Years ago I was at a Web 2.0 conference in New York, sitting next to a young guy with the sweetest baby face. Seriously, when people say someone has the face of an angel, this kid did. The speaker on stage was terrible, and the two of us spent a happy fifteen minutes quietly mocking him. Then my new friend got up, walked on stage, and blew my mind. It was Eric Ries, giving one of the first Lean Startup talks.
The idea was one of those brilliant ones that look so obvious once you hear them that you slap your forehead: what if we tested our ideas before we launched them?
I became an evangelist on the spot. I dragged my entire MySpace team to the first Lean Startup conference. I bought the first version of the book when it was just his blog posts, bound (an MVP of a book! Who wudda thunk! I stole that idea for Radical Focus). The Lean Startup movement transformed how I worked, how I taught, even how I lived.
Lean was a revolution as big as Agile. And right now, as we experience the AI revolution, we keep forgetting what it taught us.
Sure, development time is collapsing. You can build in an afternoon what used to take a team a quarter, and people are a little drunk on it. I get it; I’m a little drunk on it too. But if you need to actually make money from your idea, it is still cheaper and faster to test it than to build it and hope you guessed right. The harder part of product development wasn’t building cost or time. It was answering the question, “Will anyone pay for this?”
So why do we skip the testing? Because seeing your idea become real is heady stuff. Humans like making things. Kids build dams and bridges and sandcastles by the ocean. Lego sells well. Michaels is still in business. You don’t need to be talked into building. Building is the reward.
And we don’t love being told our ideas are lame. Building from your vision is fun. Finding out people don’t care about your vision… not so much.
The validation gap
AI output really is getting good, so let’s give the tools their due. In late 2025, OpenAI published GDPval, a benchmark where professionals blind-judged AI deliverables against work from industry experts averaging fourteen years of experience. On well-specified tasks, the best models tied or beat the human expert about half the time at launch, and newer models have pushed past two-thirds.
A month later, Scale AI and the Center for AI Safety published the Remote Labor Index: 240 real freelance projects from actual paying clients, judged by whether a reasonable client would accept the work. The human freelancers who originally did those projects earned $143,991. The best AI agent earned $1,720. An automation rate of 2.5 percent. (Both numbers are as of this writing, and both are moving fast.)
Sit with those two results. AI ties the professional on a deliverable, and manages 2.5 percent of a job. I call the space between them the validation gap. GDPval handed the models a brief written by a fourteen-year veteran, which means every task arrived pre-validated by an expert who already knew it was worth doing. The Remote Labor Index made the models do the whole job, including the part where somebody has to want the result. AI is getting great at executing a spec. The job around the spec (knowing what to build, for whom, and whether anyone will pay) hasn’t budged.
Most of those who read The Lean Startup walked away chanting MVP, but the word that mattered was validation. Ries broke one big bet into a series of small bets, each one tested before more money followed it. Build-Measure-Learn was the loop, and AI just made Build the cheap third. It didn’t touch the other two. Validated learning was the revolution; the minimum viable product was only its cheapest instrument.
What to do instead of tweaking
First: if you’re building for yourself, some tool for your own weird hobby, carry on. That is 100% fine, and it’s only now possible for noncoders. Build whatever makes you happy and skip the rest of this post.
For everyone else, test the idea before you polish the build. Mention it to someone in line at Starbucks. Bounce it off someone at work. If they react positively, find your target audience (you can pay research firms to do this for you.) Make a quick prototype of the most valuable part of your idea, show it to that audience, and ask what they’d pay for it.
There is a ton of nuance I’m skipping, like how to define your target audience and how to do good research so you don’t get false positives. But you know what? There are books for that. UX for Lean Startups, by Laura Klein, handles the research problem. And The Lean Startup itself, validated by time and by thousands of businesses, should still be your first step.