Summary

ACE Studio started at roughly $80K in monthly revenue, grew nearly 10x during HF0’s 12-week residency, and now generates about $800K a month, which 曲凯 says puts it in China’s $10M ARR tier. Joe does not attribute the result to the accelerator doing the business for the team. He compares HF0 to the performance-enhancing drug in Space Jam: “The performance-enhancing drug I gave you was actually just a glass of water.” What HF0 really provides is the conviction that a breakthrough is possible.
HF0’s core product is not a curriculum but a “container” that shuts down ordinary life and lets the team focus on a single goal. Every Monday’s Demo Dinner requires a public report of results; Tuesday’s Weekly Check-in calibrates the one KPI; and the surrounding teams’ rapid growth creates the feeling that “you’re either growing insanely fast or dying slowly.” Joe believes the residency may be “the only thing that matters.”
ACE Studio’s turning point in week 6 was not a model breakthrough but doubling its price, after which daily revenue doubled the next day. Joe explains that if revenue doubles after a price increase while everything else stays unchanged, conversion has not moved, meaning the original price was too low. 曲凯’s blunt conclusion was that the team had previously “done a terrible job.” Joe accepts that judgment in terms of the outcome, but says the real shift was treating pricing, growth, and product features as experimental components of the product itself: “You can maximize product value and still only be drawing one side of the square.”
The price increase was only the moment the team figured things out; subsequent paid-model experiments and industrialized customer acquisition turned growth into a system. The team tested 3 price points, eliminated trials and monthly plans, and Joe estimates that the combined changes produced at least a 3–4x impact. It also increased Influencer outreach to 100 people a day, measuring roughly 15% replies, 5%–10% partnerships, and 1%–2% genuinely effective collaborations, then put the best-performing videos into paid ads and reached 2x ROI on pure paid advertising.
“Data by Dinner” forced the team to compress validation into the same day, but not all of the infrastructure work from the first 5 weeks could be eliminated. Ad-account setup, tracking, conversion attribution, and the creative system still took time. Joe therefore believes that even if the team had possessed its week-6 insight on day 1, it might only have moved the initial takeoff from week 6 to week 4. The team then roughly doubled every 1.5 weeks, continued upgrading its Foundation Model, and used product-launch emails to reactivate lapsed paying users.
Joe sees the main obstacle for Chinese teams raising money from US investors not as simple discrimination but as the decision cost created by the interaction of entity structure, Cap Table, legal issues, and geopolitics. Investors have to explain why they took an investment that could go to zero because of structural risks. Unless you are “the next TikTok,” they are more likely to look at US projects first. The solution is not repeatedly asserting that the structure is compliant, but for founders to spend sustained time in San Francisco’s circles until they become “a person people know exists” rather than an unfamiliar variable.
Joe’s view of AI startups has shifted from “model updates may swamp the product” toward “there is still enormous room for productization,” with the moat built from engineer-days accumulated in the details. Video models can generate beautiful shots, but they cannot replace human expression and editing. Drawing on Perplexity’s “devil in the details” and the idea that “people always underestimate the rapper—it’s just a wrapper,” he summarizes the point in the simplest terms: “My moat is my engineer-days.”

Deep dive

1. An Email Seen a Week Late Led ACE Studio to HF0

When ACE Studio and Joe first spoke in May or June 2024, the company had just found PMF. Revenue briefly rose above $100K a month, then gradually fell back to roughly $80K because growth did not continue. More than 6 months later, monthly revenue had reached about $800K. 曲凯 accordingly said ACE Studio had entered China’s $10M ARR tier; Joe’s own wording was simply, “We make $800K a month.”
The turning point began with a sentence from PixAI founder Raven: “You should apply to HF0.” Joe did not even know what HF0 was at the time. After applying, he also failed to see the interview email for a week; only one slot remained that month, at around 3 a.m. Beijing time on the second day, and he booked it immediately.
After back-to-back online conversations with Evan and Dave, Dave asked him to fly to the US for an interview and warned him about the geopolitical risk: “Don’t IC, just come immediately,” concerned that adding more Chinese investors would make the Cap Table more complicated. Joe flew to San Francisco on day 4, interviewed in person on day 5, and received an offer that afternoon after speaking with the team: “It was just a plane ticket. I figured I’d go.”
English can indeed force Chinese founders to explain complex logic more simply, but Joe later realized that many obstacles have nothing to do with language. When 2 technically different companies discuss API integration, both sides can spend the first half-hour thinking the other side does not understand. Blaming every misunderstanding on English only erodes confidence. His advice is to “just push through it,” then go deeper until both sides reach a shared context.

