Topics Pioneers Insight Method Author
Back

E258 | Inside HF0, Silicon Valley’s Startup “Monastery,” and OpenRouter’s Early Story

·硅谷101

Rendering the long image…

Topics Commercial Pitfalls of AI-Native Applications AI Inference Gateways and Intelligent Routing AI Training Data and Evaluation Benchmarks Unique Human Value in Everyday AI
Pioneers 郭靖Joe

Summary

  • HF0’s core offering is a small, residential program that takes care of the day-to-day so founders can focus on the single thing that matters most. It takes 10 companies per cohort and runs 3 cohorts a year; Joe’s fall 2024 cohort received $1M for 5% equity. He says the hardest part of “subtraction” is “cutting out even the second most important thing,” including big-company partnerships that seem worth maintaining but move slowly in practice.
  • Growth goals need to be clear, but they don’t all have to be revenue targets. HF0 asks each company to identify one core metric, then give a 2-minute weekly update and take 1 minute of questions: ACE Studio tracks revenue and ARR; OpenRouter tracks token calls. The point is to report progress against “the one goal,” not list development work, interviews, and partnerships.
  • ACE Studio’s most direct growth breakthrough came from turning sporadic marketing into a system that runs continuously. Joe spent months contacting more than 100 influencers, while Cal AI founder Blake Anderson was reaching out to 150 a day; the order-of-magnitude gap made Joe realize the system needed rebuilding. The team now contacts about 200 people a day, with about 40 replying. The principle behind it: “Build growth like you build a product.”
  • ACE Studio’s 10x revenue increase included a short-lived peak; Joe now puts more weight on the model and product experience. Monthly revenue rose from $80,000 to a peak of about $800,000, then averaged roughly $400,000–$500,000. A pricing A/B test once lifted revenue by about 30%, but he later saw the potential cost in lost users and word of mouth. Growth only resumed after ACE built its own foundation model in 2025 and launched ACE Studio 2.0 at year-end: “Tactics alone will hit a ceiling.”
  • HF0’s most transferable asset is an organizational system built around focused goals and public progress reviews. ACE has run 4 or 5 internal accelerators on top of remote work, bringing teams together for 3 months to focus on algorithms, development, or growth; every employee has a 3-month participation commitment every 12 months. Joe says rituals can spark self-motivation; Hongjun asks whether a clear revenue target can really be achieved through pressure alone. During the first half of that HF0 cohort, no teams were growing apart from OpenRouter and Fyxer.
  • OpenRouter’s first-mover advantage comes from a unified interface and a flywheel for integrating models, though price competition could still weaken it. Its unified API lowers the cost for developers to switch models, while usage generates rankings that encourage more models to integrate early. Joe calls this “surfing mode,” but acknowledges that with equal reliability, if one channel adds a 5% markup and another is 20% cheaper, he would “definitely switch to the cheaper one.”
  • ACE’s case against Suno rests on model development speed and post-training data from professional creators. With some users’ permission, editing histories feed professional musicians’ choices back into the model; Hongjun points out Suno is also building professional tools, which Joe acknowledges, emphasizing how hard ACE’s approach is to replicate. His competitive stance is blunt: “If AI music is a big thing, we’ll compete hard.”
  • AI music may surpass human vocal capabilities in individual cases, but its range of taste remains narrow. Joe says a model can combine high pitch, power, fullness, and rasp, yet still struggles to create rough, off-key work with raw vitality. “Ask it to make a bad song and it can’t” — excelling at pop-electronic music does not mean it can recreate 1970s jazz or a live bar performance recorded on a phone.

Deep dive

1. HF0 uses 10-company cohorts for an intensive accelerator

  • Hongjun compares the scale: HF0 takes 10 companies per cohort and runs 3 cohorts a year; YC takes about 200 per cohort and runs 4 cohorts a year. HF0 also favors repeat founders. Joe thinks founders with offers from both would choose HF0 “without hesitation,” though he adds up front: “I haven’t been to YC.”
  • Joe jokes that YC’s large-scale selection process is “spray and pray.” A residential program has a physical limit: earlier-stage investing requires a larger denominator, but there’s only so much room for teams to live and work together. HF0 fixes each cohort at 10 companies, so it has to be more selective; in his view, its acceptance rate is far below YC’s.
  • The investment terms have changed over time. Joe’s fall 2024 cohort received $1M for 5%; he recalls the previous cohort getting $625,000 for 3%, while Hongjun mentions an earlier offer of $250,000 for 2.5%. These terms correspond to different stages.

