5 guests 5 episodes 3,449 words

The Taste Fortress: Will AI Commoditize Design Judgment or Make It the Last Moat?

Will AI commoditize design taste, or make human taste the last competitive moat?

For two years, designers have been telling themselves a reassuring story. AI would get good at code, then pixels, then mocks — but the last fortress, the one marked "taste and judgment," would remain uniquely human. Jenny Wen, who runs design at Anthropic, does not believe the story anymore. From the inside of the company whose models are the benchmark for "AI with taste," she thinks her profession is holding onto the myth "a little bit too tightly."

On the other side sit some of the most credible operators in product: Dhanji R. Prasanna, the CTO running 3,500 engineers at Block where an in-house agent saves employees 8-10 hours a week; Lazar Jovanovic, Lovable's first "vibe coding engineer"; Karri Saarinen, whose craft-obsessed Linear wins customers from Jira without running a single A/B test; and Marc Andreessen, who splits the design job in two and argues the elevated half has never been more valuable. Their disagreement is the defining bet of the AI-native era of product building.

If AI can now produce thousands of icon variants in seconds, critique layouts, refactor flows, and improve at what looks like aesthetic judgment every few months, is human taste on its way to being commoditized like every other production skill before it — or is taste the one thing that gets more valuable, not less, as everything around it goes cheap?

The 3 Positions

1 / 3

The Fortress Is Already Falling

Jenny Wen

Jenny Wen — Head of Design for Claude Cowork at Anthropic, previously Director of Design at Figma where she ran FigJam and Slides. Jenny leads design at the company whose own models are being held up as the benchmark for "AI with taste," and she has watched designers wrestle with tool-induced commoditization twice — first at Figma, now at an AI lab.

Jenny's argument is the uncomfortable one. The canonical design process — research, diverge, converge, mock, polish — was already obsolete before AI landed, and the Claude Code wave just made it visible. Engineers running parallel Claudes ship scrappy versions faster than anyone can make beautiful mocks, and in her own job the mocking share has fallen from roughly 60-70% to 30-40% in a few years. Against that, designers telling themselves "we'll always be the arbiters of taste" reads to her like a retreat into story.

"This design process that designers have been taught, we sort of treat it as gospel. That's basically dead. You as a designer actually do not have the time to make these beautiful mocks anymore."
Jenny Wen
▶ 00:00:00

When Lenny asks her directly whether AI will get very good at taste and judgment, she refuses the consoling answer — taste is on the same improvement curve as everything else the models are climbing, and she is willing to say so out loud even though her profession would prefer she did not.

"I think it will get better at taste and judgment and design. Yeah, I think we might be holding onto that a little bit too much and saying, 'Oh yeah, a designer or somebody will always know the best thing to ship or the best version of this.' But I do think AI's sense of taste will get better."
Jenny Wen
▶ 00:28:30

What does endure, in her view, is a narrower and less flattering thing: accountability. Someone still has to decide what gets built and own the consequences, the same way an engineer signs off on AI-written code, or a radiologist signs off on a machine read. That is judgment-as-liability, not taste-as-moat — a much smaller fortress, and the only one she will defend.

"At the end of the day, someone has to decide what is actually going to get built and what actually matters. Someone still needs to be accountable for the decision."
Jenny Wen
▶ 00:01:03

When This Applies

when you are building in high-liability domains — medical, legal, financial, safety-critical — where the radiologist analogy fully applies. The human is there to sign off, not because the machine can't see the tumor.

Taste Is the Anchor That Keeps AI From Drifting Into Slop

Dhanji R. Prasanna · Lazar Jovanovic · Karri Saarinen

Dhanji R. Prasanna — Chief Technology Officer at Block, overseeing 3,500+ engineers. Dhanji runs engineering at one of the few large companies that actually built and open-sourced its own agent, Goose, and measures its impact in hours saved per employee — so his "taste as moat" argument is grounded in running autonomous agents at scale, not theorizing about them.

Dhanji is not a designer defending turf. He is the CTO of an aggressively AI-native company, and his argument for taste as a moat comes out of what he has seen when you let LLMs free-run at scale. At Block, agents succeed on roughly 60% of well-described tasks and go off-script on the remaining 40% — building the wrong thing eagerly, twisting teams into knots automating broken processes, generating output no one asked for. The failure mode is drift, and drift compounds.

"I do think we're going to need a lot of human taste to anchor these AIs so they don't go off script to be honest. And that's really where our design lead and our design teams are pushing us to think, and that's a differentiator that I think will push us beyond this era of AI slop that everyone's talking about."
Dhanji R. Prasanna
▶ 00:37:56

He is careful to distinguish taste from aesthetic polish. For him, taste is the lens that decides what is worth building at all, paired with a sibling skill he calls "portfolio judgment" — the global sense of which projects shouldn't exist in the first place. Those are the two things Block's agents consistently fail at, and they are how you avoid drowning in your own output.

"This is the thing that AI isn't good at, it's not able to have this portfolio judgment or judgment across a global sense of what's important and what matters."
Dhanji R. Prasanna
▶ 00:38:43

Dhanji would say you can run a company with AI producing 90% of the code — and Block more or less does — but the remaining 10% of human judgment becomes more load-bearing, not less.

"It's very much anchoring it into a thing that matters to people and the thing that's tasteful and useful and has value."
Dhanji R. Prasanna
▶ 00:37:56

Lazar Jovanovic — Lovable's first official "vibe coding engineer," a job title that did not exist until he created it. Lenny has called him the top 0.1% of vibe coding skill; his working day is a live experiment in what is left when AI does the production work, which is exactly the world the commoditization debate is about.

Lazar has arguably the cleanest incentive to concede the commoditization thesis — his entire job is getting paid to produce software without writing code by hand — and he says the opposite. His central claim is that AI is an amplifier, not a substitute, and amplifiers multiply both signal and noise. Taste asymmetries do not shrink under AI; they widen. If you have judgment, AI makes you dramatically more productive. If you do not, you get to generate garbage at industrial scale.

"AI, regardless of your background, is an amplifier. If you don't know what you're doing, you're just going to produce garbage faster."
Lazar Jovanovic
▶ 00:00:29

His workflow matches the theory. Before Lovable touches a single line, Lazar spends a full day writing PRDs — master plan, implementation plan, design guidelines, user journey — because when production is free, the binding constraint shifts upstream to clarity and knowing what good looks like. His Aladdin-and-the-Genie analogy captures it: the tool grants exactly what you ask for, so the quality of the ask is the whole game. Ask to be "taller" and you get made 13 feet tall.

"I'm optimizing 100% of my time today on good judgment, clarity, quality, taste."
Lazar Jovanovic
▶ 00:00:56

Then he goes further than anyone else in the archive, with the inversion argument. Historically, commoditizing production tends to elevate the remaining hand-made version into a prestige good — calligraphy, tailoring, letterpress. He thinks code is on the same trajectory. The structural point holds either way: making something cheap makes the judgment around it rarer.

"Coding is going to be like calligraphy. People be like, 'Oh, my God, you wrote that code? That's so amazing.' It's going to be so rare that it's going to become an art."
Lazar Jovanovic
▶ 00:00:17

Karri Saarinen — