Taste Is Seeing Something Deeper Than Everyone Else

This is the sixth episode of Stolen Chat. Allen — known as "Shi Daxia" (史大侠) — studied physics, started his career as a hardware product manager at Hitachi, found his direction through a single UX class, then went on to build AI applications at Didi and work on AI and content safety at Roblox. He's been through the journey from Series A to IPO. Now he coaches product teams globally, helping AI startups find product-market fit, scale to tens of millions of users, break through fundraising bottlenecks, and compress the zero-to-one validation cycle.

Three keywords run through this conversation. Taste — not eternal, but the depth of understanding you demonstrate in a specific context. Need — you should build things that were previously impossible; that's a need. Specification — if you're just repackaging what was already possible, that's a spec.

One UX class and one SEO class

Allen graduated in physics. Half his classmates went to TSMC, but he knew clearly he didn't want to stay in a lab.

The turning point was two classes in an accelerator. A UX class taught him that when product positioning isn't clear, changing creative assets and increasing ad spend is useless — the problem isn't the assets, it's that you never figured out what you are from the start. An SEO class showed him the underlying logic of the internet — how your content gets seen by users is a channel and distribution problem.

These two classes turned him from a "project manager" into a "product manager." The difference: project managers execute tasks; product managers make decisions.

Hundreds of business models matched against cultures

Allen has a habit: when evaluating any product direction, he doesn't look at a single path. He reshapes it across multiple dimensions simultaneously. Time — what does this model look like five years ago, now, and five years from now? Geography — how does the same product develop differently in the US, Japan, and Southeast Asia? Culture — which differences are caused by culture, and which by infrastructure?

He gave the example of Brex. US corporate credit card systems have been established for years. Brex seized the momentum of a wave of unicorns emerging and built virtual corporate credit cards. But the same product in Asia — where corporate credit card penetration is extremely low — requires a completely different entry point.

"Most AI product teams lead with ops and marketing, then reverse-engineer how to adjust the business model. There's no absolute right or wrong, but if you have deeper understanding of the market and culture, you can achieve much deeper user scenario exploration."

Didi: from safety incidents to an experience system

He joined Didi's AI team in 2016-2017, when the team was single digits. He chose Didi over ByteDance or Meituan for a clear reason: Didi's entire chain can monitor a person's pre-trip, in-trip, and post-trip experience — which aligns perfectly with UX user journey thinking.

Then the carpool safety incident happened. This tragic event actually gave the AI team a critical entry point — the platform handles tens of millions of rides daily, carrying people not cargo. How do you push service and experience to the extreme in safety and balance?

Starting from the safety chain, they gradually upgraded the entire customer service system. Allen says most people understand customer service as "complaint handling" — users call in angry and venting. But fundamentally, customer service should be one step in the customer service process, focused on handling emotions and problems in the moment, not making people loop around endlessly looking for a human agent.

"The best experience? Don't tell me I need to do anything else. Just solve it for me."

Take invoicing as an example. Stage one: user calls in saying they can't get an invoice, customer service handles it. Stage two: system predicts you'll have an invoice issue and proactively offers a self-service channel. Stage three: when you come in, it simply tells you the problem has already been resolved.

From reactive response to proactive prediction to direct resolution — these are AI's three levels in service experience.

Roblox: never thought of itself as a metaverse company

After Didi, he joined Roblox. The reasoning was simple: once all internet efficiency gains hit their ceiling, attention will inevitably flow to entertainment.

He made two phone calls to his aunt and uncle in the US — two people who couldn't use smartphones told him "Roblox is a good company because the neighbor's kids all play it, and some even make their own games."

When someone who doesn't need to understand technology can sense the product's power — that's enough.

At Roblox he worked on AI and content safety products. Two insights different from Didi:

First, when building products for kids, you can't guess what they'll like. Roblox's principle is maximum freedom with minimal intervention. The moment you tell users "step one: build a house," they get trapped in the house-building template. A platform that sparks creativity should let users explore on their own.

