The MVP Didn’t Die. It Got Stolen From the User.

MVP used to be the first line of code a founder wrote. In 2013, a team spent a week in a café building a page that could only upload photos and collect likes. The server crashed three times on launch night. Between fixes, someone refreshed the backend and saw a stranger uploading a breakfast photo at three in the morning, and in that moment, they knew they weren’t guessing.

Ten years later, another founder, the night before a pitch, added a gradient loading bar to the demo. Investors won’t wait more than three seconds staring at a blank screen, he said.

That’s the shift. It’s not that “Lean Startup failed.” It’s that the MVP got pulled out of the user’s hands and stuffed into the pitch deck. In an AI cycle, the technical barrier is low enough to spin up a demo in a day; the capital window is so tight that “if we don’t close next week, we can’t follow on”; and the user? The user swipes away by the fourth second of no response. Three rhythms colliding: engineers getting faster, investors getting more impatient, users getting slower.

Last winter, we met a team building an AI chat app. The dashboard looked good: 400,000 sign-ups in the first month, average session length 22 turns. The founder projected the chart onto the wall and asked what valuation we thought he could negotiate. We didn’t answer. We just asked: “That person who dropped off after the 23rd turn—did they come back the next day?”

He was silent. The silence lasted about five seconds.

Later, we traced that user’s path ourselves. Someone spent late nights setting up an entire fictional world for the AI, talked for four evenings, and didn’t open it again on the fifth. The custom character name still sat in the dropdown, but she chose “Default Assistant” in the end. The last line of the chat log was the AI saying, “I understand. Your world is fascinating.” She never replied. This wasn’t a technical problem. It’s not even “unclear demand”—she just never found a reason she had to come back. That reason is something a curve can’t draw.

Internally, we call those projects “hot-start traps.” The next-day retention curve on the dashboard looked textbook-perfect until it fell off a cliff on day seven. That day was exactly when users exhausted their free quota and the paywall popped up. A colleague parked the cursor on that cliff, took a screenshot, posted it in the post-investment group chat, and typed only one line: “What he said wasn’t worth it wasn’t the money.”

That line later got taped onto the whiteboard in our meeting room, next to another sticky note, handwritten: “The sentence the Perplexity team stuck next to their monitor: ‘What did he use us to look up today?’” Two notes side by side. One about why people leave. One about why they come back.

Speaking of Perplexity, one detail got lodged in our heads. Early on, they didn’t do growth. For close to six months, the team only polished a single motion: after a user types a fuzzy question, how fast does the system surface the first citation? Not how fast it delivers an answer, but how fast it lets the user see where the answer is coming from. There was a test version that squeezed citation load time from 1.8 seconds down to 0.9 seconds. That week, the share rate for results jumped 12%. No press release, no blog post. That 0.9 seconds was their MVP.

Not every team works that way. At a Demo Day last year, one project got over ten thousand registrations on the spot. Three people were mobbed by investors at the booth for three hours. Three months later, we walked past their office. Nobody at the front desk. A property management payment demand notice stuck on the glass door. HR was processing departures remotely; the last approval comment read: “Computer returned, access card left in the drawer.” Their MVP validated “can get attention” but had nothing to do with a user’s real morning.

Then there’s another kind of team—they turn the MVP into an interface that stays open. In Mistral’s Discord channel, messages often popped up at three in the morning. Once, a developer pasted an error code and asked: “Is this error because I’m stupid or the model’s stupid?” Within five minutes, an official reply: “Your approach is off. Lower the temperature to 0.3 and try again.” No PR language, no brand speak. We stared at that reply for a long time. It was clearer than any pricing strategy: they were testing, with every single response, the boundary of “what debugging tolerance developers are willing to tolerate.”

On the capital side, the way we judge has also retreated to very small motions. One of our partners always starts a founder conversation with the same question: “Is your first user still logging in?” He says he doesn’t look at the explanation that follows—only at the split second before it. Whether it’s a frown, a sigh, or a flicker in the eyes, that moment is more honest than all the data that comes after.

Another colleague, when doing a post-mortem on Character.ai, projected their user journey onto the screen and dragged the progress bar back and forth, staring at a single metric: the gap before a user’s second message. Among those who didn’t open the app again within 24 hours of their first session, 90% never came back. He highlighted that gap in red and wrote underneath: “Not a lack of interest—it never crossed into non-negotiable need.” That sentence was later written into our internal project evaluation template, replacing the old line “user stickiness needs further validation.”

Another question in the template got rewritten, too. It used to be: “Does the team have rapid iteration capability?” Now it’s: “How long did it take the team to recognize the last abandoned hypothesis as wrong?” One founder answered us: three days. On the third day, they noticed users weren’t clicking that prediction button at all. That same night, they pulled it from the navigation bar and replaced it with a search box. He said the first search query after the search box went live was “how to delete prediction history.”

In an era where everything is accelerating, the teams that didn’t end up in the graveyard all did the same thing: they broke the MVP back down from a “display interface” into a single question. “At which exact action did the user hesitate?” Then they stared at that hesitation until it hardened into certainty.

Just before Spring Festival, one team didn’t launch a new version. Instead, they invited five users into the office, handed each a cup of convenience-store coffee, and asked them to complete the same task. One user paused on a certain button for about four seconds, then clicked somewhere else. The tech lead didn’t say anything. He just wrote down those four seconds in a notebook and added a line: “We assumed he knew this was search.”

The coffee went cold. Outside the window, someone set off fireworks. He closed the notebook and said, “Tomorrow, change that button into an input field.”