2. HF0 Uses a 10-in-3,000 Filter to Find Teams Already Beyond the Accelerator Stage

Roughly 3,000 teams applied to that HF0 cohort, and only 10 were admitted. Joe infers the selection logic from the business model: with so few slots, the safest choices are either repeat founders or products whose data is already well beyond the normal accelerator stage. ACE Studio had roughly $100K in monthly revenue when it applied, which was itself a rare signal.
Joe acknowledges that YC is the world’s best-known accelerator, with its alumni, community, and steady stream of top companies providing powerful validation. But with 3 batches a year and roughly 100–200 teams per batch, it is closer to “spray and pray”: scatter the water first, then pray. Looking at how much concrete help a single team receives, he says he is “actually quite skeptical.”
His view of HF0 is the opposite. Another company in the cohort also grew 10x, while other teams grew 3x and 5x, so Joe believes the mechanism genuinely works. But effectiveness does not mean HF0 finds the direction, builds the team, or teaches founders how to run the business: “If you need an accelerator or an investor to teach you how to do business, you shouldn’t be doing that business.”

3. The “Container” Pauses Life, Then Turns the Impossible into Team Consensus

HF0 does not call itself an accelerator but a residency, emphasizing that it provides a “container.” For 12 weeks, ordinary life is almost completely shut down; time, space, and daily rules are rearranged. “Everything is distorted,” leaving only the task of finding and advancing the single most important thing for the company.
Before opening on September 18, the teams were taken into the mountains for several days of rituals, including meditation and carefully designed activities resembling corporate outdoor training. The point was not to teach growth tactics, but to ensure that members entered the house fully switched on and focused, instead of continuing to debate why they had moved to Silicon Valley.
Joe explains the mechanism through Space Jam: Jordan gives his teammates a performance-enhancing drink and their abilities surge, only for the movie to reveal that it was just water. HF0 cannot do the business for a company, but it can prove that something is possible. Just as later LLM builders did not need to steal code after the first LLM had appeared—the possibility had already been demonstrated.
This kind of conviction is difficult to reproduce on your own. If Joe rented a house in China and asked the team to work in isolation for 3 months, every question about cost, location, and weekly-reporting frequency would be reasonable, but each would gradually drain the energy. When a third party announces that “you have been selected by Silicon Valley’s toughest and best incubator,” everyone becomes willing to collectively “brainwash themselves.” His sharpest summary: “Reasonable things cannot create breakthroughs in this world.”

4. A Single KPI and Weekly Delivery Force Development to Translate into Business Results

Every Monday at 6 p.m., Demo Dinner required each team to explain in 2 minutes what result it had achieved over the previous week. HF0 did not want a list of features built; it wanted to hear something like, “We achieved 20% week-over-week growth this week.” The peer pressure of public reporting gradually became Joe’s “biggest source of motivation and biggest fear.”
During Tuesday’s Weekly Check-in, an HF0 partner would spend about 30 minutes with each team, compressing the Demo Day target into a simple KPI that could be tracked cumulatively. Most teams stopped tracking user count, retention, and feature count simultaneously and chose revenue instead. ACE Studio set a monthly revenue target of $500K, more than 5x its level at the time.
“Data by Dinner” pushed iteration to the limit: think of an idea in the morning, build it during the day, and see the data that evening. The goal was not to ensure every tactic worked, but to increase experiment throughput—trying 1,000 things quickly makes it more likely that one will work. “I’m building it; I can deliver next week” was no longer an excuse for having no result that week.
HF0 also arranged roughly 45-minute one-on-one external Office Hours rather than asking guests to give generic talks. Meals, workouts, massages, and Friday Family Dinner at the house were designed to remove day-to-day friction. Even lunch was deliberately low-GI to prevent post-meal drowsiness. Everything served the same goal: keeping attention on the company.

