I'm going to say something that sounds fringe today and will be obvious in a few years:
Vibe coding will just be coding
People laugh at "vibe coders" today. Insecure apps that get hacked. Launching on localhost:3000. Terrible UIs.
But pause: people with zero technical knowledge are shipping things.
The will to build is universal, and the tools will catch up. As they mature, their ceiling will rise past what hand built software can do along all dimensions.
At Create, we're betting on that future. We're building the tool that lets you "vibe code" your way to production apps.
Vibe coding
Andrej Karpathy said it best a few years ago post ChatGPT: "the hottest new programming language is English"
He then coined "vibe coding":
Two paths are converging:
- AI IDEs like Cursor. Designed for engineers. They raise the ceiling on what developers can do. But the floor is harsh - hence the jokes as regular people use them.
- Text to app platforms like Create. Designed for "builders". We slam down the floor. The open question on how far we can push the ceiling without dropping into an IDE. Low end disruption says once the toy clears production requirements, you never go back.
What's next
We're moving from keystrokes to intent. In a few years, "I built this" will mean "I described it, guided it, and shipped it"
Some things I'm thinking about:
- Production is the bar. Auth, data, payments, deploys, monitoring, analytics, app store approvals, scaling backends. This is what makes building software people use hard. The winning tools will help you push real end to end software, not mockups.
- Owning the runtime matters. Marcus and I often say - code only matters if it's run. There's infinite open source. But which of it is deployed in production. With what else? Code that's running, together, becomes its own signal you can reinforcement learn on. That doesn't change with infinite generation if you don't own a runtime millions of apps are built on.
- Nights and weekends. I read a stat that 67%+ of YC already use AI tools for MVPs. What the weirdos do on the nights and weekends, others will do in their day job in 10 years. Maybe the "idea guy with distribution" is the archetype founder in 2 years.
- Security isn’t a blocker, it’s missing tooling. Once threat modeling, dependency scanning, and runtime checks are native, the “AI-built apps are unsafe” argument ends. Plus, humans also write insecure code. It should be easier to guardrail AI if you own the runtime.
- Autonomy is climbing. Today’s tools feel like coding in English. Cool, but not exactly transformative - like automatic transmission in driving. But we're gradually pushing out the time coding agents can act on their own without supervision. Full self-driving for software means agents that reason, instrument, self-correct, and ship inside guardrails.
- Language needs more resolution. We only have plaintext prompts and raw code. Most aren't precise enough. Expect better UX to express intent with nuance.
- The math favors speed. Build costs are already dropping from ~$50K to <$5K. Time-to-MVP shrank from quarters to weeks. Velocity and distribution can be moats. Whoever iterates faster than the market can learn faster.
- Overheads still exist. There's still a fixed cost to product thinking. Once you find a great flow or interface, you amortize it across users. So there will always be room for people with taste and product judgement. AI accelerates them, but doesn't eliminate them. Copycats miss that people want new things.
- Taste wins. Infinite output will become trivial; software that feels intentional is rare. The best systems will ship human-quality design by default. Taste right now is held out as a type of last bastion of human skill. It feels very learnable by AI models, so it might abstract to the human choices amongst infinite AI choice.
- Feedback loops. There's a great talk from an Apple engineer who obsesses over making personal tools to increase his own feedback loop in development (because it increases the speed of his learning) Vibe coding is already doing this. You can just ship something to try it out. Scrap it if not. You might kick off 10,000 variations of something.
- Levels of abstraction. If the AI writes all the code, how do tools need to be redesigned to let the human jump to different levels of abstraction & still keep the mental model of the program in their head. Especially if they've never coded before. This feels unsolved and like a fun UX challenge.
If you find this interesting, come chat with me in SF.
