41. Are AI Giants Buying Up Workers’ Screen Recordings? — Michael | Final Round / Hellyeah

Summary

• Some frontier AI labs are already knocking on doors to buy knowledge workers’ desktop recordings, and 3 signals have validated the digital-employee thesis. Qwen’s September and October 2025 papers trained digital employees on YouTube Excel tutorials with almost no preprocessing and “outperformed frontier models by 3x”; after raising $20B, xAI “stopped working on AGI internally,” using visual computer-use agents to take over legacy systems for customer support, with customers ranging from Tesla to JPMorgan; Anthropic paid a premium to acquire a company that trains digital employees at scale on video, then launched Cowork immediately.
• Raymond says Final Round invented the interview-copilot category and is now Hellyeah; Michael chose not to be a data supplier but to go all in on building digital employees, taking the opposite lesson from Scale AI. Michael’s words: “If he’d known transformers worked, he wouldn’t have done data labeling—he would have trained GPT himself after labeling the data.” The foundation is 2 years of interview-copilot usage from more than 20 million users, a revenue run rate above $10M, and the demand-side reality that “you won’t believe how many people in this world are looking for jobs.”
• The dividing line between a digital employee and a skill is autonomous decision-making, proactivity, and self-learning, while the real bottleneck in distilling yourself is self-knowledge. “The way you understand yourself and the way other people understand you are completely different,” which is why the system needs a VLM to observe the full workflow and conduct daily-reflect sessions that probe causality. The co-founder’s framework is “unity of knowledge and action” (知行合一): reasoning and actual behavior must match for the result to become a real person’s digital copy.
• The business-model benchmark is the agency: human-service firms earn gross margin from labor arbitrage, while digital employees earn it from software leverage and workflow memory. “In theory, once you distill a veteran from Ogilvy, you can replicate and use him infinitely,” and “as long as you renew, he won’t leave.” The ideal form is an SDK living in the code repository, event-driven enough to capture signals such as global CAC rising roughly 30% on the first day Israel struck Iran. Early customers are more than 20 leading North American tech companies; non-tech prospects are not expected to start serious discussions until Q4, mainly because they worry about capability boundaries and where human-in-the-loop controls are needed.
• Meta’s work-recording plan drew enormous pushback, but it also helped Hellyeah close a new financing round after the company had been doing it for 6 months. Raymond’s summary: “When you do it, they think you’re crazy; when Zuckerberg does it, they think he must have a good idea.” Michael’s defense is that acceptance depends on how intrusive the product feels, as well as branding and positioning: JPMorgan has long recorded every single keystroke and stored it for 10-15 years, while OpenAI Codex’s Record and Replay is internally called Chronicle; Raymond says Xiaohongshu has many related posts. Data from Individual plans is used for training by default.
• U.S. labor-market data have not visibly deteriorated because headcount can be fabricated and hiring processes themselves are expanding. A mid-level PM role at Stripe requires 14 interview rounds, with the process stretching to 5-6 months—extending Final Round users’ payment cycles. On layoffs, Michael’s view is that “it’s not really because of AI; they know their organizations are bloated,” with Elon’s decision to cut 80% of Twitter and then “discover that Twitter still ran like this” setting the precedent; Raymond adds that it could not have been because of AI.
• The endgame’s most important inversion is that foundation-model companies are distilling users’ work behavior while still charging those users. That is why Michael says OpenAI’s Q2 gross margin could be -90%: deep discounts are getting everyone onto Codex, and “if it doesn’t get the usage data and Record and Replay, it won’t be able to compete with Anthropic in Q3”; India gets 2 years free for everyone, while Singapore gets a 50% discount. Michael: “We are critical fuel for its second growth curve. Without us, it can’t be used.”

Deep dive

1. “You Won’t Believe How Many People Are Looking for Jobs”: From Jarvis to the Interview Copilot

• Raymond says Final Round invented the interview-copilot category and the company is now Hellyeah. Michael’s starting point was not a grand AI thesis: in summer 2023, he wanted to build a real-world Jarvis, starting with a meeting copilot that users repurposed for sales and online dating. After one user emailed, “Can I pay for this?”, he casually quoted $99 a month—and the user actually paid. The team then turned it into an interview copilot. Michael calls it a new category; in 2 years, it has reached more than 20 million users and a revenue run rate above $10M.
• In 2025, enterprise customers started coming to them instead: the interview copilot worked so well because “a resume is an advertisement and an interview is a performance,” while a perfect resume could be generated in 10 seconds. Hellyeah then built a trustless hiring platform that is essentially job simulation: AI generates a task, the candidate does real work, and AI scores the entire process. The work-trial analogy is “a bit like living together before marriage, rather than getting married after a blind date.” More than 20 colleagues at the company were hired this way.

