Product conversations used to begin with two questions: can we build it, and how should we build it?
AI is rapidly lowering the cost of both. An idea that once required a team and several weeks can now become a working prototype in a matter of days.
As making gets easier, the difficult questions move earlier:
What should we build in the first place?
When dozens of directions look promising, which one should we try first?
If the data is weak, should we improve the product or stop?
With so many possible things to make, where should we spend our time?
I recently listened to several episodes of Lenny’s Podcast with Instagram head Adam Mosseri（在新标签页中打开） (opens in a new tab), iPod creator and iPhone co-creator Tony Fadell（在新标签页中打开） (opens in a new tab), and Andrew Ambrosino, who leads the Codex desktop app at OpenAI（在新标签页中打开） (opens in a new tab).
They were discussing different products at different stages, but their conversations kept returning to the same question:
When technology becomes more capable and the number of possible answers keeps growing, what should people rely on to make a decision?
I have started to think that taste in the AI era is no longer just visual sensitivity. It is a broader form of judgment: knowing what is worth making, what to leave out, and where a product should go.
Who decides when there is no answer to copy?
For a product that already exists in the market, we can study data, competitors, and established patterns.
A true version 1.0 rarely has that luxury. There is no reliable historical data for something that has never existed.
Tony Fadell described the argument inside the original iPhone team over a physical keyboard versus a virtual one. The team ran many tests, but the tests could not produce the only correct answer. Someone still had to consider the technological direction, the user experience, and the product as a whole, then make the call and accept the risk.
That kind of decision is not a random hunch.
It comes from accumulated experience, an understanding of users, a feel for technical limits, and intuition sharpened by being wrong many times.
Data helps us understand what has already happened. Where no answer exists yet, someone still has to decide where to go next.
When output becomes cheap, filtering becomes expensive
Generating dozens of approaches to the same feature is no longer difficult.
The difficult part is deciding which one addresses a real problem. Which details are worth keeping? Which ideas are merely novel but make the product more complicated? Should a feature stand alone, or should it disappear into a larger user journey?
Product teams once struggled to turn ideas into working software. Increasingly, the scarce skill is finding a direction among too many plausible options.
That is why I like to think of the future product manager as a curator.
A curator does not create every object in an exhibition. They establish a standard, select the work, arrange the relationships between pieces, and shape the experience of the person walking through the room.
Products work in a similar way.
AI can keep producing prototypes, copy, interfaces, and code. It cannot decide why a feature should exist, how it belongs in the larger system, what a user truly needs to see, or what should be removed.
I began to see validation speed as a form of taste
While listening to these conversations, I kept thinking about the AI products I have worked on recently.
For Mengya, I quickly assembled the product logic, interface, recruitment, and testing flow. The first users were willing to register and have fairly deep conversations with the AI. We found a clear moment of value: when someone described their life in a scattered way, the AI could help them feel seen and organized.
But day-one retention was only 15%, below the threshold I had set.
I decided to close the project.
Closing it did not make the experiment worthless. It confirmed two things. People did want to understand themselves through conversation, but one good conversation was not enough to make them return.
I kept conversational AI in the next product, Snowball, but stopped asking the conversation to carry the entire experience. The dialogue helps understand the person and break down a goal. A persistent product structure carries long-term goals, daily actions, and feedback.
When I joined the work on Fanshu Intelligent Edition, I also moved quickly on distribution. In the final analysis, 77 creators could be identified accurately, and the campaign brought in roughly 100 to 130 new users.
The result did not produce the sustained growth we expected.
Looking more closely at the data, the 18.2% next-day retention rate made me realize the immediate problem was probably not a lack of traffic. It was the first experience for people who were not already Fanshu users, and whether they had a reason to come back.
These experiments changed how I understand judgment. It does not mean guessing the right answer at the beginning.
It means forming a useful hypothesis before all the information is available, putting it in front of real users quickly, facing the result honestly, and deciding what to keep and what to let go.
Validation speed is also a form of taste.
Taste becomes visible in specific decisions
AI will keep making creation faster.
We will have more prototypes, more content, and more features. We will also face more directions that all appear possible.
The scarce capability may no longer be generating more options. It may be developing a standard for choosing among them.
It is knowing what problem a user actually needs to solve, where a feature belongs in a larger system, what deserves more craft, and when it is time to stop.
Taste does not guarantee that someone will always make the right choice.
It means noticing earlier where you are wrong, deciding faster what should remain, and being willing to take responsibility for the direction.
AI gives us more power to create. It also makes one question harder to avoid:
Of all the things I could build, what do I actually believe in?