In a Juejin column last week, 怕浪猫 (Scared Stray Cat) dropped a fact that made many AI product managers uncomfortable: a real product goal looks like "40% of test users return within 3 days"—never "users will love it." What this line hits isn't the methodology itself, but the same mistake the AI industry is making en masse right now.

What this is

怕浪猫's serialized Juejin column "Product Design from 0 to 1, Full-Process," Chapter 8, makes one core argument: product decisions aren't driven by gut—they're a three-step "hypothesis → validation → decision" loop.

The article breaks down three cognitive biases most likely to derail teams—

  • Validation bias: you only test on people predisposed to buy in, so results look great
  • Survivorship bias: you only look at users who stayed; the ones who left are already gone from the data
  • Confirmation bias: humans are wired to hear what they want to hear

怕浪猫 offers an operational template: write the idea as a testable hypothesis (with metric, threshold, and time window), test at the lowest possible cost (from desk research up to an MVP—the minimum viable product—ranked low-to-high cost), pre-write what counts as failure before you run the test, and don't move the goalposts after the experiment.

Industry view

This methodology isn't new, but in AI it cuts deep. We've noticed that over the past six months, as large-model products have launched in waves, many teams' practices have landed exactly in the three traps 怕浪猫 names: testing with a self-selected enthusiastic cohort and declaring "the direction is right," reading retention only off the best users, and treating a fractional bump in model performance as a "breakthrough."

Counterarguments deserve airtime. A serial founder told us privately: "This methodology works great on toC products, but on toB AI products, customer engagement is structurally low—you ask them to validate a hypothesis and they can't even articulate their requirements." A second voice, from a product lead at a major tech company: "Our most expensive cost isn't engineering—it's the time window. By the time you've validated, the race has already changed."

Our judgment: 怕浪猫's methodology isn't for deciding what to build—it's for avoiding the pitfalls you only discover after the fact. In an AI market drowning in product homogeneity, dodging one derailment can be worth an entire R&D budget.

Impact on regular people

For enterprise IT leaders: the next time a vendor tells you "our users love the product," ask one more question—which users, in what scenario, with what control group?

For individual careers: the AI tools you're currently using (writing assistants, meeting-notes generators, PPT builders) almost certainly fell into these three traps. When a feature ships and immediately dies, the reason may not be "AI doesn't work"—it's that no validation was done before the decision.

For the consumer market: in the next 12 months, expect a wave of AI products to exit quickly. Whether a product survives depends less on its feature list than on whether it was built from a clear, testable hypothesis.