Recently, while researching product cold starts, I ended up reading a lot of material on product competition. There were things I thought I understood before, but after reading them, I realized I only knew of them—I didn't truly understand them.
The Old Playbook Stopped Working
The old playbook for building and competing with products was simple: whoever had the bigger voice and more resources won. Go big, plaster ads everywhere, and users would naturally follow.
This approach started to crumble in the AI space around early last year.
On January 20, 2025, DeepSeek released the R1 model. Almost no one had heard of this company before. No brand recognition, no ad budget. They simply quietly uploaded the model to the open-source community and pushed a chat feature in their app.
Then it happened to coincide with Chinese New Year.
In about seven days, DeepSeek's user base exceeded 100 million. It topped the App Store charts in both China and the US, servers were overwhelmed, and at one point they even suggested users try other AI assistants. Yes, the servers couldn't handle the load—they were actively diverting traffic away. But DeepSeek didn't seem too bothered, just kept open-sourcing and updating.
The most representative comparison at the time was Doubao. Doubao is ByteDance's AI app, likely the biggest domestic investment in terms of resources, and it had long held the top spot for DAU among AI apps. Before it, Kimi had reportedly poured hundreds of millions into marketing for growth—investing in Bilibili, Douyin, and the education market. But as soon as Doubao made its move, Kimi's voice dropped significantly.
Then DeepSeek emerged, spent nothing, and surpassed Doubao outright.
You thought those who spent big were safe, then someone who spent nothing overtook you. That was a pretty big shock.
Iteration Speed That's Absurdly Fast
Later I went and checked the version numbers for Claude Code and Codex. I was a bit shocked.
Claude Code's latest version is 2.1.87. From its first release in February 2025 until now, it has shipped 365 versions total—averaging one every 1.1 days. If you open it today, there's a good chance you'll find something new.
Codex's latest stable version is 0.117.0. Since launching last April, it has released 132 stable versions, plus twenty to thirty alpha iterations per version—basically moving every single day.
These aren't small tools. Claude Code has over 10 million weekly npm downloads; Codex has over 3 million. At this scale, iteration speed is measured in days.
The lateral speed is even more exaggerated.
Last Thursday (March 27), DingTalk open-sourced workspace-cli, granularizing core features so AI agents could directly manipulate calendars, todos, and messages. The next day, Feishu (Lark) open-sourced larksuite/cli, covering 11 business scenarios and over 200 commands. One day after that, WeCom (WeChat Work)'s wecom-cli also appeared on GitHub.
Three companies, three days, almost exactly the same thing. When facing a valid direction, follow-up speed is measured in days. So-called "first-mover advantage"—in this environment, how long can it last? Maybe two to three weeks.
The Unscalable Way to Cold Start
While researching cold starts, I discovered a pretty counterintuitive pattern: many products that later became huge didn't get there by throwing money around early on. Instead, they relied on things that were completely unscalable.
In 2013, Paul Graham wrote an essay called "Do Things That Don't Scale". It had a huge impact in the startup world. The core idea is: don't think about scaling in the early days of a startup; use unscalable, manual methods first.
Sounds like motivational fluff until you read the case studies.
Stripe's two founders, Patrick and John Collison, when someone said "I could try your payment product," didn't send a link and let you figure it out. Instead, they said "Give me your laptop" and integrated Stripe into your code on the spot. Paul Graham thought this move was brilliant and named it the "Collison Installation", later teaching it in every Y Combinator batch. Airbnb did something similar. Early New York listings had terrible photos—the kind taken casually on phones. Brian Chesky and Joe Gebbia flew there, rented a camera, and went door-to-door helping landlords shoot professional photos. Listings with professional photos saw booking rates jump by 2.5x.
DoorDash was even more direct. The founder started out as a delivery rider himself, getting PDF menus from restaurants, using his own phone number as the customer service line, and riding out to deliver orders himself when someone placed one. Pinterest's Ben Silbermann noticed that many early users were design enthusiasts, so he went alone to offline meetups for design bloggers and recruited people one by one. Tinder took it even further. Whitney Wolfe went to college sorority parties to promote it, and the entry requirement was having Tinder installed on your phone.
These tactics share one thing in common: the founders got their hands dirty, serving users one by one, doing things that absolutely don't scale. But precisely because of this, that earliest cohort genuinely felt "this thing is different," and then spontaneously spread the word for you.
Think about it, and DeepSeek's path is actually the same thing. It wasn't relying on paid acquisition; the product itself was so good that people couldn't help but tell others.
Two Weeks of Lead Time, Then Start Over
Putting all of this together, my understanding of product competition has definitely changed.
In the past, you created distance through resources and channels—whoever had more ads and a louder brand captured more market share. That advantage could last a long time. Not anymore. The infrastructure for distribution is too developed now; something truly great can spread at almost zero cost. DeepSeek hit 100 million users in seven days—the fastest ever. The supply side is completely different too. It used to take months to build a feature; now Claude Code ships one every day, and if DingTalk open-sources something today, Feishu follows tomorrow.
Even if you do well, your lead time might only be two to three weeks. If you don't keep getting better during that window, users will switch without hesitation. That's exactly the state between Claude Code and Codex right now—one ships a feature, the other matches it, and users are switching between both.
The New Deciding Factor
So lately, I've been feeling less and less inclined to "watch the competition." I used to think this was basic product management—competitive analysis, differentiated positioning, finding whitespace. Now, in the AI race, I think spending too much time watching what others are doing is worse than spending time with your own users. The landscape you see today could change in two weeks.
What I think actually matters now is whether the product itself works, and whether users can feel that it's "genuinely different." DeepSeek spent nothing on promotion; Claude Code barely does any user acquisition. Yet users are flooding in. Then there's the relationship with early users—how to find the people who truly think this thing is cool and make them co-builders. Stripe's "Collison Installation" established exactly this kind of tight relationship between founders and users. In the early days, that relationship is far more valuable than ten thousand users from ads.
The AI era has indeed given product builders more choices and freedom—to focus more on the thing itself, rather than constantly competing over who has more resources or who can shout louder. Of course those still matter, but they're no longer overwhelming advantages.
At the end of the day, there's really just one question: Is what you're building creating value, or consuming other people's attention?
Once you figure that out, a lot of the confusion disappears.