digital copy of a real person.

Raymond: The first vertical you picked was growth marketing. That decision was made in the first half of 2026, when coding was the consensus — from a reinforcement learning signal or GTM standpoint, why not coding?

Michael: Coding certainly has the strongest RL signal, but it’s too crowded: Claude Code is competing there, Cursor was already over $300 million in revenue at the time, and we genuinely didn’t have the capital to fight. We had three considerations in picking a vertical. One, the RL signal has to be strong enough. Two, there has to be a clear real-world commercialization setting that gets us large volumes of data — digital workers can’t rely on signing pilot contracts with enterprises, they have to actually operate in a commercial environment, not the grand-and-beautiful case of “the company bought Harvey but nobody uses it” (though Harvey later broke through and has gone very deep in legal). Three, the team’s own passion — we like marketing, our consumer product grew on a fully automated self-improving agent system, so we have a uniquely strong first-mover advantage.

Raymond: Then why not sales?

Michael: The feedback cycle in North American B2B sales is longer than in B2C marketing. Worse, we’ve seen too many sales-AI companies end up maintaining a few hundred human salespeople to sell it — 11x, the various AI SDRs, all good friends of ours, but it doesn’t make sense: if your sales AI were really that good, your sales AI should be selling your sales AI. Which is exactly what we do.

Raymond: Like Claude Code being developed by Claude Code. On marketing, your GTM back then was extremely wild: a hacker house in San Francisco with over a hundred influencers invited, stripper ads, and the Trump billboard after he was elected — hugely controversial but very eye-catching. All of that requires human creativity. Could your agents today come up with something like a Trump billboard?

Michael: I think they can. When AI outputs one-shot, everyone’s web pages look similar; but give it enough context plus taste skills and the design suddenly has taste. Our system has seen enough things; fundamentally it’s large-scale permutation and combination. But more importantly: the large clients we work with now don’t need that much creativity. Creativity is for small companies going from zero to one; from one to ten and ten to a hundred, what you need is systematic and programmatic growth, omni-channel. Look at Wispr Flow — early on they also did a lot of influencer marketing, and now they only do ads, branding, billboards and partnerships. We’ve been burned by viral marketing ourselves: one video takes off, users flood in, and then there’s no hit the following week and it goes cold. A startup can accept that; a large company wants consistent growth — going viral doesn’t help them much. Previously a company would hold a QBR, the CMO and CEO would sit down and pick priorities: P0 might be a 500% return, P10 only 5%, and resources only cover P0 through P3. Now with AI you can eat every priority, even if P10 returns a tiny amount.

Raymond: GEO, for instance — optimizing so future models can pick up your corpus, with no obvious return today. Squarely a P10. But the digital worker is idle anyway, so it does GEO in passing, and over time there’s a compounding effect.

Michael: Exactly. When labor is 10x or 100x, a great many opportunities you previously didn’t dare touch become worth trying, and the cost of trying is very low — there are expensive tokens and cheap tokens, and you can have Claude write it and GLM execute it.

Raymond: Like Apple designing in California and assembling at Foxconn in Henan.

Michael: Everyone will operate this way over the next year or two. Chamath started a company called 8090 that wants to be the Accenture of the AI era, an AI Palantir for enterprise, helping large companies rebuild software at scale — and his logic is precisely that Chinese open source models are the best: frontier models do planning and architecture, and all actual execution uses Chinese open-source models.

3. Thanks to Zuckerberg: “when you do it you’re a lunatic; when Zuck does it he has vision”

Raymond: Among your customers, what’s the split between tech and non-tech companies?

Michael: Right now it’s all tech companies — over 20 leading North American tech companies as early customers. We’ll start conversations with non-tech in Q4, and their pushback has two parts. One, they aren’t clear on what digital workers can achieve or which parts need a human in the loop — traditional industries need someone to carry the blame and need an internal champion to push the deal forward; succeed and you get promoted to VP, fail and it’s a failed procurement. Two, it relates to “distillation”: we’ve run pilots of this screen-capturing-to-digital-worker pipeline inside North American enterprises, but many traditional companies no longer have a founder CEO, and a professional manager finds it very hard to push a story like “distill the employees.”

