Insights

Steve Blank is Right (About MVPs and Vibe Coding)

By Saad Ahmed, Founder of Cosmic · September 24, 2026

Steve Blank's Stanford class found that AI lets founders build a finished-looking product on day one, before they've learned what customers need. His teams pivoted later, mistook polish for product/market fit and ended up with what he calls learning debt. Cosmic's answer is to slow the start down: discovery and framing on day 1, fast building after that, and real users testing the build twice.

Key takeaways

AI makes building fast and cheap, which moves the bottleneck to judgment about what to build and for whom.
A polished product on day one is what Blank calls evidence theater: it feels like validation without any customer contact.
Teams that start with a finished product pivot later and learn less, which Blank calls learning debt.
Go slow to go fast: settle the narrative, information architecture and users before letting agents build at speed.

Steve Blank is right. The speed at which AI can create mocks, UIs, prototypes and end-to-end apps is blazing fast, but it comes at a detriment to learning what the user really wants.

As I mentioned in my other articles, the three things I learned about the challenges of building with AI are what Cosmic is now trying to solve:

Agents default to their training. Codex built our marketplace a finished-looking product before we'd agreed on what the product was. The job became curating the narrative arc for the human while building a system an agent can traverse. (Part 1 of the marketplace build)
Agents build demos. The agent kept handing back screens that described the work instead of letting someone do it, and it cut the evidence that made the product worth using. A deck about the product is one of my best tools. A product shaped like a deck is a story nobody can act on. (Part 2)
Functional is the easy part. My own side build, Tandem, worked in 20 days. Getting it to production grade took two more weeks: load times, 28 security vulnerabilities, rate limiting. Useful, Usable and Joyful still take product work. (Take your vibe-coded app to production)

For those who don't know Steve Blank (really… must I introduce him?), he's the Stanford professor whose Customer Development method started the Lean Startup movement adopted by so many startups, corporations and business-school incubators. His Lean LaunchPad class is now taught in hundreds of universities. Here's what he had to say in his LinkedIn post earlier this week.

Every team showed up with a finished product

In the first class of Spring 2026, the teams didn't present PowerPoints or wireframes the way they had for 15 years. Every one of them demoed a complete product, built with AI, the kind of thing Blank's team used to hope to see at the end of week 10. The teaching team was giddy.

Then he wrote: "We were wrong, wrong, wrong."

The product got ahead of the learning

Students used AI instead of customers for insights and validation, and assumed the answers were correct. MVPs became sales pitches instead of experiments. Teams collected compliments instead of disconfirming evidence, and the people they interviewed reacted to the product instead of explaining their problems.

Blank has a name for it: "AI had created evidence theater."

Pre-AI teams pivoted three or four times over the quarter. This year, teams pivoted late if they pivoted at all, because a polished product felt like product/market fit. As Blank puts it, "It wasn't that AI was hallucinating – the teams were." He calls the result learning debt: the code worked and the deck was polished, but the teams couldn't defend their assumptions or explain the edge cases.

Judgment is the new bottleneck

His conclusion is that the MVP, as an artifact of learning, is dead. The bottleneck has moved from the time and cost of building a product to judgment about what to build and for whom. In his words, that choice "becomes the whole game."

He adds two things founders should hear. Customers now use the same AI tools to build their own alternatives just as fast as a startup can. And with building nearly free, a startup's moat has to be discovered, the same way the customer is. He promises a Part 2 on how the class is changing.

What this means for how we build

I've seen versions of this on my own builds: a finished-looking product before we'd settled the idea, and a demo standing in for the real thing. It's why I'd use the racing adage: go slow to go fast. Spend more time up front on narrative design, information architecture and user research, then speed up with that context.

That's how Cosmic's sprint is set up. Day 1 is discovery and framing: the outcome, the users and what the week is for, before anyone builds. The agents build fast on day 2. Then real users test the build on days 3 and 4, and what they do reopens discovery. The speed Blank describes is real, and we use it. We just don't let it replace talking to the people the product is for.

Frequently asked questions

Is the MVP dead?

Steve Blank argues the MVP as an artifact of learning is dead: when AI can produce a working product on day one, having one no longer proves anything about customers. What still counts is the evidence from talking to them.

What is learning debt?

Blank's term for what teams build up when AI lets them skip customer discovery. The product works and the deck is polished, but the team can't defend its assumptions or explain the edge cases.

How do you build fast with AI without skipping customer discovery?

Use AI to build and to sharpen your questions, and get the answers from real users. Cosmic's sprint does discovery and framing on day 1, builds on day 2 and puts the build in front of real users on days 3 and 4.