When All Can Build Products, What Determines Success? AI PMs' Four Discriminative Skills

In 1984 Apple’s Macintosh democratized desktop publishing, allowing anyone to design and print. Steve Jobs later observed that while tools became accessible, taste did not, noting that 90% of creations were aesthetically poor. Forty years later AI agents, low‑code platforms, and AI IDEs have similarly lowered the cost of building a minimum viable product, but the core problem remains: creating is easy, making something great is hard.

Industry commentary reinforces this shift. The New Yorker (Mar 2024) called “taste” the new buzzword of the AI era, and Paul Graham is quoted saying that when anyone can make things, differentiation hinges on what you choose to make. Axios (Apr 2024) reports AI companies now market their products as “tasteful,” underscoring that when technology is no longer a barrier, taste becomes a new competitive dimension.

The article distinguishes the AI era from the desktop‑publishing era: the former democratizes creation rather than just expression . In the past, tools like PageMaker required knowledge of layout, color, and typography; today AI handles those tasks, shifting the question from “can you use the tool?” to “is the result good?” This leads to a flood of low‑taste products—generic landing pages, flashy but ineffective apps, template‑like offerings—while high‑taste products remain memorable.

What Is Taste?

Many conflate taste with aesthetics, but the article defines taste as judgment ability . Paul Graham’s Taste for Makers links good design to simplicity, timelessness, solving the right problem, and meticulous detail. For AI product managers, taste manifests in four discriminative abilities:

Distinguish user‑stated needs from true needs (e.g., “AI portrait” vs. “see a better version of myself”).

Separate “features” from “value” (a dozen AI capabilities rarely outweigh a single well‑executed one).

Tell “flashy” from “useful” (e.g., a 3D transition that slows tasks vs. a subtle UI that speeds them up; Notion’s quiet “Ask AI” feature exemplifies usefulness over flash).

Separate personal preference from what target users will actually purchase.

These abilities require understanding user contexts—what works for a 40‑plus female audience in lower‑tier markets may differ from a designer’s personal taste.

How to Cultivate Taste

The article presents a formula: taste = extensive exposure to good products + analysis of why they work + repeated personal trade‑offs + judgment that serves target users, not oneself. It proposes a four‑step training loop:

Input and dissect: Review ~100 similar products, not for visual appeal but to identify first‑impression cues, engagement hooks, monetization signals, and cheap‑looking elements. Reverse‑engineer decision logic rather than feature lists.

Force deletion: Practice saying no; removing features reduces cognitive load. The “AI generation cost fallacy” makes it tempting to keep every generated option, but true taste means discarding nine out of ten proposals.

Build a curated sample library: Organize UI components, landing pages, and interaction patterns by scenario, emotion, and user stage, then use them as prompts for AI to align with your standards.

Calibrate with real user behavior: Rely on usage data—where users linger, drop off, pay, or share—rather than surveys, because behavior reveals true taste.

Why “Volume‑Betting” Fails

Some teams adopt a “launch many, hope one explodes” mindset, mistaking rapid iteration for scientific experimentation. True experimentation accumulates knowledge; volume‑betting treats each launch as an independent gamble. AI lowers the creation barrier but does not lower the distribution, operations, compliance, or taste barriers, as Tencent Research Institute notes.

Consequences of volume‑betting include:

Increased homogeneity—without taste, ten AI‑generated products look alike, diluting market impact.

Attention fragmentation—spreading focus across many products prevents deep user insight and experience refinement.

Higher opportunity cost—users have limited slots; a single well‑crafted product outperforms ten mediocre ones.

The recommended strategy is to build one product and perfect its taste, focusing on a vertical, delivery‑oriented approach that leverages deep domain knowledge.

Encoding Taste into Systems

AI product managers should move from “human‑in‑the‑loop” to “human‑on‑the‑loop,” encoding judgment into design systems, prompts, and quality gates. Vague instincts (“it feels off”) must be translated into concrete rules (e.g., headline ≤ 8 characters, error messages must include next‑step guidance). Only when taste reaches the level of enforceable constraints does it scale.

Team‑Level Transmission

Because a single person’s taste cannot sustain a product line, the article advocates three mechanisms for team alignment:

Clear review standards that reference explicit quality criteria.

Design systems that codify spacing, typography, color, and component behavior.

Non‑negotiable quality red lines (e.g., all user‑facing errors must suggest a corrective action).

These turn personal judgment into shared agreements.

Deeper Layers of Taste

Taste varies across generations and cultures—what feels premium to Gen‑Z may seem gaudy to older users. The article stresses that taste is not universal aesthetics but precise empathy for a target segment. It also highlights a meta‑ability: sensitivity to change. As user expectations evolve (e.g., dialogue‑driven interfaces giving way to forms), taste must adapt, staying dynamic rather than static.

In conclusion, the article reiterates Steve Jobs’ insight: “Taste is your moat.” In the AI era, the moat is the ability to sift through AI‑generated options and select the truly great one, a skill that must be continuously practiced, dissected, and calibrated.