H. Jun Huh

Hobby Capital

July 7, 2026

Translated by Claude Opus 4.8

1. In the AI era, knowledge work loses its value while domain knowledge becomes what matters.

  1. Knowledge gets produced fast and consumed fast. We now live in an age where AI automatically collects and summarizes primary data, making information delivery trivial. AI creates content and AI reads content. This age arrived a long time ago.
  2. Agentic AI is replacing not just knowledge but the actions built on top of it. AI can interpret medical knowledge, legal knowledge, and even physical domains, and it can hand you a set of instructions to act on.
  3. What AI doesn’t have is domain knowledge. Every industry (finance, hobbies, medicine, you name it) has high-context areas. The raw data AI gathers on its own doesn’t explain them.
  4. Agentic AI that rides on top of domain knowledge can actually run useful work.1 If you get AI to collect high-context information and build an action workflow around it, you end up doing things that neither an ordinary person nor an ordinary AI could pull off.
  5. You have to go after human-dependent fields. The go-to examples are Karrot (physical relationships), Hervey (licensing barriers), and Anduril (defense). In these fields, an AI workflow creates a compounding structure.
  6. So does each individual person’s data matter? No. Personal context data isn’t a moat. Information about a single person’s life, their behavioral patterns, and so on migrates easily. When a new service comes along and you want to switch from ChatGPT to another LLM, you just ask GPT for the information it has about you. Personal context data is easy to extract.2
  7. You have to separate declarative knowledge from tacit knowledge. Take pour-over coffee. A preference like “I like coffee with acidity” is verbalized, so it moves easily between LLMs, but the feel for reading how much the grounds bloom when you pour water and adjusting the thickness and speed of your next pour accordingly can’t be verbalized at all, which makes it impossible to extract.

2. Hobbies are the domain knowledge that’s about to take off.

  1. A hobby is an activity you have to acquire through your own body, where the point is the process itself rather than the result, and where value only arises when the person doing it is you. That’s why most hobbies live in the physical realm. And the physical realm isn’t something AI can easily replicate.3
    1. Sports, instruments, exhibitions, being a barista, going to shows, wine: most of the hobby fields we know of are physical.
    2. With work, the result is the point, so you can swap out who does it and the value holds. With a hobby, the process and the experience of “I do it myself” are the point, so the moment you swap out who does it, the value disappears. Even if a robot finished a hike for you, it would mean nothing. In other words, humanoids don’t replace hobbies. They take away your work and give you back your hobby time.
  2. Hobbies have depth. The knowledge you’ve acquired through your own body becomes the barrier to entry. It functions as a kind of self-satisfaction. Like the “easy to learn, hard to master” structure in game design, the easier you make it to get started, the more people flow in, and the pyramid of mastery above them actually grows larger. The reason lowering the barrier to entry (which I’ll get to below) is itself a business opportunity also comes from “easy to learn.”
  3. A hobby isn’t knowledge work. The goal isn’t productivity, it’s pure fun. And that feeling is something AI can’t deliver.
  4. Hobbies come with identity and community. The core of it is the fun of spreading a hobby and talking about it with other people.