5. Almost No Growth in the First 5 Weeks; Conviction Survived on Public Pressure and Responsibility

One company entered the residency at roughly $20K in monthly revenue and then grew 30%–40% every week, catching ACE Studio by weeks 4 or 5. ACE Studio, by contrast, rose only from just over $80K to just over $90K and then just over $100K, growing 1%–2% a week. Joe still believed it would eventually take off, but began retreating to the hope that it might “just start gaining traction by the end of the 12 weeks.”
The pressure also carried the burden of representing a Chinese team. Joe wanted to prove that Dave had not made the wrong choice and did not want HF0 to conclude that it should stop recruiting Chinese founders. When ACE Studio stalled, he even hoped the other 2 Chinese teams would grow first: “At least that would prove Chinese teams aren’t the problem; we’re the problem.” But none of the 3 teams showed significant growth in the first 5 weeks.
ACE Studio had 6 people on site. Joe considered his team the hardest-working one, with its bedtime moving from 2 a.m. to 3, 4, and 5 a.m., then even 7 a.m., eventually reverting to China time. Other teams were also working roughly 80 hours a week.
This was not an independently sustainable management model, but a deliberately induced state for a limited period. With every company around them rising, the team found it increasingly difficult to explain stagnation through insufficient resources. “They don’t have an extra brain, or 24 extra hours a day,” so the only possible conclusion was that ACE Studio’s method was still wrong.

6. The Week-6 Breakthrough Came from Exhaustive Search; the Highest-Probability First Move Was Simply Raising Prices

On October 31, with the program more than halfway through, Dave used the line “some teams break out, some teams hang out” to distinguish companies beginning to break through from those settling into mediocrity. Joe wanted to replace revenue as the KPI, but Dave refused: changing the metric meant “you compromised; you got scared.”
A Quora co-founder had previously shared a case after Demo Dinner: when Quora hit a growth bottleneck, it listed 20 possible solutions and tested them in order of probability, only breaking out again with the 11th. Joe did not retain the now-obsolete tactic itself, but the broader lesson: most people try 5 or 7 things before declaring that growth is impossible, without ever exhausting the search space.
That night, the 4 core members walked through a San Francisco park, listing every possible move and choosing the first one by probability. When they opened their eyes the next morning, daily revenue had doubled. The tactic had no dramatic technical sophistication: “It was simple. You just raise the price, and you raise it by 2x.”
曲凯’s rebuttal was direct: if the highest-probability move worked immediately, the team had previously done a terrible job. Joe accepts that conclusion on the results, but says the 2 situations were not operating in the same coordinate system. The mental model had shifted. Previously, adding 10 singers better suited to Western tastes took 1–2 weeks to develop, followed by another 1–2 weeks to find Influencers and promote them; doubling the next day was impossible. Now growth and product value were 2 sides of the same square, jointly determining its “value area.”

7. Pricing Is Not a Backend Parameter but Part of the Product and Business Model

Several Office Hour guests opened by asking Joe about his pressing model, prompting the “domestic product manager” Joe to realize that he had previously spent more time analyzing human behavior and social relationships, while pricing amounted to copying the market. Another trigger was more concrete: an annual membership for shared bikes in San Francisco cost more than ACE Studio’s professional SaaS subscription.
Doubling the price was not a permanent answer. The team subsequently tested 3 price points before finding its sweet spot. The key principle was that the absence of a universal standard does not mean something cannot be optimized. Price, payment structure, growth systems, and user-experience features all belong to the product and can all be tested.
The team also eliminated trials because the US market contains large numbers of users paying with fake cards to get free access. It then eliminated monthly plans, though Joe did not elaborate on the specific reason. He estimates that the various pricing and payment-model changes together contributed at least a 3–4x impact, confirming that the initial price had been materially too low.

8. Once Influencer Acquisition Was Industrialized, Content and Paid Ads Became the Second Growth Pillar

ACE Studio originally contacted only dozens to a few hundred Influencers a month and treated one-by-one outreach as unsustainable manual work. Blake Anderson, who specializes in Influencer marketing, asked: “How many Influencers can you contact in a week?” The correct order of magnitude, he said, was 100 a day.
High-frequency outreach made the funnel measurable for the first time: roughly 15% replied, 5%–10% ultimately partnered, and about 1%–2% were genuinely effective. That small group was enough to cover the entire cost. The insight was not which script Blake used, but that only at this scale could the team legitimately ask whether the channel worked.
The operations team built and repeatedly iterated an SOP, eventually maintaining roughly 20 email templates tailored to different audiences, including Black and Latino communities, while measuring each template’s conversion rate. Growth stopped meaning “everyone reach out a little more” and became a production system with input volume, segmentation, and feedback data.
Among the multiple Influencer videos produced each week, some consistently emerged as high performers. The team then put the winners into paid ads and reached 2x ROI on pure paid advertising. At the same time, the Foundation Model and breakthrough features continued to improve; every update triggered an email to former paying users who had churned, making revenue “sprout another spike.”

