Open-Sourcing My Mistakes: Six Months of AI Entrepreneurship

Part 1: The Traps I Fell Into

Six months in. A lot of tuition paid. Fortunately, relative to our funding, the cost was manageable — but if we had been a YC company with a smaller runway, we would already be dead. I want to lay out my decisions honestly, open-source my mistakes. In hindsight, right and wrong are easy to assign. But I want to recreate the fog of the moment, because the hard part of startup decisions is never the retrospective attribution — it is the ambiguity of the present.

Decision 1: Go vertical, not general

Our first big call was to build a vertical agent, not a general one. The logic was clean: general AI is a big-company game, and a startup charging in has no advantage. I still believe this was rational. But if I am honest with myself, about 10% of that decision came from fear — fear of entering a hyper-competitive market.

That fear is insidious. It never appears in your consciousness as “I’m scared.” It disguises itself as reasonable analysis: “this direction isn’t differentiated enough,” “big companies will crush us here.” But underneath, it is just helping you avoid uncertainty.

Decision 2: Native Agent vs. Agentic Workflow

After going vertical, we faced a sourcing use case. Agent accuracy was low. The temptation was to build a hybrid — borrow from traditional software engineering, wire up a workflow, and get accuracy up. We compared the two approaches head-to-head: the workflow won on accuracy.

But I made a simple bet: choose the path that benefits from models getting stronger. As models improve, agents improve. A workflow’s ceiling is the flow you hard-coded. So we chose native agent.

In hindsight, this decision meant we could never serve enterprise customers well — they want predictable, stable, accurate. That is exactly where agent architecture is weakest. But at the time, this was not clear.

SaaS vs. Agent: two opposing philosophies

Traditional software is built on cybernetics — Norbert Wiener’s 1948 theory. A feedback loop: measure, compare, adjust. Same input, always the same output. SaaS is entropy reduction: a product manager talks to a hundred users, abstracts a workflow, locks the path down. Users follow the track.

Agents are the philosophical opposite. They are stochastic, not deterministic. The path is not pre-written — it emerges from the agent’s interaction with its environment. Every execution may differ.

SaaS gives you a track. Agents give you a wilderness. SaaS wins on certainty. Agents win on exploration.

But here is the truth: most people, if you take them off the track and drop them into a wilderness, will freeze. I am not dismissing SaaS — it is still a worthy game. I just do not find that game exciting.

Two types of user feedback

After shipping a GUI + LUI hybrid product, feedback split into two camps. The first wanted a clear workflow, more buttons, more dashboards — basically a SaaS. The second said: just tell me the goal and send me the result.

A pattern emerged: the first camp was operators and executors. The second was founders.

When I catalogued every feature request from the first camp, I realized I was echoing five-year-old SaaS products in our space. Those incumbents were asking how to add AI to their SaaS. We were asking how to add SaaS to our AI. Different starting points — converging features.

This observation was extremely painful, and took courage to admit: I had been repeating someone else’s path while telling myself I understood AI better.

So we made a pivot: we rejected most feature requests, blocked the button-and-dashboard demands, and focused on serving users who wanted autonomy over tooling.

The Mistakes I Must Own

1. Treating “theoretically correct” as safety, not hypothesis

I think this is a curse. Founders with strong analytical ability often build the worst products — because when a framework explains things elegantly, you mistake explanatory power for commercial traction. You end up satisfying your own vanity, not the user’s willingness to pay.

2. Underestimating the wedge product

A repeatable, fast-to-deliver, word-of-mouth single-point product — I severely underestimated this. Meituan started as a Groupon clone. ByteDance started as a news aggregator. They grew into super-applications not because the starting point was sexy, but because it was sharp enough to pierce a real need. I was too allergic to building anything that looked like a SaaS tool. The result: I never secured a starting point at all.