Raymond: I think it’s hard to push at any company. Meta pushed a plan this year: your screen is recorded while you work and all your conversations are recorded and analyzed — 24/7 monitoring of every step of how you work, which amounts to every employee distilling themselves step by step and gradually eliminating themselves. Meta has a founder CEO and still couldn’t push it: enormous pushback, first on employee privacy, second, you’re plainly going to fire me, so why would I also distill myself and hand it over?

Michael: Which is why we should thank Zuckerberg. Previously a lot of people in the market questioned our “train AGI on screen recordings” logic as fanciful; the moment Zuck did it, a great many investors came asking, “Michael, how’s that going?” — and we closed a new round quickly as a result. We were in fact doing this six months before Zuckerberg.

Raymond: When you do it, he thinks you’re a lunatic; when Zuck does it, he thinks he must have vision. Honestly, I was very uncomfortable watching your work trial promo videos on YouTube. I can’t accept someone watching me while I work, and I can’t accept today’s work becoming a screen recording used for training. Distilling myself — taking what I’ve written, distilling it and putting it on GitHub for others to use — no problem. But you distilling me, and I’m unhappy. It’s very subtle.

Michael: The key is whether the thing is intrusive. Lots of big North American companies have long done this, just for compliance: at JP Morgan every work computer has every single keystroke and every interaction with backend applications recorded, used to check anti-money-laundering and insider trading, and it has to be retained 10 to 15 years before deletion.

Raymond: Put that way, it’s true — back when I was in banking this setup was already frightening, it just wasn’t being used to train digital workers. Today is only one step forward. ByteDance also open-sourced a product called Mind Context that records your screen and analyzes your work; and there’s Rewind. Notice that when Zuck does this, what people think of is surveillance, a police state; Rewind’s line is “don’t let a single day of your life go to waste” — same thing, different framing. Another example: OpenAI shipped Record and Replay for Codex in May or June, and there were tons of posts on Xiaohongshu showing off that feature — I don’t believe OpenAI bought placement on Xiaohongshu. Why do people scramble to hand themselves over in some places and refuse absolutely in others?

Michael: It depends on branding and positioning. Codex’s angle is on the benign side — internally they call the product Chronicle, as in a historical record, meant to upload all of humanity’s desktop work into history. It’s a grand story, and for now they’re coaxing people into uploading voluntarily.

Raymond: Is that permitted under the privacy policy?

Michael: Yes. Individual plan data gets bundled up for training; Enterprise plan can opt out of training — but you have to go and opt out yourself.

Raymond: Paying that much every month and also serving as someone’s fertilizer is fairly painful.

4. Agencies sell person-days; a digital worker never quits

Raymond: You’re effectively now a giant labor contractor, a contractor of many digital workers. You do performance marketing — what’s the fundamental difference in what you finally deliver versus a traditional agency like Ogilvy or BlueFocus?

Michael: An agency’s commercial core is selling person-days and service: more people means more revenue, but knowledge is very hard to accumulate. The most typical case is an influencer marketing team — once the influencer gets big, there are usually two paths, either they become the boss or they marry the boss, and otherwise they leave. When your point of contact at an agency changes, delivery quality very likely drops sharply. Our goal is the opposite: every delivery accumulates into the system, and the next one is faster, more accurate, better. An agency’s gross margin comes from labor arbitrage — packaging the intelligence of people you can’t find or who are very expensive on the market and selling it to you at a lower price. Our gross margin comes from software leverage and workflow memory: once you’ve distilled an old hand at Ogilvy, in theory he can be copied infinitely and used infinitely.

Raymond: Distill Mad Men all over again and you’d get a lot of wisdom out of it.