2. Frontier Labs Want to Buy Screen Recordings: 3 Signals Validate the Digital-Employee Thesis

• In late 2025, some frontier labs approached Hellyeah to buy knowledge workers’ desktop recordings as reinforcement-learning environments. There was no market where work recordings could be purchased directly. Michael’s first signal was Qwen: it trained digital employees on YouTube Excel tutorials with little preprocessing and “outperformed frontier models by 3x.”
• The second signal was xAI. Michael says it had long believed the best AGI path was a visual computer-use agent: “Like its cars, it can see like a human and operate like a human.” After raising $20B, it stopped pursuing AGI and moved into deploying digital employees. Its first down-to-earth use case was customer support: no MCP, no APIs, no replacing legacy systems—“I’ll just send an employee to work, except this employee is super fast.” Customers range from Tesla to JPMorgan. The third signal was Anthropic paying a premium to acquire a video-training company and launching Cowork immediately afterward.
• The logic behind going all in was the opposite lesson from Scale AI: “He didn’t know transformers worked at the time. If he had known, he wouldn’t have done data labeling—he would have trained GPT himself after labeling the data.” So Hellyeah would not be a data supplier; it would keep the multiplier for itself.

3. A Digital Employee Is Not a Skill: Distillation Requires “Unity of Knowledge and Action”

• The difference from everyone simply burning skills into n8n workflows comes down to 3 things: making its own decisions, proactively taking on work, and having a self-learning growth path. A skill is only one step; the system also needs memory and recursive growth.
• The barrier to distilling yourself is underestimated. Gary Tan and Karpathy have both publicly promoted the idea, but “the way you understand yourself and the way other people understand you are completely different.” Hellyeah’s approach has a VLM observe the work, generate questions where the AI cannot understand what is happening, and run daily-reflect sessions that probe causality: “Why did you click here? Was there context beyond the screen?” Only when knowledge—reasoning—and action—actual behavior—match does it become a real person’s digital copy.

4. Choosing the Lane: No Coding, No Sales; Save Creativity for Zero-to-One

• Coding has the strongest reinforcement-learning signals but is too crowded—Cursor’s revenue had already surpassed $300M at the time. Sales has a more fundamental problem: Michael has seen too many AI SDRs sold by human sales teams of several hundred people. “If your sales AI is really that good, it should use its own sales AI to sell its own sales AI.” Hellyeah chose growth marketing based on 3 criteria: strong RL signals, a real commercialization path rather than a “big, beautiful” product like Harvey that enterprises buy but nobody uses, and a team with passion plus a first-mover advantage from building its own growth system.
• Raymond challenges the idea with the creativity behind a Trump billboard: can an agent come up with that? Michael’s answer has 2 layers. With enough context and taste skills, it can. But large customers do not need that: creativity is for zero-to-one; from one to 100, what matters is systematic, omnichannel, consistent growth, and “going viral doesn’t help them much.” Raymond cites GEO as a P10 example, and Michael agrees that once labor becomes 10x or 100x more productive, opportunities like this can be handled on the side: “Let Claude write it, let GLM execute it.” Chamath’s 8090 thesis follows the same logic: frontier models handle the architecture, while execution runs on Chinese open-source models.

5. Thank Zuckerberg: “When You Do It, You’re Crazy; When Zuckerberg Does It, He Has an Idea”

• Meta’s plan to record work screens and conversations met enormous pushback over privacy, alongside the obvious objection: “You’re clearly planning to fire me, so why would I distill myself and contribute it to you?” Michael says that after Zuckerberg pursued the idea, many investors who had previously dismissed screen-recording-based AGI as fantasy came asking about Hellyeah’s progress. The company then closed a new financing round quickly; it had started working on the idea 6 months before Zuckerberg.
• Michael believes the key variables are how intrusive the product is, along with branding and positioning. Raymond contrasts it with JPMorgan, which has long recorded every single keystroke and stored the data for 10-15 years to investigate money laundering and insider trading. Rewind’s pitch is that “not a single day of your life should be wasted,” and Raymond says Xiaohongshu has many posts about Codex’s Record and Replay. Michael adds that Codex calls it Chronicle internally: “It wants to upload all of humanity’s desktop work into history.” Under the terms, Individual-plan data is used for training by default, while Enterprise customers must actively opt out of training. Raymond’s reaction: “You pay that much every month and still serve as someone else’s fertilizer.”

6. Agencies Sell Person-Days; Digital Employees Live in the Code Repository

• The commercial core is different. An agency’s gross margin comes from labor arbitrage, while knowledge is difficult to retain; when the point of contact changes, delivery quality will probably drop sharply. A digital employee’s margin comes from software leverage and workflow memory: “In theory, once you distill a veteran from Ogilvy, you can replicate him infinitely.” After 90 days, an agency gives you a clean-cut deliverable. Hellyeah wants to give you a digital employee that understands the market, product, users, and channels—and “won’t leave as long as you renew.”
• The hardest delivery problem is not content generation but getting the context quickly. The ideal state is for the employee to “live in the code repository”: run npm install hellyeah, scan the codebase and database to build context, and even push a PR to standardize the customer’s data schema. That breaks down the wall between marketing and product/engineering and lets the system establish the causal chain behind “why did the advertising data drop?” in a second.
• One event-driven example: on the first day Israel struck Iran, CACs across all platforms globally rose by roughly 30%. When Microsoft cut 5,000 jobs, the system automatically generated a landing page targeting those 5,000 people. “This capability used to exist only in hedge funds. Why shouldn’t ordinary companies have it?”
• The early customer base consists entirely of more than 20 leading North American tech companies. Hellyeah expects to start talking to non-tech companies in Q4. Their main concerns are how far a digital employee can go and which steps still require a human in the loop; career managers also find it difficult to push through the idea of distilling their employees.

