Summary

FinalRound AI reached about 7M total users with a 20-person team, has paying users in 150+ countries, and crossed $10M in annualized ARR last quarter. Michael does not hide the fact that job-search products naturally have low retention; he defines the growth challenge as continuous acquisition: more than 1B people look for work globally every year, “a huge market, a very fluid market.” The company calculates ARR by multiplying the previous quarter’s revenue by four, while also citing more aggressive methods such as multiplying the highest single-day revenue by 365, the highest hourly revenue by 24 and then 365, or the highest hourly revenue by 20.
The real product wedge was not a grand AI career platform, but a user with an interview in 10 minutes who was willing to pay $99 immediately. The team had initially explored Jarvis-like real-time conversational intelligence that could offer proactive prompts, eventually narrowing the general Meeting Copilot into an Interview Copilot that keeps the bot invisible to interviewers. Michael’s judgment was direct: “This is a very, very painful pain point.”
The 12-week jump from $1M to $3M ARR came not from a single breakout hit, but from turning counterintuitive experiments into an organizational capability. The team raised the monthly price from $99 to $150, shut down the free trial, rebuilt the brand, and kept compounding 5%–10% improvements in details such as button placement and color. It now runs 100+ experiments every week, allowing product, engineering, operations, and marketing to allocate traffic and test their own hypotheses.
FinalRound’s leverage with 20 people comes from agent-driven growth infrastructure, not merely adding an AI button to existing workflows. The team generates 4,000–5,000 long-tail SEO pages a week, tests 500 ad creatives, maintains long-term relationships with 100+ influencers worldwide, and turns signals such as layoffs into pages, podcasts, social content, and campaigns within 10 minutes. When hiring, it even asks: “Can you build an AI that replaces all of your early colleagues?”
Behind the $6.88M seed round and 13 institutional investors was Silicon Valley capital’s renewed view of young AI application founders around 2024. Michael observed that established VCs once preferred executives from major tech companies and foundation-model companies, then began to believe that young hackers could rebuild traditional sectors such as HR, payments, and healthcare; investors who rejected the application layer a year earlier came back to follow up at the start of this year. He describes young founders challenging established teams as “乱拳打死老师傅”—wild punches taking down the master.
The team accepts extreme intensity, but Michael ties its legitimacy to equity and high reward while preserving mechanisms for exiting the high-pressure state. The office is designed for 24-hour use, with beds and a shower; team members can “go straight to sleep when tired, and work whenever awake.” In response to 曼祺’s questions about extreme hours, Michael acknowledged that traditional 996 is criticized precisely because it lacks high reward, while offering unlimited PTO and proactively telling people showing signs of burnout to take 2 weeks off.
The larger growth opportunity is in B2B: use AI to simulate a real company and its first week of work, then rebuild resume screening and traditional interviews. Candidates collaborate with real or virtual roles in a programmable workspace, and employers evaluate not only the result but also which tools were used and how the result was achieved. The product is iterating with design partners and plans to launch in September. Michael calls it an opportunity that could take the company from $10M to $1B, while explicitly noting that enterprise adoption speed, the first career category to target, and how aggressively the consumer product can evolve will all affect the outcome.

Deep dive

1. FinalRound Started in the US and Defines Itself by the Global Market, Not as a China-to-Overseas Expansion

曼祺 places FinalRound on a third path to globalization: neither a China-based company serving overseas markets nor a domestic company that scaled first and relocated wholesale, but a non-ABC Chinese founder building directly in Silicon Valley. Michael, 27, graduated from UIUC, went on to Yale for an MBA, dropped out in 2021 to start a company, sold it, and began this venture in October 2023.
Michael defines the product as an end-to-end career AI platform: AI resume editing, automated applications, and Avatar mock interviews before the interview; a real-time Interview Copilot during it; and a focus on mid-level and senior-level knowledge workers in North America and Europe. “Give job seekers AI superpowers” is the shared narrative across the product line.
The company has about 7M total users, with paying users in 150+ countries. ARR crossed $10M last quarter, with only 20 employees globally across San Francisco, Shanghai, and Bangalore. Michael emphasizes that it is first and foremost a US company, but hiring is location-agnostic: “Wherever you are, if you’re an AI hacker, you can join us.”