Michael: Simply put: if 90 days later the customer receives a clean-cut deliverable, that’s an agency; if 90 days later the customer owns a digital worker that understands the market, the product, the users and the channels, that’s what we actually want to build. And digital workers aren’t shared across companies: the agent selling sausages understands sausages very well, the one selling microphones is deeply versed in consumer electronics, and the knowledge base, influencer connections and KOLs contacted are all different. After 90 days you get a knowledge base, and this team will never quit — excellent employees do leave; this team, as long as you renew, doesn’t go anywhere.

Raymond: What difficulties have you hit in actual delivery?

Michael: The biggest challenge isn’t getting AI to generate content, it’s getting AI to obtain context quickly — to get the core critical data. We think the optimal living state for a digital worker is inside the code repository: it’s an SDK, npm install hellyeah, it scans your repo to understand the business logic, scans your database to understand the business data, and builds context from that. This fundamentally breaks the divide between marketing, product and engineering: previously a marketer optimized channels, users came in and the product was weak, and neither side could construct causality. Now with a global view it can analyze how product changes affect marketing performance, and how imprecisely targeted traffic in turn affects product metrics. It can even change the repo — push a PR to help the customer standardize the data schema, though of course it only takes effect if they merge. That’s a lot of permission, so the first hurdle is how to get enterprises to trust us and allow access to that context, and we’re screening early flagship customers very carefully.

Raymond: With permissions that broad, how do I know you won’t take my data and build a competitor to me? Though I understand the value. Previously the person running ad buying — often a humanities graduate in China — notices the numbers dropped today and has no idea why; in fact the product shipped a new version and they don’t understand the product that well. The product person has to go to the data warehouse to pull numbers, two days pass getting it back, then a weekend intervenes, and the seven-day ad budget is already spent. At that point everyone remembers, “didn’t we discuss something last week?”, laughs it off and disperses. This happens daily at internet companies.

Michael: Right. With this system, causality is constructed within a second and the next ad campaign gets replanned — even if the decision is only “pause it until I figure this out,” that’s enormous value to the company. We call this characteristic event driven: proactively monitor events and respond according to the event. It’s a bit like a hedge fund — an earthquake or a volcanic eruption somewhere affects local share prices, but that capability previously existed only at hedge funds. Why shouldn’t an ordinary company have it? When we built our consumer product we had a system: Microsoft lays off 5,000 people today, and the system automatically produces a landing page targeting those 5,000 — don’t worry, we’ll help you find your next job. Fully automated event-driven marketing. Large enterprises badly need this: on the first day Israel struck Iran, CAC on every platform globally marked up roughly 30% — everyone anticipated future costs rising, so bidding went higher. Big companies have platforms like Smartly at a few million dollars a year doing optimization; small companies have no real-time processing capability and receive information long after the world has already moved. A frontier lab sees things two or three months before an ordinary San Francisco engineer, San Francisco sees them two or three months before the market, and people in AI see them six months before ordinary people — the gaps in between are enormous.

Raymond: It feels like I live in 2026 and everyone else has made it to 2027. So who will your second, third and fourth digital workers be? From Final Round through work trial you’ve collected vast amounts of interview and work data, and you could do any industry.

Michael: I’ll keep you guessing; there are internal discussions. But the form may not be confined to the traditional job divisions of knowledge work: since the divide between marketing and product has been broken, could a new species called product marketer appear that swallows both roles at once? Many teams no longer have designers; they’re all design engineers. There’s a new title in Silicon Valley called AI builder — as long as you have a Claude Code account you can do everything. Engineers used to split into front end, back end, iOS, Android; now it’s one person.

Raymond: Would you want to replace people in senior roles? Bankers, lawyers?

Michael: Those are very high-unit-price roles. I wouldn’t say we want to replace them, but we definitely want to distill them. I love the feeling of a control room — have you ever been in a shopping mall’s control room?

Raymond: I have, and I don’t like it. Why would you be interested in a control room?

