If You’re Not Prototyping with AI, You’re Going Too Slow

Last week, I attended YC’s W25 demo day, where 160 startups presented to a packed room of the world’s top investors. At the start of the program, Y Combinator CEO Garry Tan shared an astonishing fact that he had posted on Twitter/X a few days earlier,

If you haven’t heard the term “vibe coding” before, then you’re to be forgiven. It was only coined 6 weeks ago.

YC Demo Day was just the latest in a string of events that has quickly brought vibe coding to the forefront of the tech world. The week before, Clockwork Labs unveiled Spacetime DB, a new relational database with an integrated back-end server.

So what?

By creating a system that can seamlessly execute application logic inside the database, Spacetime DB unlocked the ability for code — and, thus, AI-generated code — to fully deliver an application back end.

<Side Quest> As a database geek, I went down the rabbit hole a bit on this one. Spacetime DB isn’t a complete replacement for existing relational databases (for example, it doesn’t support complex analytics, like those I helped build at Aster Data). Rather, it is purpose-built around the subset of high speed / low latency features required to support real-time applications. You can think of it as the natural successor to MongoDB, which was originally positioned as a turn-key back end database for web applications. Pretty cool, if you ask me. </Side Quest>

Clockwork Labs is specifically targeting online games as the initial market for Spacetime DB, which makes a ton of sense. But what does that have to do with vibe coding? Well, if you’ve spent anytime on “startup Twitter/X” in the past few weeks, you’ve likely come across an increasing number of posts from people showing off simple games that they built using AI.

That gives us YC startups and amateur game developers as two groups that are early adopters of vibe coding. And what do they both have in common?

They’re both prototyping.

Prototypers are the ICP for Vibe Coding

One of the most important early tasks for founders is to get from initial concept to prototype as quickly as possible, so that they can validate their core hypothesis. Until you can get a real-life implementation of “the idea” in the hands of potential users/customers, it’s impossible to really know if you’re on to something (which is why The Mom Test is required reading for all founders).

When I was a founder, the journey from idea to prototype was measured in months or even years. The emergence of AWS significantly reduced both the time and cost of developing new software, but it still took most startups several months to go from concept to prototype. “Clickable prototyping” tools, like Figma, provided an incremental improvement, but they mostly ended up as tools for non-technical founders and settled into an ICP in larger, more mature companies.

This early prototype of DataHero took more than 6 months to build

The arrival of AI and vibe coding is once again changing the game.

Even in the early innings, LLM-based coding has drastically reduced the journey from idea to a functional prototype for many startups (something I predicted last year). And we saw this on full display at YC Demo Day.

Although I don’t know this for certain, I have a strong suspicion that the 40 or so YC companies that had generated 95% of their code using LLMs were almost all still at their prototyping / early user feedback stage (a strong hint was the frequency with which certain founders proudly shared that, “we only started writing code X weeks ago…” in their pitch). That’s not to lessen the impact of AI-based development. Rather, it’s to emphasize that the initial PMF of vibe coding very clearly falls within the early prototyping stage. And I suspect it’s going to stay that way for the foreseeable future.

As amazing as AI-based coding tools are at accelerating developers, they’re nowhere close to a point where they can develop complex, production-quality applications on their own (I’ll dive into this assertion more in an upcoming post). For now, I’ll simply point to the current state of AI image generators as a proxy:

So…which one is facing backwards? (Also, I asked for founders…not McKinsey consultants.)

I often share that one of my litmus tests when meeting new founders is the question, “Can you beat my friends?” Well, my friends are all now building prototypes using AI. So if you’re a software founder and you’re not already leveraging these tools, stop what you’re doing and spend a week trying them out.

Because if you’re not prototyping with AI, you’re going too slow.