2. An Email About an Interview in 10 Minutes Compressed the Jarvis Vision into a $99 Product

After selling his previous company, Michael saw a flood of conversational applications during the wave when GPT-3 and GPT-3.5 had “just come out,” but believed the chat interface was still behind the curve: users had to enter text, voice, or images before the AI returned an answer. What he wanted to build was real-time conversational intelligence—a system that could proactively observe, listen, and help.
The most concrete prototype came from Jarvis in Iron Man: “See what Iron Man sees, hear what Iron Man hears, and then proactively support him.” Meeting Copilot, launched in October 2023, could enter a meeting context, understand the visual scene and conversation in real time, and be configured for work meetings, sales meetings, or even online dating.
The product had no subscription plan at the time; free sessions required users to invite 3 friends. One user wrote in saying they had an interview in 10 minutes and asked whether they could simply pay. Michael casually quoted “$99 a month,” far above what most AI applications charged then, and the user subscribed immediately. “That’s when I knew this was a very, very painful pain point.”
The team then narrowed the broad exploration into FinalRound AI and Interview Copilot. The first-generation UI was a real-time teleprompter for interviews, and the team attracted early attention through provocative short videos on Instagram and TikTok. The product moved from asking “What should AI look like?” to serving a clearly defined, high-intent paid use case.

3. The Invisible Assistant Triggered Cheating Controversy, and the Team Chose to Capture Existing Demand First

Interview Copilot does not add a bot to the meeting room; only the user can see the assistant. Michael calls this the key difference from the AI note-takers of the time, noting that next-generation products such as Granola later also stopped making the bot visibly enter meetings.
曼祺 asked whether this design constituted cheating. Michael acknowledged that the product “challenges conventional thinking to some extent,” but his decision rule was not to eliminate controversy: “Every innovation challenges conventional thinking to some degree. If we don’t do it, someone else will.”
On Cluely and its “cheat on everything” marketing, Michael located the difference in the interaction model. Cluely started with coding interviews and required users to see a question, take a screenshot, and activate it manually; FinalRound continuously listens to the conversation and watches the meeting hands-free, then proactively prompts the user. Later, when discussing other products, Michael said Cluely’s most admirable qualities were its “consequence-free marketing” and its founder’s independent personality; Riley was the name 曼祺 had mentioned earlier when introducing Cluely.
Asked whether interview-assistance tools could eventually be banned by policy, Michael offered no firm answer: “It’s more about watching how the market evolves. Every product carries regulatory risk.” The risk has not been resolved; it has simply become part of the conditions under which the team continues.

4. A Job-Search Product Does Not Depend on Permanent Retention; ARR Depends on Methodology and Continuous Acquisition

曼祺’s challenge went straight at subscription quality: job seekers may interview intensely for 1 month and then stop using the product for 7 or 8 months, while fast-moving products often claim ARR before they have even been live for a full year. She therefore asked whether the $10M truly represented stable revenue or merely a short-term peak.
Michael accepts the limited retention candidly: “In any case, the amount of time people spend looking for jobs is limited.” The team does not treat long-term retention as the primary problem; it focuses on continuously reaching and helping more job seekers. By his market definition, more than 1B people look for work globally each year—a massive, highly fluid market.
FinalRound calculates ARR as “the last quarter multiplied by four.” Michael calls this relatively conservative, having heard of more aggressive methods based on multiplying the highest single-day revenue by 365, the highest hourly revenue by 24 and then 365, or the highest hourly revenue by 20. In his view, ARR primarily expresses company momentum and business volume, and cannot be interpreted separately from real users, social content, and usage feedback.