Michael: A camera room has a hundred screens, each with something different going on. For someone with fairly severe ADHD like me, that’s deeply satisfying to the dopamine: I only have one pair of eyes, but I can watch a hundred things happening at once. Future digital workers follow that logic: a crowd of digital workers — internally we call it relentlessly, tirelessly — working like workhorses, and a human taking accountability. Investment banking and legal still need a great deal of human in the loop, and problems there are severe, which is why Harvey’s product, from a pure product standpoint, is very basic: ask PDF, fix your grammar and sentences, no whole-document rewriting. When we pick our next vertical, the first thing is that it has to be able to accept fairly radical change; legal and investment banking can’t, which doesn’t match our team’s temperament — for another team it might be a great vertical. We don’t just look at whether a market is good, we look at whether it suits us.

5. Stripe runs 14 rounds for one PM: the truth about the labor market

Raymond: What you provide is essentially labor as a service. How do you view the current job market?

Michael: The job market splits into high end and low end, white collar and blue collar, and by region — the Middle East is living happily on UBI, so-called UHI, Universal High Income; developing countries are in real pain, and Southeast Asia has livelihood problems driven by oil prices. I’ve been to over a hundred countries and I’m very sensitive to this, with strong empathy — which is also one of the biggest challenges I’ve hit as a founder: too much empathy, not cold-blooded enough. If I were as rational and cold-blooded as Dario, I might perform more efficiently. Back to the question: the labor market overall is certainly cooling in the short term, but AI will bloom entirely new jobs. “High-end data labeler” is now fashionable in Silicon Valley: OpenAI is hiring lots of investment banking people to label data. Some think that once you’ve finished labeling you’ve finished uploading yourself and will be discarded — that isn’t the logic. Every person is a living individual capable of learning and growing; what you distill is only your current self, your past self, and as long as you keep learning, distillation has to keep happening. Long term I’m fairly bullish: humans won’t have to spend time on low-intelligence matters and can fully enjoy high-intelligence gaming and thinking. The short-term impact is obvious.

Raymond: Since Cursor and vibe coding took off I’ve thought America was headed for mass layoffs — one engineer can do five people’s work. But over the past year the US labor market hasn’t visibly worsened, and this year it’s even improved. How do you read that?

Michael: It depends which metric. Headcount can be faked: many large companies’ job counts on their websites are watched by institutions — lots of hiring makes people think the business is strengthening and the stock rises — so they permanently maintain a beautiful hiring pipeline, and how many people are actually stuck in the middle isn’t clear. Another example: guess how many rounds Stripe — a very large fintech that supposedly moves very fast — runs for a mid-level PM?

Raymond: Five or six, I’d say.

Michael: Fourteen. There’s a written test, and an AI interview — HR daren’t say “since we used an AI interview I don’t need a human interview,” they only dare add two AI rounds on top of the existing process, and then run the whole line as well. Americans also like “this meeting overran, let’s move it to the same time next week,” so the whole hiring timeline stretches to five or six months — and by then a lot has changed in that person’s life; they may no longer be right for the job. Which is why Final Round users pay for longer than the market assumes: one company at 14 rounds easily takes two or three months, and with three or four companies in hand you’re paying for half a year.

Raymond: I hadn’t thought about that. I tried an AI interview once myself: facing the screen, able to see my own face, and it asks, “Raymond, what do you think is the most important transaction in your life so far, and what role did you play in it?” All behavioral questions, nothing about how to compute a DCF.

Michael: Today’s AI interviews skew behavioral; technical is coming imminently. There’s another interesting topic: can the work trial you did at company A be seen by company B? We think the property rights belong to the product and to the candidate; companies don’t see it that way — North American corporates are all fairly high-handed.

Raymond: On layoffs. Meta is cutting heavily, and Oracle and Microsoft cut earlier. Is there a pattern?