7. 14 Rounds for a Stripe PM: Why Labor-Market Data Look “Unchanged”

• Raymond is puzzled that the U.S. labor market appears to have improved since vibe coding took off. Michael says it depends on the metric: headcount can be fabricated, and large companies keep a “beautiful hiring pipeline” on their websites to support the stock price. A mid-level PM candidate at Stripe goes through 14 interview rounds because HR is afraid to use AI interviews to replace humans and will only add more process. The timeline stretches to 5-6 months. That also explains why Final Round users pay for longer than the market assumes: having 3 or 4 companies in play means paying for 6 months.
• Michael’s view on layoffs is blunt: “It’s not really because of AI; they know their organizations are bloated.” AI is the perfect excuse, and Elon’s decision to cut 80% of Twitter in one stroke and “discover that Twitter still ran like this” set the precedent. Raymond adds that this happened in late 2022, just after GPT emerged, so it could not have been because of AI. Chinese tech giants can cut 80% without vibe coding; they simply do not discuss it publicly.
• Michael’s view on workers preserves the original hedge: the market will “definitely get colder” in the short term, but he is relatively bullish over the long term. OpenAI is hiring people from investment banking as “high-end data labelers,” while distillation captures only “the current version of yourself”; continued learning requires repeated redistillation. Raymond pushes back: future improvements may already be covered by someone else’s present, leaving incremental gains at the societal level too small to matter. Michael concedes: “That’s why you have to learn faster and more efficiently.”

8. China-U.S. Differences: Competing on Intelligence vs. Labor, and the 10x Cross-Border Opportunity

• The arbitrage is concrete. An AI learning device costs RMB200 on Taobao and at least $300 in the U.S.; “there’s definitely a 10x price gap.” Eight Sleep—a North American white team paired with a Shenzhen supply chain, whose mattress shipment to DOGE went viral—had its latest round led by Sequoia China and entered China. One month later, a Chinese substitute appeared and was “just as good.”
• The investment and organizational read is that “90% of Chinese investors invest based on consensus,” a bit like Michael’s grandparents trading stocks: buy whatever is hot. Silicon Valley wants to create new categories, and capital commands a higher premium there. The root problem with going overseas is that trust cannot be extended—“you always feel that non-Chinese people are lazy.” More fundamentally, the paradigms differ: “Anthropic has launched more products and model iterations in the past 2 quarters than all Chinese tech giants combined. This is no longer a competition in headcount or hours; it is competition with intelligence as the unit.”

9. Flatter Organizations, the Meeting-Router CEO, and “Foundation-Model Companies Should Pay You”

• The first organizational change is already visible: no middle managers, and new cross-functional roles such as product engineer, design engineer, and AI builder. Michael’s own turning point came with a PPT whose outline he had prepared to hand to a designer. One second before sending it, he passed it to Claude; 5 minutes later, it had produced a polished deck. “Do we still need designers? That’s a pretty tough conversation.”
• The spicy version of the CEO thesis is: “AI won’t replace every CEO, but it will replace the kind of CEO who acts as a meeting router.” Once an enterprise brain aligns information in real time, the meeting-router’s work of getting everyone on the same page may no longer be needed. What remains is making tough decisions and serving as the cheerleader for irrational goals: AI can tell you rationally in 3 minutes that a rocket cannot be built, while Elon Musk will find a way to push it through.
• The closing inversion is the one to remember. Raymond believes every conversation you have with ChatGPT is part of its reinforcement-learning loop—and “it still isn’t paying you; you’re paying it.” Michael agrees with Raymond’s crazy idea that foundation-model companies should pay us to use their products: “We are critical fuel for its second growth curve. Without us, it can’t be used.” Alex Karp says foundation-model companies are monopolies. “He’s someone I really dislike, but I agree with him on that.”
• Michael also says OpenAI’s Q2 gross margin could be -90%, with deep discounts used to capture Codex usage data: “If it doesn’t get Record and Replay, it won’t be able to compete with Anthropic in Q3.” India gets 2 years free for everyone, Singapore gets a 50% discount, and the Middle East gets the same treatment. The U.S. is also beginning to push its own open-source models: Prime Intelligence has just raised $130M to build “America’s DeepSeek,” but distribution remains the problem.