Second, UGC platform explosions are unpredictable. When Squid Game blew up on Netflix, hundreds or thousands of user-made versions appeared on Roblox overnight. Allen says Roblox didn't make the game — Roblox is the platform. But precisely because it's free enough, users turned trending content into their own creations.

"Roblox never thought of itself as a metaverse company. Internally, no document or terminology ever mentioned the word 'Metaverse.'"

Cross-market growth: every market runs on different logic

Allen shared a Japan B2B case.

Japan has lifetime employment. Regular employees making mistakes on reports face no serious consequences. But senior professional managers sign labor contracts that demand extremely high certainty. This is why Japan's consulting culture is the strongest in the world — not because Japanese people love consulting, but because executives face huge risk in making decisions and need third parties to share that risk.

So if you're doing B2B in Japan and they say "we've never heard of your company," the deal is essentially dead. Not because your product is bad — because for an executive, introducing an unknown vendor is an enormous career risk. That's why Japan is covered in outdoor advertising — you need decision-makers to have already "heard of" you before they make a decision.

"Same B2B, completely different playbook in every direction. You really have to look at it holistically — culture, internal company structure, and the perspective of the people being employed."

Content going global: you can't explain why it goes viral

Allen has worked on Roblox's gaming platform and on novels and comics going overseas. His take on content going global is sharp:

"You can find every reason why something doesn't go viral. But you can't explain why it does."

So over-relying on editors and human intervention for content selection has limits. Hitting 5 out of 10 projects is extremely difficult. Ultimately you have to give power back to local creators and users and let the market give you feedback.

Games and content have a core difference: games have almost zero language barriers — if it's fun, it works. But cultural content — novels, short dramas, comic dramas — has clear migration paths. Some travel from the US to Korea/Japan to Southeast Asia. Some go directly from Korea/Japan to Taiwan and Southeast Asia. You need to understand this penetration process to know what content to use in which market.

Plaud and Monica (pre-Manus): three criteria for category innovation

Allen helped Plaud (recording card) and Monica (predecessor to Manus, a Chrome extension) with their go-to-market. He has three criteria for selecting products to work with:

First, he has to use it himself. If he uses it and believes in it, convincing others is simple. He brought a few Plaud devices to a corporate dinner, demo'd "press a button during a call to record, then send the transcript directly to the other person." Every single device at the table — including his own — got sold on the spot.

Second, communication and education costs must be low. Open a laptop, install Monica in ten seconds, demo immediately. No app download, no waiting. Products with high education costs are painful to push.

Third, iteration speed must be fast. He requires partner teams to interview 50-100 customers per month minimum. Big companies can't deliver on a daily basis. Startups can — that's the advantage.

AI era doesn't need PMs? False premise.

Allen's judgment: when productivity becomes simple enough — Anthropic handles most of the work for you — the next most important decision is "which battlefield should I invest resources in to go deep."

That's the product manager's most important job: making decisions from the user's and market's perspective.

He used a pendulum metaphor. Early on, AI product managers were rare — route planning and pricing were enough. Then fine-grained operations took over, PMs subdivided endlessly — Facebook might have thirty or forty PMs plus designers on a single page. Too scattered. The world swings like a pendulum back the other way. Past era: PM as product gatekeeper. AI era: back to a few people making key decisions.

"No matter how the pendulum swings, the core is always the same: how do you have a good insight today to make a decision."

Taste, specifications, and needs

Finally, Allen defined "taste."

"Taste means you see something with extraordinary depth. It's not eternal — it's the deep understanding you demonstrate in a specific context and moment. People with taste naturally let go of unimportant things and focus on what matters — that's why their work goes deep."

He gave the example of Granola — the meeting recording product. Most user interviews yield the same answer: "give me a real-time word-for-word transcript." But someone with sense realizes that in a 1-on-1 meeting, the core is exchange and sparks. If there's a transcript running alongside, people get distracted. The real need isn't a transcript — it's how to preserve the value of the exchange without disrupting the exchange itself.

"Many teams are doing things that were already possible before. That's not a need — that's a specification. You should build things that were previously impossible. That's a need."