Michael: They expanded too aggressively before and the organizations got bloated; after bringing in AI they found a perfect excuse — “we don’t need this many people, because AI.” It isn’t actually because of AI; they knew themselves the organization was bloated. In some sense Elon started it: he cut 80% of Twitter’s few thousand people in one go and found Twitter kept running exactly the same, with no change at all — so other companies imitate. The capitalist system pursues maximum shareholder profit, so they follow naturally.

Raymond: Twitter’s layoffs were late 2022 and early 2023, when GPT had just appeared; it can’t have been because of AI — he simply felt the company only needed 20% of the people. Chinese internet is the same: the fifty-thousand and hundred-thousand-person giants today simply don’t need that many people; without vibe coding you could cut 80%, and with vibe coding it’s only more. Tencent and Alibaba are probably both cutting, they just don’t discuss it publicly because the social impact is too large.

Michael: So China also needs to “go to Mars” — create entirely new markets and entirely new jobs. So what should working people do over the next three years? First, still, learn to use AI — it sounds cliché, but what the market eliminates isn’t people who don’t use AI, it’s people who can’t use AI to reach their own goals. Fundamentally this is the same ferocious competition: twenty or thirty years ago 80 out of 100 got you into Harvard, and now a perfect score might not, because everyone’s average keeps rising. Same for startups; I’ve invested in quite a few friends’ companies myself and what I weigh most is fast learning and fast iteration — when tokens are cheap the cost of being wrong is very low. You’ll notice San Francisco is now very restless: the old environment was a Linear-style focus on one thing for ten years; now a team arrives and runs ten products at once, each with budget and headcount, ship it if it works, kill it if it doesn’t. OpenAI has four or five thousand people and probably five thousand projects (laughs). China is even more extreme — do you know how many office software products China has? Every big company has three different Buddies. And ElevenLabs, an audio lab, whose newest product generates ad creative — talent and money have both spilled over, and the boss says it’s fine, go explore. The newest Chinese meme: this summer students aren’t taking internships at all, they’re all founding companies — spend $200 on Claude Code and you can build 30 products. There will be more one-person companies in future, or small teams of a few people carrying large revenue.

Raymond: I now feel that friends around me who can’t vibe code are almost a different species; it’s already hard to hold a conversation.

Michael: But the opportunity is still there. My grandparents’ generation — Doubao has been pushed pretty well in China. Will American grandpas and grandmas use ChatGPT? No. There’s a big opportunity in there.

6. A 200-yuan AI study device on Taobao sells for $300 in America

Raymond: Will you have a China play?

Michael: We’re an American company with no China business. But I have a global mindset, with colleagues all over the world, and I’m somewhat more familiar with North America. I’m open to any exciting opportunity.

Raymond: My first reaction after looking at your business is a sweeping claim: China has far too much AI hardware. Search Taobao and it’s all AI bears, AI dinosaurs and AI pen holders with a small model inside. If Hellyeah’s growth marketing is genuinely that good, this stuff can be sourced directly on 1688, and the only question is how to properly push it overseas — SEO, web pages, ad buying, the whole set. If the model is right, in theory that’s a 90%-gross-margin business.

Michael: We are in fact helping several leading Chinese going-global companies with preliminary plans, skewed to e-commerce. We understand the American market well, and we understand the American AI market well, and the premium in there is substantial: an AI study device that sells for 200 yuan on Taobao sells for at least $300 in America — a 10x price gap certainly exists, with adjustments needed for cost, sales and consumption habits. There’s an interesting product called Eight Sleep, a water-cooled mattress selling for several thousand dollars: a white North American team with a Shenzhen supply chain. Their hottest moment was giving DOGE a few dozen free mattresses to support them sleeping in the office, which went straight viral.

Raymond: That’s good marketing, Republicans would all support it. But it’s also far too expensive; it deserves a cheaper equivalent — doesn’t China love building cheaper equivalents most of all?

Michael: Here’s the exciting part: Sequoia China led Eight Sleep’s latest round, Eight Sleep entered China selling at an enormous price, and a month later the Chinese equivalent appeared, doing it just as well. Anything without an exclusive patent, Chinese people learn and execute extremely easily, and the Shenzhen supply chain is fiercely competitive. In some sense it also gives consumers more choice — you can drink Coca-Cola, or you can drink Pepsi.