5. An Executive Unemployed for 8 Months Reconnected Growth Metrics with Product Value

The paid user Michael remembers most clearly was a senior professional seeking a Director of Product Marketing role, with children and a family, who had been interviewing for 8 months. After reaching a tenth-round interview at a large, high-growth payments company, the candidate was rejected—an extremely costly failure for a role requiring a specific referral where opportunities were already scarce.
After FinalRound launched AI Job Hunter, it mass-applied on his behalf. He got a new interview within 7 days and an offer on day 17, while using the full product suite before, during, and after interviews. Michael summarized the outcome: “In the eighth month of unemployment, he got an offer for his ideal job.”
The feedback strengthened the team’s sense of mission. On one side of the labor market, Meta spends “hundreds of millions of dollars” hiring talent; on the other, large numbers of highly capable professionals are unemployed for various reasons. FinalRound is trying to help more people land the jobs they want and improve the connection between opportunity and talent.

6. HF Zero Took a 5-Person Team from $1M to $3M ARR

The product quickly crossed $100K ARR; at a $99 monthly price, that required only about 100 paying users, or 3–4 new payers a day on average. The team was still bootstrapping with its own money when it joined HF Zero in April 2024 and began turning scattered product momentum into a systematic high-growth startup.
HF Zero is primarily for serial founders, selecting 10 teams per cohort. Its Alamo Square residence was a large house where “the floor creaked when you walked,” but it offered 3 meals a day, a private gym, and a boxing coach. Michael says boxing made him “more confident and more aggressive.”
The company had about 5 people when it entered the program, mainly Michael and his co-founder living inside the accelerator. Over 12 weeks, ARR rose from about $1M to $3M, while monthly revenue climbed from nearly $100K to nearly $300K.
Michael felt the geographic differences in concrete terms. Palo Alto concentrated engineers from major tech companies, discussing compensation and packages “so large they don’t feel like starting a company”; San Francisco centered on cutting-edge AI applications and had a higher density of young founders. He even cited the view that “the average age of the previous YC cohort should have been under 20,” presenting it as a personal impression rather than verified data.

7. Growth Is Not Inspiration; It Is an Experiment System That Lets Counterintuitive Conclusions Win

The HF Zero period had 2 tracks: systematically running A/B tests, new products, and marketing channels, while also taking actions that looked unreasonable at first glance. The most direct advice from an advisor was to raise prices. Although the team thought $99 was already expensive, it raised the monthly price to $150, with no obvious collapse in willingness to pay—a single price change representing roughly a 50% increase.
Other experiments included shutting down the free trial, rebuilding the brand, and testing referral codes and scarcity. Michael says the new brand was created by a design team that had “worked with Tesla and OpenAI.” But he emphasizes that experiments do not guarantee correctness: “Some work, some don’t.” Their value is letting real traffic, rather than preset logic, determine the outcome.
Most growth does not come in the form of 2x, 5x, or 10x jumps. It comes from small 5% and 10% improvements compounding continuously. In some tests, whether a button sat on the left or right, or was red or black, produced a 50% conversion difference. Product taste must be validated by data rather than left as an aesthetic debate.
The company now runs 100+ experiments simultaneously every week, using an internal AI-native A/B testing platform for rapid launches and rollbacks. Operations, marketing, and development can all propose product ideas and receive a share of traffic: “If the experiment is good, we promote it; if it’s bad, we go back to the old version.”