3. Going forward, the goal is to lower the barrier to entry for hobbies.

  1. A clear example: over the past few years, the share of live classical performances has gone up.4 People increasingly want and enjoy high-context hobbies of a serious caliber. Even so, you need to understand why classical music is hard, and there are three big reasons.5
    1. You have to understand the history.
      1. You have to know the historical backdrop against which the composer wrote the piece, who the composer is and what kind of life they lived, who’s conducting it now, what the world’s famous orchestras are (the Vienna Philharmonic, the Berlin Philharmonic, the Royal Concertgebouw Orchestra), and which recordings are the great ones. This isn’t about memorizing facts. You have to approach it as if you’re living in that era yourself.
    2. You have to understand form and theory.
      1. You have to understand the basic movement structures of classical music: sonata, ternary form, scherzo, rondo, and so on. There’s a fun in matching the form as you listen and understanding it, and it helps you catch the hidden meaning in a piece. On top of that, you have to know the multi-movement forms like the symphony, the concerto, and chamber music.
    3. You have to enjoy the music itself.
      1. This part is a hobby from the very start. Even if you know nothing about the history or the form, if you feel a rush the first time you hear it, that’s where you begin.
  2. Outsourcing knowledge work to AI and the arrival of humanoids give time back to humanity.
    1. Why is it that even after humanity got its hands on computers, we ended up working more than the humans before us? It starts from the limitation that the one doing the work is, in the end, still a human. Humanoids and AI are the first case of moving who does the work off of humans.
    2. You don’t have to reach far for examples. History already has cases of time being given back. In the Roman era, slaves freed Roman citizens from labor, and in the 17th century, butlers and servants freed aristocrats from housework, cleaning, and laundry. Of course the amount of capital poured into them was considerable, so only aristocrats could afford it, but going forward, humanoids and AI will create the slaves and butlers of a new era. And since they run on nothing but electricity, I expect the payoff relative to the upfront cost to be high.
    3. Part of AI’s productivity gains gets redistributed into leisure spending.
      1. It’s true that the evidence on how AI’s productivity gains get allocated across time is mixed. There’s research showing that at the stage where AI is a “complement” to labor, working hours actually go up (NBER, +3.15 hours per week).6
      2. But in personal-life domains and in jobs where output is fixed, the time saved doesn’t get clawed back into production and instead leaks into leisure (Stanford’s study of 200,000 households,7 the Bank of Korea’s survey8). In other words, the key to time coming back is the “shift in who does the work” from 3.2.1, and the inflection point is the moment AI crosses over from complement to substitute (agents, humanoids).
      3. The demand side, meanwhile, is already proven. Leisure’s share of spending has climbed from 9.5% to 13% over ten years (Visa),9 and the shift toward experience spending is a structural trend (McKinsey).10 How fast the time comes back is up for debate, but the direction (that the returned time and income flow into leisure) is backed by the data.
  3. Hobbies are expensive. And demand keeps growing.
    1. Taste has been capital for a long time now. Even for the same service, people pick the product that fits their taste, and people’s tastes diversified ages ago. The waves of customization, interior design, and minimalism prove it. Now taste is being elevated to the level of a hobby. Taste is a preference; a hobby is an activity.
    2. Even so, good taste is expensive. Ivan Zhao, the CEO of Notion, has said he’ll hire11 people with good taste. Good taste is still an expensive area that takes study. Now, AI has removed the floor. Everyone puts out something halfway decent. Conversely, the ceiling has become infinite. Customers who know what makes a good product have gotten more sophisticated. What separates a good product from a mediocre one is taste.
    3. Good taste comes out of a range of hobbies. You have to know people’s hobbies (what they like and enjoy doing) to sharpen your eye for good taste. You need hobby knowledge you’ve experienced and made yourself, not just something you’ve watched and understood conceptually, for good taste to come out.
    4. The more AI lets anyone consume content and produce mediocre stuff, the scarcer body-and-time-intensive, acquired hobbies become, and scarce things become markers of status and identity.

4. And here’s where it actually became a business.

In the end, hobbies become a business in three ways. Community becomes the moat, or the habit becomes recurring revenue, or the domain knowledge becomes an AI workflow.

  1. Strava, where the hobby activity itself is the community hub.12 Even if users train, find routes, and track their biometrics on other apps, in the end they upload to Strava for the recognition of their records and to compete on segments. The data and features are scattered across many apps, but its position as “the hub where the people who run together gather” is what counts. Direct evidence for the proposition that the moat is data bound to a community.
  2. Chess.com, where a hobby’s habit-forming nature equals recurring revenue.13 Its AI coach aims to go beyond a plain algorithmic evaluation and explain why, with explanations personalized to the user’s particular playing style and their recurring mistakes. It gave people who take up chess as a hobby the possibility of getting better, and that turns into monthly recurring revenue. That said, passing on top-tier skill is still the domain of a human coach,14 and this is exactly why the mastery pyramid of “hard to master” from 2.2 holds up.
  3. GOATY, where a hobby’s domain knowledge helps improve an AI workflow.15 While most golf apps stop at measuring, GOATY runs a loop that diagnoses the root cause and tracks whether it worked. Its structure is to collect the user’s high-context information and propose it as an action workflow. It’s the first form of “collect high-context information -> action workflow” from 1.4, implemented in the realm of a physical hobby.
  4. Garmin, where the same logic works even in hardware.16 In an age where AI commoditizes the software layer, whoever owns the hardware data layer (the sensors) is structurally protected.

In the end, recurring revenue comes from a hobby’s habit-forming nature, and the moat comes from data bound to a community. If hobby businesses up to now have stopped at removing barriers to entry, the next stage is an AI workflow that assists the hobby, or (once the humanoid era arrives) the seat that assists the hobby in the physical realm.

Having a hobby matters.

Understanding hobbies means understanding the roots of human behavior. The fact that people act toward the goal of a hobby without any financial motivation is grounds for calling it an open game with diversified goals. If you can touch on people’s different desires and tastes somewhere in there, you start to see business opportunities.