Raymond: You have an office at MoSpace, Shanghai’s AI incubator, so you’ve watched a lot of Chinese AI application companies up close. What’s the difference between China and the US?

Michael: Silicon Valley wants to create entirely new categories; China isn’t interested in creating new categories and wants fast commercialization. Dreaming really is easier in Silicon Valley, and the capital premium is higher too — too much money, and too many talented people who need that money. China skews toward consensus investing: 90% of Chinese investors invest by consensus, a bit like my grandparents trading stocks — whatever’s hot, buy that. Two different logics, and on paper returns it’s hard to argue who’s right; but young founders genuinely do prefer dreams over fast commercialization. Silicon Valley has also produced a new mechanism: funds like General Catalyst don’t take large equity stakes and instead do lending — help you with marketing, growth and customer introductions, and take a cut of revenue. When you’re very sure something works, not diluting your own equity is actually better for the founder.

Raymond: You could do something similar.

Michael: We’re thinking about it, but financial products are high risk and not something a startup can take on casually.

Raymond: Going global is consensus — everyone knows it’s easier to make money abroad, but very few genuinely pull it off; MiniMax’s Talkie counts as one. Why?

Michael: First you need overseas background and genuine understanding of the market: the founder has to be international, and so do the middle managers. ByteDance has had a painful time — I have many friends working there: global expansion, and local employees versus expatriated employees certainly have cultural and language conflicts, and you have to keep moving fast amid enormous internal chaos, repairing the plane while flying it. A lot of Chinese companies going global find it very hard to accept the foreign approach to work. The first culture shock is “why isn’t this 996” — in fact plenty of American companies do run 996. The second is that trust can’t be extended: Chinese people don’t fully trust outsiders, and there’s a persistent sense that non-Chinese are lazy. My wife’s company has a Brazilian colleague and she was shocked, saying it was the first time in her life she’d been out-worked by a Brazilian — expectations were simply too low, the assumption being that showing up is good enough, and then he did a lot of things she hadn’t managed. Every person is an independent individual and you have to accept difference. More fundamentally: the products and model iterations Anthropic shipped in the past two quarters exceed all Chinese tech giants combined — this is no longer competing on headcount or hours, it’s competing with intelligence as the unit. China actually has plenty of intelligence; it’s just that everyone has been competing on labor. Competing on intelligence requires some new design in organizational structure.

7. AI won’t replace every CEO, but it will replace the CEO who is a meeting router

Raymond: You’ve gone from hiring all the way to labor solutions, so you must think a lot about organizations. How does organizational structure in the AI era differ from before?

Michael: First, no middle managers. Second, lots of cross-role positions get empowered into existence by AI — product engineer, design engineer. Our team used to have designers, and then engineers took design into their own hands and found they could do it. There was a trigger: I was making a deck once, had written every page’s outline and requirements, and was about to hand it to the designer to polish into something presentable. A second before sending it, I thought about it and passed it to Claude — five minutes, and an extremely refined deck appeared in front of me. That’s when I started thinking about whether we still needed designers. It’s a fairly tough conversation. The gaps between roles keep narrowing, and I need a team capable of doing things outside their own capability range, or at least knowing how to attempt them: an engineer can be very technical with no product sense, but when a requirement lands, he knows how to find the product sense through multiple rounds of conversation with Claude.

Raymond: People still use tools. So I think Richard Liu’s logic is right — set up a school for 600,000 couriers and turn them into operators and maintainers of embodied intelligence. The education AI transition requires will be extremely important, but the market lacks that solution; nobody knows what education is needed.

Michael: There’s a saying in AI circles: learn slowly enough and you don’t need to learn at all — things change far too fast. I live in enormous AI anxiety every day: even working five or six days a week, AI companies worldwide have very likely shipped twenty or thirty changes worth my attention, in products and features, and it’s simply impossible to keep up.