8. The $6.88M Financing Bet on Both Revenue and a Window for Young Founders

After the accelerator, the company raised $6.88M, or nearly $7M, in a seed round backed by 13 institutions. Michael believes studying and living in the US improved his communication and participation in local activities, while his technically rigorous co-founder strengthened credibility. But investor conviction ultimately came from traction, trust in the founders, and broader industry change.
Some investors had been watching since the product launched in October 2023 but had not yet developed enough conviction. Michael believes subsequent traction and growth gradually gave those investors the confidence to commit.
Michael observed that before 2024, established VCs favored executives from major tech companies and foundation-model teams, while young founders were often assumed to be suited to social or dating products. By 2024, capital had begun accepting young people using AI to rebuild HR, payments, finance, and healthcare. He uses Cursor as an example and attributes the shift to AI requiring the old development logic to be overturned—essentially, “wild punches taking down the master.”
The shift in investor preference also reached the application layer. A firm that “looked down on the AI application layer” a year earlier came back to ask about FinalRound at the beginning of this year because the fund had started studying applications. Michael’s impression is that the US application startup wave was not meaningfully earlier than China’s. Foundation-model companies generate larger revenue figures, but that does not invalidate the commercialization potential of applications.

9. Small Angel Checks and Thousand-Person Music Parties Are Both Silicon Valley Relationship Assets

Michael has personally invested in 28 companies, writing checks of perhaps $3K, $5K, or $10K, funded from his salary and savings. He compares the logic to “gold lying everywhere”: one person cannot personally pursue every opportunity in 24 hours, so small amounts of capital can support founders who are qualified and sufficiently passionate.
These investments are also a form of socializing, like walking a dog together or going to a music festival, except the subject is startups and exciting ideas. If 100 people each invest $5K, the cash itself remains modest, but Michael believes “the support of 100 people” creates energy and follow-on opportunities far beyond the size of the checks.
When choosing funds, he does not equate brand with service quality. Established firms such as Sequoia, a16z, Khosla Ventures, and Google Ventures can create FOMO and introduce customers, but may spread their attention thin across too many deals. Newer or more focused funds such as the XYZ Ventures he mentions, First Round Capital, and Goodwater may offer more direct operator experience.
FinalRound’s hacker house also converted one floor into a dance floor, with a co-founder who is a “full-time CTO, part-time DJ” providing the music. The space attracts about 800–900 founders, investors, engineers, researchers, and designers each month. The events deliberately avoid technical discussions: “Starting a company is already hard enough. If you’re going to hold an event, don’t talk about this stuff.”

10. San Francisco’s 996 Comes from Fear of Missing the Window, but Michael Admits It Must Deliver High Reward

Michael believes 996-style work in the US has clearly intensified in the AI era. Elon Musk, followed by return-to-office changes at Amazon, Meta, and other companies, reshaped the post-pandemic remote-work norm. FinalRound’s intensity is not driven by competitors working late; it comes from the fear of missing the window: “Everyone is capturing these opportunities, and if we don’t, we’ll regret it.”
When searching for an office, the team explicitly required 24-hour access, with somewhere inside or nearby to sleep, shower, and eat. In the early hacker house, the rule was “sleep when tired, work when awake.” The formal office still has beds and a shower, but Michael emphasizes that he does not want people “working late for the sake of working late.”
曼祺 pointed out that publicly praising extreme hours is generally unpopular in China, and early employees may have almost no life outside work. Michael’s answer is high risk, high reward: employees hold equity and share in the company’s breakthroughs and valuation growth. Traditional 996 is criticized precisely because employees make extraordinary contributions without receiving extraordinary returns.
曼祺 described Windsurf as Google absorbing the core team in a move that was effectively quoted, and questioned whether the remaining team’s options and shares could be realized. Michael criticized Windsurf’s conduct as “very dishonorable” and speculated that its founders may have burned out after years of pivots. FinalRound offers unlimited PTO; requests for 2 weeks off are approved immediately, and he actively encourages the team to take leave. His experience visiting 100+ countries during college has also convinced him that travel can generate strategic inspiration.