Lately I’ve been brewing pour-over coffee every morning. I’m still shaky on all of it, from telling apart the differences between Colombian and Ethiopian beans and which beans to pick for good flavor, to distinguishing sweet, sour, and bitter, to how flavor changes with grind size, to the technique of pouring, to water temperature, and on and on. There’s a mountain of knowledge to learn, and since the history and world of coffee run deep, I’m studying as much as I can. And in the middle of all this, the number of products you’re supposed to buy. At the points where I felt something was inconvenient while doing pour-over, products that solve exactly that already existed. So what about the software side? You could easily build plenty of apps that help you brew pour-over coffee. When everyone (every person, every company) is focused on AI and productivity apps, I think there’s an opportunity in starting from the nichest field of all, hobbies.

These opportunities are the kind you only find by doing it yourself. Finding the friction in a hobby takes hands-on experience. If you only read about it in books and dig in purely on logical consistency, the problem definition gets overly complicated and the solution turns bizarre. That’s how you end up building a product nobody uses. On top of that, the process of solving problems while learning a hobby is fundamentally simple. It’s simple because you have the solution you want and you can weave it into a product.

In this piece I’ve interpreted hobbies at a macro level. Pinning down exactly what a hobby is risks getting in the way of the independent interpretation that comes from ambiguity. Think about what hobby you enjoy, and maybe somewhere in there you’ll spot an opportunity.

One more thing: when you read this, I’d suggest doing it not from the present moment but from a vantage point five years out, at the moment humanoids are just becoming as widely available as cars. Or imagine it from the moment AR glasses go mainstream two or three years from now. If you take that and connect it to the thrust of this piece, I think you’ll be able to “imagine” what hobby capital is.

Footnote

  1. https://github.com/nousresearch/hermes-agent ↩

  2. Ardent Venture Partners argues that basic memory, workflows, and agent orchestration are now table stakes, and that the moat that survives belongs to applications that encode the undocumented, tacit rules of a domain (Feb 2026). In other words, memory isn’t the moat; domain knowledge is. https://ardent.vc/blog-posts/the-moat-just-moved-areas-of-opportunity-in-ai-native-software-d34b7 ↩

  3. This is the combination of consumer business types 3 and 4 in The 4 Key Skills Entrepreneurs Must Focus On. ↩

  4. Per KOPIS, the number of classical performances in 2025 (+3.3%) and the number of shows (+9.6%) are on an upward trend. https://www.kopis.or.kr (Ministry of Culture, Sports and Tourism / KAMS, 2025 Performance Market Ticket Sales Analysis Report) ↩

  5. I used classical music as the example because the approach here applies to ordinary hobbies too. Knowing the background, understanding the concepts, and enjoying the subject are the core of every hobby. ↩

  6. NBER Working Paper “AI and the Extended Workday” ↩

  7. Stanford SIEPR, analysis of browsing data from 200,000 U.S. households. ↩

  8. Bank of Korea BOK Issue Note “Does Adopting AI Raise Productivity? An Analysis of the Effects Over the First Three Years” (June 2026) ↩

  9. Visa Business and Economic Insights “For the fun of it: The evolution of leisure spending” ↩

  10. McKinsey “Are you experienced?” ↩

  11. Ivan Zhao, “The Refounder” (Sequoia Capital Podcast). Zhao defines talent as “capability × taste × agency,” and says that because LLMs have leveled out capability, Notion optimizes its hiring for taste and agency. https://sequoiacap.com/podcast/notions-ivan-zhao-the-refounder/ ↩

  12. Strava filed for an IPO in early 2026, with a valuation discussed at roughly $2.2–3 billion. It has 150M+ users, but its paid conversion rate sits at just 10–15%, and competition with Garmin, Whoop, and Oura is flagged as a risk. ↩

  13. AI chess coaches like Aimchess, Chessvia, and Chess.com’s Coach Chester run on a $12–15/month subscription model, showing how a hobby AI turns into recurring revenue. ↩

  14. A strong human coach watching how the user thinks, understanding their psychological tendencies, and giving real-time feedback is still judged to have better improvement efficiency than any AI subscription. This is the same layer as “tacit knowledge can’t be extracted” in 1.7. ↩

  15. Across 1,896 members of GOATCode.ai (GOATY), the average GOAT Score improvement is +29, with roughly half improving by 5 or more. Most competing apps stop at measuring. ↩

  16. Garmin’s defensive moat is a biometric stack (sleep, HRV, Body Battery, and so on) that only works on its own hardware. In an age where AI commoditizes the software layer, whoever owns the physical data layer of the sensors is structurally protected. ↩

← Back