Raymond: And some people use an iPhone 4 for four years. Not everyone needs Fable; Opus 4.5 can also transform your life completely. Last question: your employees work on the digital worker business, so they must also realize they could be replaced, be distilled. What is a CEO’s value in this era?

Michael: My own computer has screen recording on too. Everyone can ultimately be replaced — but what can be replaced is your current self, not your future self. As long as you keep learning and keep growing, there’s always new knowledge that the present stage of distillation can’t capture.

Raymond: I’ve heard a contrarian version: you distill your own past, but the large model distills a great many people — your future progress may already be covered by other people’s present. You strive away, and the actual increment at the level of society is negligibly small.

Michael: True, which is why you have to learn faster and more efficiently. Being a CEO looks fancy, but what you actually do every day is tough decisions — the easy, pleasant decisions have already been happily made by colleagues, and what reaches me at the end is all the unpleasant stuff. Either the AI is unpleasant or I am; someone has to make these tough decisions. The big future direction is an agent with the OKRs set up, and it goes and executes on its own — give it a goal, it decomposes the KRs and does the work, like a real employee. I’m quite looking forward to that day. But it isn’t possible yet, and human nature is irrational: you say you want to build a rocket and the AI tells you in three minutes that it can’t be built; it’s extremely rational and unable to execute. Elon Musk will say it’s fine, I’ll find a way to push it through. What a CEO has to do is broadcast the irrational goal and encourage everyone — to a degree, the CEO functions as head cheerleader. Something spicy: AI won’t replace every CEO, but it will replace the kind of CEO who is a meeting router — a lot of professional-manager CEOs do nothing but relay messages, you align with him, he aligns with you, meetings without end. In a year or two, enterprise-level brains appear, information becomes unusually open and transparent, and structures get flatter and flatter — an American engineer can directly access the decision data of someone at the European headquarters, and context alignment becomes real-time and automated, so you no longer need a meeting router.

Raymond: The most frequent and most time-consuming phrase in big-company jargon is “align” — which fundamentally means two departments’ interests don’t match and are being forcibly yoked together.

Michael: AI employees naturally help with that: they live natively between the two departments, with no ego and no emotions.

Raymond: I learned a lot today. This era has arrived regardless — whether you like it or worry about it, you may already be on the path to being distilled: every conversation you have with ChatGPT is essentially its reinforcement learning loop.

Michael: And it isn’t even paying you yet — you’re paying. One or two hundred dollars a month to let someone else distill you. So Alex Karp is right that large model companies are monopolies — he’s someone I dislike a great deal, but on this point I concede it.

Raymond: Never mind Anthropic today; Zhipu is distilling you, GLM is distilling you, Doubao is too. It’s no different from scrolling Douyin back then: Douyin distilled your aesthetic preferences, and today what’s being distilled is your work behavior. The future is here, it’s just unevenly distributed. So I have a crazy idea: large model companies should pay us to use them.

Michael: Agreed, they genuinely should — we’re important fuel for their second curve; without us it doesn’t get off the ground. That explains why OpenAI’s second-quarter gross margin may have been -90%: huge discounts had everyone racing to use Codex, gross margin looks terrible, but they had to push it through — without the usage data, without Record and Replay, they couldn’t compete with Anthropic in Q3. So far the effect is clear. The landscape is fairly frightening: America is now starting to push its own open source models too, like Prime Intelligence, which I saw on Twitter today just raised $130 million to be the American DeepSeek. The problem is distribution — OpenAI can pay for distribution.

Raymond: Then Prime Intelligence may need to talk to Hellyeah about growth hacking.

Michael: Growth hacking is simple: push the thing to where it’s cheap. OpenAI is in fact distilling Indians — it and Perplexity both gave all of India two years free, because the usage data is worth more than the subscription fee, and they can’t collect money in India anyway. They signed a half-price agreement with Singapore, and the Middle East as well. Absolutely wild.

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.