11. The Path from $3M to $10M ARR Came from Resystematizing the Entire Company

Growth after June 2024 could no longer be explained by isolated experiments. Michael’s list includes continuously launching products and features, improving global deployment stability, building a marketing team covering the US while understanding Europe and Asia, and simultaneously developing influencer, SEO, in-house creator, performance marketing, PR, and data-flywheel capabilities.
Michael says the financing also reflected a clearer mission and growth direction. The team was not only building an end-to-end career AI platform; it was also exploring B2B products and a moonshot project aimed at the future of hiring. The more complete growth infrastructure collectively created the momentum from $3M to $10M ARR.
In hiring, the company asks candidates: “Can you build an AI that replaces all of your early colleagues?” Answering yes is a reason to be hired. Michael is also thinking about how to replicate 10 or 20 versions of himself so human time can shift toward judgment and new opportunities.
Internal projects are led directly by domain experts rather than by one PM overseeing 10 projects at once. The marketing team has dedicated engineering support to agentize SEO, paid acquisition, and influencer work. Michael insists this should not be called merely AI automation, but “fully agent-driven.”

12. 4,000–5,000 Pages a Week Turn Real Interview Data into Search Distribution

With user permission, the SEO system organizes interview questions and answers from the platform into a forum and generates interactive pages targeting companies, roles, and professional terms. Someone searching “How does [company] interview?” may land directly on FinalRound. The system currently generates about 4,000–5,000 pages a week, a scale a manual team could not match.
The team has also built free tools for AI resume writing and recruiter communication, capturing interaction needs that Indeed and LinkedIn cannot currently satisfy. These tools generate traffic and collect user information, allowing later products to serve the same users more accurately.
SEO is no longer aimed only at ranking on Google, but also at appearing in answers from ChatGPT, Gemini, and similar systems. Michael sees substantial room for “growth hacking”: timeliness, focus on a single question, and whether a page directly completes the user’s task matter more than indiscriminate volume.
Some friends’ companies have subscribed to FinalRound’s internal marketing tools. They pay only “normal subscription fees,” which are not material to current revenue. The team values the external validation and refinement more, and is waiting for the right time to open the infrastructure to more founders.

13. Event Signals Become Content, Pages, and Campaigns Within 10 Minutes

Michael borrows an analogy from hedge funds: funds capture opportunities from event signals, and ordinary companies can also automatically detect events and generate actionable items. The key is not chasing a trend after the fact, but building an end-to-end system from signal detection through production and distribution.
His example is: “Microsoft lays off 20,000 people today.” The system immediately generates landing pages, podcasts, and social posts for the affected population, then automatically distributes them. The process from signal detection to executable content takes about 10 minutes. The “20,000 people” figure was Michael’s business example.
曼祺 asked why a large language model would retrieve this content. Michael’s answer was timeliness, long-tail coverage, and task relevance. CNBC may have a news page reporting the layoffs, but FinalRound’s page can directly help laid-off workers find jobs, so he says its related pages can typically achieve higher rankings.
More than a dozen friends’ companies are already using the system. Michael summarizes it as: “What a hedge fund can do, we can do too,” while acknowledging that its event granularity is not yet at the level of a real fund.

14. More Than 100 Influencers Are Managed as a Long-Term Channel, Not a One-Off Content Supply

FinalRound works with 100+ influencers globally, but the relationship is not “Here’s $1,000, shoot the video, and leave.” The team invites them to Bay Area events and rents large houses in Las Vegas, transforming them into immersive creative spaces filled with brand logos and swag.
The collaborations remain paid and include flights and lodging, adding a travel experience on top of the commercial partnership. About 5 influencers a week travel in from across the US. The local house manager oversees the property, manages the influencers, and tells them how to shoot.
Michael wants to do this at a lower cost than major brands spend hosting creators around F1 or sporting events, while making KOLs value the opportunity more. The branded environment and on-site support improve efficiency, replacing the model of finding a creator ad hoc and paying $5,000 for a single video.