3. Projecting my personality onto the product

This is the most hidden trap. When I discussed product positioning, a huge amount of self-narrative was embedded in the argument. Why did I hate toolification? Why did I want the product to have a global view? How much of that came from user needs and how much from my own aesthetic preferences and identity? I found myself “preserving optionality” to avoid the pain of choosing. But startups demand the opposite: pick a direction, test it, accept the result, move on.

4. Seeing the endgame and refusing to start

The visionary’s disease: because you can see an opportunity logically disappearing within a year, you refuse to pursue it. But you underestimate distribution friction and trust. Users do not care that your product might theoretically be replaceable in twelve months. They have a pain point today, and they will pay today.

Similarly, many smart founders see market homogeneity and interpret it as “no opportunity.” But homogeneity means the opposite — everyone has spotted the need, but no one has found the deepest factor yet. Five thousand group-buying websites were homogeneous to the extreme. Meituan still emerged from that field.

Too little verification. Too much extrapolation. Too much love for narrative beauty. Too much fear of an unsexy starting point. The result: I never even secured a starting point.

Part 2: Why I Believe AI Startups Are Playing an Open Hand

On this visit to San Francisco, I finally went deep into the local startup ecosystem — hackathons, incubator events, Founders Inc.’s Artifact Festival. Walking the floor, I had one overwhelming feeling: I could have guessed every project here without leaving my apartment.

Marketing agents, sales agents, PM agents, browser agents, agents that cancel your appointments. A fatigue set in. Are we hitting an imagination ceiling? Is every card in this game face-up?

Lessons from the mobile era

I looked to the mobile internet era for answers, because it was strikingly similar — all cards were face-up there too. Yet from that ocean of obvious opportunities, ByteDance, Meituan, Instagram, and Snapchat emerged.

ByteDance survived because it found the first-principles factor: recommendation algorithms powered by a data flywheel. Users bring better content, not the other way around. By the time Baidu shipped its news feed in 2017 and Tencent launched Tiantian Kuaibao in 2015, ByteDance already had three to five years of behavioral data and model iteration. Late entrants could not buy back that time.

Meituan won the Thousand Groupon War not by outspending but by out-operating. Wang Xing used KFC locations to gauge a city’s spending power, Taobao indices for online buying habits, cinema counts for local activity. He only entered cities that met the threshold. Competitors blanketed hundreds of cities, burning cash in markets that generated no meaningful transactions. When the funding winter hit in late 2011, 96% died.

Instagram found the deepest factor in a crowded photo-sharing market: mobile-first simplicity. Four steps — shoot, filter, caption, share. It deliberately ignored the web, betting that phones would become the center of everything.

The winners did nothing new. News aggregation, group buying, photo sharing, instant messaging — all existing categories. They won by finding the deepest factor in each category and executing it beyond anyone’s reach.

But does this logic still hold in AI?

Honestly, the more I think about it, the more pessimistic I become.

The mobile era had one massive structural advantage that does not exist today: a billion-scale new-user increment from device adoption. AI has no equivalent. There is no new device category penetrating rapidly. All AI products are fighting over existing internet users — competing for attention from a fixed pool.

Worse, the foundational layer is feudal. Training a frontier model costs approaching billions. Foundation model companies are not neutral platforms like Apple — they build competing applications themselves. Your infrastructure provider is also your biggest potential competitor. Meanwhile, engineering capability has been fully democratized: anyone with Claude Code can ship a product in a weekend. Monopolized infrastructure, democratized applications. Startups get squeezed from both ends.

Want to build network effects? Scale largely follows capital — big companies always have more. Want a data flywheel? ByteDance succeeded because incumbents were three to five years late to react. Today, OpenAI sees a promising direction and ships in three months. Want technical moats? You use the same APIs as everyone else.

In this era, startups have no structural advantage. Almost nothing constitutes a moat — except taste and community.

When features can be pixel-perfect copied — Instagram Stories was a pixel-perfect copy of Snapchat — what keeps users is not the feature but the belonging. Facebook had two billion users and still could not replicate Snapchat’s place in Gen Z’s heart. In the AI age, the feeling that a product gets you may be the only card that cannot be cloned in a weekend.