2. The monastery mindset means letting go of the second most important thing

  • Joe revises the idea that HF0 is about “material minimalism”: its keyword is “subtraction,” but the living arrangements are generous. A chef, laundry, a fitness coach, massages, and temperature-controlled beds are all there so founders “don’t have to worry about anything” except their business.
  • Hongjun mentions that Dave once spent time in a monastery and asks whether that inspired the “monastic mode.” Joe doesn’t recount Dave’s experience as a monk in Taiwan, and doesn’t know whether HF0’s first cohort was in Taiwan; he says he was also surprised when he first heard about it. What he felt directly was Dave’s ability to build an immersive environment.
  • The line that struck Joe most during the first week was: “The hardest part of subtraction is focusing only on the one thing that matters most to the company, even cutting out the second most important thing.” Founders already know they shouldn’t party every day. What’s hard to give up are opportunities that also look important.
  • His example is big-company partnerships: even when progress is slow, founders still feel they should maintain the relationship. Hongjun asks whether someone else could take it on. Joe pushes back: that person is also a company resource with their own most important task. “If you want to do everything and handle it all with perfect polish,” you won’t be doing the most important thing. “I think entrepreneurship is always about grabbing one corn cob and dropping another.”

3. Two-minute updates force companies to define real progress

  • Before the program starts, the teams go on a retreat in the mountains. They meditate, talk, imagine where they want to be in 3 months, then work backward to set weekly goals. Evan, the HF0 founder who leads day-to-day business discussions, helps each company identify one core metric.
  • At Monday’s Demo Dinner, each team gets 2 minutes to present and 1 minute for questions. The metric depends on the company: ACE tracks revenue and ARR; OpenRouter tracks weekly token calls. A team buying land in Canada to build a physical Web3 city tracks construction progress.
  • Joe says the update should be clear in a sentence or two: growth of 10%, a decline of 5%, or no growth. Building a feature or conducting interviews is an activity list. HF0 asks: “What progress did you make this week against your one goal?”

4. Influencer marketing goes from hit-or-miss to a system

  • At first, Joe thought influencer partnerships were unpredictable: an outreach might get no reply; a reply might not lead to a partnership; a partnership might not generate reach. He contacted more than 100 people over several months. Blake Anderson, who built Cal AI, told him he contacted 150 a day. Joe immediately realized, “We’re off by an order of magnitude,” which meant the system had to be broken somewhere.
  • Blake’s answer was: “Once you scale up something unpredictable by an order of magnitude, it becomes predictable.” Start by widening the outreach funnel, then improve reply rates, email templates, and conversion rates. That’s how partnerships can become a repeatable channel.
  • ACE has since iterated the system to its third or fourth version and now uses AI extensively. The team contacts about 200 people a day and gets about 40 replies. What began as sporadic outreach has become sustained work run by a dedicated team.
  • Around the same time, the team began scaling paid ads, including bidding, conversion funnels, and extensive A/B testing. Joe’s takeaway then: “Build growth like you build a product.” Growth, like product and algorithms, takes investment in building a system.

5. Growth tactics opened up new possibilities and revealed their limits

  • When Joe first encountered Silicon Valley’s growth playbook, he started attributing every problem to go-to-market. Hongjun asks whether his attention then had moved beyond product improvement. Joe says the product was also being updated, and growth during HF0 was tied to major features; he just thinks he later overcorrected in his focus on growth tactics.
  • Raising prices was his most extreme thought experiment: keep the product unchanged, increase the price 10x, and if the paid conversion rate doesn’t fall, revenue rises 10x. In practice, a pricing A/B test lifted revenue by about 30%. But he now thinks the move may also have cost ACE potential users and the word of mouth they could have generated. “Energy is conserved.”
  • ACE’s monthly revenue climbed from about $80,000 to a peak of about $800,000, but that peak was driven by a combination of tactics. It later averaged roughly $400,000–$500,000. Joe doesn’t treat the peak as a sustainable run rate.
  • He has since shifted his focus back to user value, especially model performance: “If you can deliver a SOTA model, growth gets easier.” Growth only resumed after ACE built its own foundation model in 2025 and integrated it into ACE Studio 2.0 at year-end. A growth system matters, but product experience remains central.

