In the spring of 2010, about a dozen people at a brand-new company called Xiaomi set out to make a smartphone. None of them knew hardware, and they had no phone of their own to build on. So they started with software: a phone system built on open-source Android, running on other companies’ phones.
Even that was too big. A phone operating system, the founder Lei Jun said later, was a huge engineering job, far beyond a team of a dozen or so. So they picked the four functions that mattered most on a smartphone at the time, calling, texting, contacts and the home screen, and set out to get those right first. In two months the first version of MIUI was done.
Getting it onto a phone meant flashing it over the phone’s own software. Xiaomi set up a forum and asked for volunteers. A hundred people, Lei Jun said, were willing to take what he called a huge risk. To thank them, Xiaomi made their names into the boot screen. On August 16, 2010, the first version was officially released.
A hundred users, four functions done well first. As launches go, it was almost nothing.
Four functions and a forum
What happened next mattered more. By Lei Jun’s account, Xiaomi did no promotion at all, and the number doubled every week: 200 in the second week, 400 in the third, 800 in the fourth. And every week there was a new version. Xiaomi’s 2018 listing prospectus says the company updated the MIUI developer version every week from August 2010, building in direct feedback from its fans.
Getting four functions right first was what a dozen people could manage. What they bought Xiaomi was real people, on real phones, saying what they thought.
Build the smallest real thing that can teach you something, and put it in real hands. That is a minimum viable product. Eric Ries, later the author of The Lean Startup, defined it in a 2009 blog post as the version of a new product “which allows a team to collect the maximum amount of validated learning about customers with the least effort.” He added a warning: “MVP, despite the name, is not about creating minimal products.” A product development firm called SyncDev says it first used the term in 2001, and that it later spread into popular books, Ries’s among them.
A cook tastes the soup before serving the pot. The spoon isn’t a smaller dinner. It’s how you find out whether the pot needs salt while there’s still time to add it. The four functions were Xiaomi’s spoon.
The word in the middle
The easy word to skip is “viable.” When Sony rushed out the first Walkman in 1979, its insides came from the Pressman, a mechanism already proven over 500,000 units, so that the new idea wouldn’t be judged on a faulty machine.
When a first model breaks, it’s hard to tell whether users rejected the idea or the fault. MIUI’s testers had signed up on a forum to install something unfinished. Shoppers in 1979 had not. The less your first users chose to be testers, and the more a failure costs them, the more “viable” has to carry.
The test that liked everything
AI products hit the same trap from another side. A test can have real users and plenty of numbers and still learn little, if it measures the wrong thing.
In April 2025, OpenAI had a new version of GPT-4o, the model behind ChatGPT, ready to ship. Its offline evaluations, the company wrote later, “generally looked good,” and an A/B test with a small number of users suggested that those who tried it liked it. Some of OpenAI’s expert testers said the model’s behavior “felt” slightly off. OpenAI launched. By its own account the update made the model “noticeably more sycophantic,” and on April 28 the company began rolling it back.
OpenAI didn’t blame a single change. Several, each of which had looked beneficial alone, may have combined to tip the model; one was a new reward based on users’ thumbs-up and thumbs-down. Its A/B tests, it wrote, “didn’t have the right signals” to show how the model was doing on sycophancy. Launching on the positive signals from users, it wrote, was “the wrong call.”
If you’re piloting an AI tool with your team, give it a forum as well as a rating button. Start with a few people who chose to try it. Before they begin, write down what would make you stop, including something a thumbs-up can’t show, such as the tool agreeing with a figure you know is wrong. Then read a sample of the actual conversations against what the task required. Among OpenAI’s own fixes: in some cases, an opt-in “alpha” phase, to hear from users who want to give direct feedback before launch.
Twelve modules or one page
Imagine you run onboarding at a company that hires a handful of people a month, and the two-day welcome course gets poor reviews. The plan on the table is an online academy: twelve modules, videos, quizzes, half a year of work.
The smallest version that could teach you something is one page. It lists the twenty questions new hires asked most last year and who answers each, plus a 15-minute call at the end of week one. Next month’s five new hires get it. Before they start, you write down what would count as failure: two of them still asking their manager how to book leave in week two.
Nobody asks about leave. On the calls, three of the five raise the same thing instead. They couldn’t tell which of the company’s forty chat channels mattered, and nobody had told them. Not one of the twelve modules covered it.