9. The Accelerator Habit Did Not Persist, but the Team Began Reproducing Extreme States in Cycles

3 months after leaving HF0, the habit of delivering at the weekly limit had not survived. Joe’s candid answer was, “It didn’t,” because the process had been too exhausting. But the founding team’s way of thinking had clearly changed: different stages of a business require different tempos, and extreme acceleration is suitable for periodic use rather than year-round operation.
In the future, they may run 1 R&D acceleration cycle and 1 growth acceleration cycle each year. The former would spend 3 months asking only whether the team could deliver the next generation of revolutionary technology; the latter would ignore everything except whether revenue was rising. They retained HF0’s subtractive discipline while acknowledging that a company cannot operate at residency intensity every day.
Their next experiment is to rent a large seaside house in Huizhou and have the core R&D members live and work together for 3 months, targeting the next-generation AI singing-voice model and their own music model. Choosing the seaside rather than a cheaper suburban villa is meant to make members feel, the moment they open the door, that “this place is different”—creating the physical conditions for conviction first.
Joe believes peer pressure is merely one objective effect of HF0. More fundamentally, the team could always sense the distance to its goal and had to produce a visible result every week. Burning Man and the Bay Area work similarly: entering a different set of time, language, food, and rules makes people believe they have arrived in “another world,” where magic can happen.

10. US Fundraising Pursues “Fewer Than 3” Points; Chinese Teams Must First Digest Structural Friction

At HF0 Demo Day, each company had only 2 minutes, so the main presentations for 10 projects took about 20 minutes. But the teams spent an entire week preparing those 2 minutes, receiving daily instruction, practice, and feedback. The objective was not to explain every detail of the business, but to make investors remember fewer than 3 highlights and agree to a follow-up meeting.
Joe observes that the US fundraising process is highly standardized: a 2-minute pitch is typically followed by a roughly 30-minute meeting, in which the investor introduces themselves in 2 sentences, the founder immediately delivers the key points, and the meeting ends after 2 or 3 questions. Investors still care about the ceiling, team, founder motivation, competition, and whether a large company could replace the startup—the questions are broadly the same as in China. The difference is that the answers must be simpler and sharper.
To describe the friction facing Chinese teams, Joe compares it to “an entrepreneur from Papua New Guinea raising money from Chinese investors.” An unfamiliar registered entity, a Cap Table made up of local investors, non-fluent language, and trade conflict on top of that all prompt the institution to ask, “Why should I look at your project?” Unless you are “the next TikTok,” the cost of verifying legal structure, long-term residency, and geopolitical risk can be enough to shelve the deal.
Investors can accept losing money after making a reasonable decision, but they struggle to explain to LPs that they lost capital because they failed to investigate structural risks—a mistake that could even affect their careers. Joe therefore emphasizes that founders need to go to San Francisco and keep appearing at its parties and in its communities. Only after repeated meetings will investors confirm that “there is this person, and there is this project,” lowering the cost of trust.

11. AI Models Are No Longer the Product; Engineer-Days Accumulated in the Details Are the Moat

Joe’s view of AI productization has shifted directionally. He once remained uncertain about whether model updates would swamp applications and make product work meaningless; now he thinks that view is “increasingly wrong.” Models still matter, but the final opportunity will return to simple, obvious, labor-intensive productization.
Early video models inspired visions of replacing video and even film. In practice, they are closer to tools that generate beautiful shots. What users actually consume is the creator’s expression behind the work, which still requires a powerful interface to organize and finish the content. AI has solved much of the generation layer, but the final product is still “a next-generation video-editing tool,” not a standalone model experience.
Joe sees Cursor and Perplexity as representative productized companies and strongly identifies with the “devil in the details” line from Perplexity’s Pitch Deck, as well as the idea that “people always underestimate the rapper—it’s just a wrapper.” When investors repeatedly ask about his moat, his answer is: “My moat is my engineer-days”—large amounts of non-revolutionary work that make the experience incrementally better, with enough secret sauce that a new entrant would need another 3–5 years to rebuild it.
On management, 曲凯 brought up Silicon Valley’s recent founder mode, Airbnb’s founder’s early practices, and the failure of the professional-manager system. Joe added that Silicon Valley is once again paying attention to micro-management, which he remembered being taboo when he studied management. 曲凯 then cited Elon Musk’s nano-management and his consecutive late nights with HR after acquiring Twitter, arguing that the key is still to identify the crux of the problem and solve it. 曲凯 had previously compared these basic elements to sunlight, air, and water.

US-China Relations: History, Culture, and AI Competition, with Noah Smith, from Econ 102 Epstein Files Flop, State of the Market, Autonomous Robots, Trump's Gold Card, Friedberg on Jeopardy