6. Public progress reviews create pressure, but breakthroughs may come late

  • HF0 emphasizes hard work, but the Demo Dinner doesn’t directly judge teams that aren’t growing. When strong people see others making progress, they ask themselves why they’re falling behind. Joe thinks the key is to “set up a ritual that makes people push hard,” so they trust the environment and engage more fully.
  • Hongjun’s challenge is concrete: setting an ordinary person a goal of making RMB100,000 in a month doesn’t mean pressure will get them there. Joe acknowledges that startups fail and miss expectations. “If the person isn’t good enough, replace them” is the direct answer, but it may also be time to change the method, change direction, or reconsider whether the goal is right.
  • The cohort started on September 18. In the first half, OpenRouter and only one other team, Fyxer, were growing; none of the others were. Dave even worried he might have misjudged the teams. On the last day of October, Dave said the cohort had reached a dividing line: some teams would “break out,” while others would just “hang out.” On November 1, ACE doubled its daily revenue, then grew another 50%, and doubled again.
  • Why did several teams turn at the same time? Joe says he doesn’t know; maybe the work they’d done earlier finally paid off. ACE had already built an influencer system and booked many creators. On November 1, one person posted a video and quoted a price; ACE also launched consumer-facing campaigns and took paid-growth actions. Several things combined to drive growth.

7. Internal accelerators embed focused work in a remote company

  • ACE began winding down its Beijing office in 2024. Based on its experience during the pandemic, the company decided remote work was more efficient and better suited to the geographic uncertainty of expanding overseas. After HF0, it added internal accelerators to its remote-work model; the first was even called “ACE0.”
  • The format is to bring the relevant engineers together for 3 months around one goal, borrowing HF0’s retreat, Demo Dinner, weekly updates, and daily support. For one session in Huizhou, ACE rented a more than 400-square-meter seaside villa with a pool to build its own music foundation model. Joe says the setting and sense of occasion helped people get into the work.
  • The algorithm accelerator in early 2025 produced the first model, which gained some traction after being open-sourced. ACE launched a second version at year-end and continued updating it internally. In the second half of 2025, the development team also held focused sessions by the Haihe River in Tianjin and in Tianjin Binhai, rebuilding ACE Studio as 2.0, a “free musical canvas.”
  • ACE has more than 30 people and has held 4 or 5 accelerators focused on algorithms, development, or growth; they can also run in parallel. Hongjun worries about employees being away from home for too long. Joe explains that every employee has a 3-month participation commitment every 12 months, which the team calls “military service.” There are 4 accelerators a year; that doesn’t mean everyone attends all 4.

8. OpenRouter started with 4 people and moved fast

  • Alex and Louis were in the fall 2023 cohort, exactly 12 months before Joe. Louis, who is Vietnamese, didn’t get a visa at the time and had to participate remotely. Once his visa came through in fall 2024, HF0 invited him back for another round. He was in residence with Joe but didn’t present at Demo Day because OpenRouter wasn’t raising money then.
  • When Joe met OpenRouter, it had 4 people, 2 of them part-time. ACE’s CTO was surprised: they seemed to have no test environment. They wrote code and “YOLO’d it straight to production,” then fixed bugs live. Louis would hold his laptop in his left hand and code with his right, writing while walking and eating.
  • Although everyone had a good room, Louis turned a temporary lounge in the basement workspace into his bedroom, filling it with blankets, pillows, and toothbrushes. ACE’s China team often worked until 5 a.m.; Louis was also regularly working until 2 or 3 a.m. Joe remembers the close quarters and the intensity.
  • When Louis reported token growth, he would say revenue was only “enough for two cups of coffee.” Hongjun mentions at the start of the episode that Stripe acquired OpenRouter for $7B. Joe says Evan posted the acquisition news in the alumni group and their cohort chat with the caption: “enough for coffee.”