15. Offline Ads Build Awareness; Online Acquisition Is Engineered Around 500 Creatives a Week

Offline campaigns follow an “all-out attack” approach. Ad lines include “interview confidently like our president,” with a photo of Trump, and “stop working for jokers, get a new job.” Placements cover San Francisco, South Bay, and East Bay commuter hubs, extending to major markets such as New York, Texas, and Chicago.
曼祺 asked whether billboards really work. Michael acknowledged that the team cannot precisely track whether an individual visited because of an ad; it treats billboards as top-funnel awareness and estimates impact using data such as fixed foot traffic. The specific budget and ROI are confidential, but his qualitative assessment is “greater than 1, and very positive,” or the spending would not continue.
Online acquisition is not simply a matter of spending more. Of 10 creatives, perhaps only 1 is excellent, and with only 2 there is almost no room for optimization. FinalRound tests about 500 different creatives every week, generated in bulk by AI and edited by humans, then A/B-tested across platforms, countries, time slots, and budgets, with funds automatically shifted toward the best-performing combinations.
Michael does not assume Google will optimize for the advertiser because the platform also wants customers to spend more. Compared with some Chinese teams that can field dozens of people for manual campaign management, the high cost of Silicon Valley talent forced FinalRound to build agents earlier. Human time can then move toward messaging, new channels, and higher-potential opportunities.

16. The B2B Product Uses “Virtual Work” to Rebuild Resumes and Traditional Interviews

Michael believes consumer products are breaking the old hiring process: AI Job Hunter can apply to thousands of jobs overnight, diluting the signal in resumes and traditional screening. The B2B answer is not another interview Copilot, but a programmable workspace where candidates prove their abilities through actual work.
FinalRound itself keeps only 1 simple interview round, then invites candidates to work together in the office for 1 to 5 days. Michael found that people with resumes that look excellent or terrible can perform in exactly the opposite way on the job. Real collaboration exposes ability, habits, and fit faster than a question-and-answer interview.
曼祺’s pushback was that this approach is difficult to scale: large companies cannot work alongside every candidate, and candidates cannot spend 2 or 3 days with every company. The productized answer is to have AI simulate a company’s first day or first week, with candidates interacting with different roles and completing tasks before the system generates evaluation signals. Depending on the customer, the roles can be all human or entirely AI-simulated.
The product is iterating with several large-enterprise design partners and plans to formally launch in September. Michael says the goal is to help organizations the size of JPMorgan, Google, and Microsoft run workshops systematically, viewing it as an opportunity to grow from $10M to $1B. The decisive variables include how quickly enterprises accept a new hiring model, which career category to enter first, and how aggressively the consumer product can evolve.

17. The Market Is Still Not Zero-Sum, and Talent Evaluation May Be Rewritten

Michael estimates that 40–50 companies in the US, China, India, and Europe are already imitating FinalRound. Cluely has the greatest visibility and is executing best; another team even swapped FinalRound’s red logo and entire design language for blue or purple and launched directly. He believes this kind of copying “affects the entire innovation ecosystem.” But the consumer market is still early, and examples such as Lovable, Bolt, and v0 lead him to conclude that it is not zero-sum yet.
He has not seen many comparable teams on the enterprise side, but says Anthropic and OpenAI are exploring related directions internally and offering them as a service to large enterprises; this is his account of an industry observation. If someone builds it faster and better, he would even be willing to partner, join, or acquire the company, because “if I’m the only one building it, that’s usually a little strange.”
Michael leans toward premium pricing. At $10 a month, a product may generate a lot of “vanity revenue”; at $1,000 a month, it is more likely to reveal real demand. The original examples did not specify the currency. For enterprise customers, 曼祺 proposed a base fee plus a hiring-outcome fee. Michael agreed and added that he prefers pricing around “If I help you hire 1 person, how much will you pay me?” The market will ultimately determine the price.
The real challenges are 3-fold: scaling from a small team to simulations of hundreds, thousands, or even enormous organizations; deciding which knowledge-worker category to enter first; and determining when the market will accept the new system. Michael’s latest view is that once AI lets everyone deliver outcomes, “the outcome is no longer important.” Talent value will shift toward the path taken, the tools selected, and the collaboration process, which is why the team continues recruiting AI hackers and reinforcement learning specialists globally.

20VC: GPT5: Sam Altman's Masterplan or a Gift to Anthropic Mutually Assured AI Malfunction [Dan Hendrycks]