9. A unified API creates a flywheel, but price can still change the choice

  • Joe says OpenRouter figured out that it wanted to be “the unified interface for LLMs.” It may initially have leaned toward a consumer-facing interface like Poe, but later saw greater demand from developers: every model switch meant changing API code and releasing a new version. A unified API removes that friction.
  • During the period Joe observed, the team prioritized integrating more models quickly, efficiently, and reliably. It didn’t focus much on fundraising or monetization; he says OpenRouter had raised money only from HF0 and possibly an earlier angel investor. These choices all served the goal of becoming the shared entry point.
  • Developers gathering on the platform encouraged new models to integrate early; usage then generated reports and rankings that attracted more models. Joe calls it a “usage benchmark.” Hongjun mentions that OpenRouter adds a 5% markup to model prices. Joe says it didn’t start out by buying low and reselling at a markup.
  • Hongjun says OpenRouter caught a wave of demand. Joe agrees, calling it “classic surfing mode”: “The bigger the wave, the higher they go.” But he acknowledges that a price war could undermine the advantage: if reliability is the same and one interface adds a 5% markup while another is 20% cheaper, he’d definitely switch to the cheaper one as a developer.

10. ACE is betting on catching up in models and proprietary editing data

  • Joe doesn’t sidestep the model competition with Suno: “We need to beat Suno on models as quickly as possible.” Professional tools and general music generation serve different product use cases, but both depend on music-generation models underneath. He compares the long-term opportunity to Mount Everest at 8,848 meters: Suno may be at 50 meters and ACE at 2 meters, but that doesn’t mean ACE can never catch up.
  • On the pace of progress, he says Suno may have trained models for 3 years, while ACE has spent 1 year and is already close to the level of Suno 5.5’s previous-generation model. He believes ACE is catching up quickly, while leaving room for uncertainty: “Maybe our technology is already better than Suno’s; we just need time to catch up.”
  • With some users’ permission, ACE collects editing histories for post-training. Joe says the professional product and its work with top creators such as Hans Zimmer create an environment that is hard to replicate. Hongjun points out that Suno is doing this too. Joe agrees, emphasizing the difficulty of copying ACE’s approach, not that Suno can’t enter the space.

11. Music expertise helps spot errors; user preferences shape taste

  • Joe says the team has “the best music-generation researchers in the world.” As an example, he says Suno’s model is also built on a representation foundation created by one of ACE’s researchers. He offers this as evidence of the team’s technical depth, while acknowledging that ACE still needs time to catch up.
  • Hongjun asks whether researchers’ preferences for classical, rock, or pop could determine the training data and the model’s style. Joe says no: dataset construction follows a rigorous methodology. Knowing music matters more for judging whether a caption correctly describes a style, instrument, or theremin sound, or is hallucinating.
  • What really shapes taste is human preference in post-training: what users like or dislike, and whom the platform chooses as a taste maker. Joe says you first need those users. ACE has professional creators such as Hans Zimmer, who can provide a source of feedback.

12. Surpassing vocal ability doesn’t mean covering every taste

  • Asked whether AI can surpass the best musicians, Joe begins, “I don’t know how to define that,” and says the public and the market ultimately decide whether a work is good. He prefers a narrower claim: individual songs “may already far surpass humans.” For example, a model can combine an unusually high, powerful voice with fullness and rasp, beyond what a human can sing.
  • That ability hasn’t solved the problem of a narrow distribution. Joe thinks a model can be stunning at pop-electronic music but still make 1970s jazz sound like electronic pop. Excellence in some directions doesn’t mean it can handle every style.
  • His clearest counterexample is: “Ask it to make a bad song and it can’t.” Rough, off-key work that sounds as if it were dubbed from a phone recording, yet has raw vitality, is still difficult to generate. He likes the songs Li Zhi performed in bars in Nanjing and that audience members recorded on their phones; later studio arrangements and production, he feels, “didn’t have the same feel.”
  • Joe also cites the moment when Shan Yichun stopped midway through “A Letter to Film Artists” on Sing! China. It was a mistake, but the emotion she showed made him like the song. He thinks imperfection can make a work: some people want perfection, others want imperfection, and the range of preferences AI music can cover today is still limited.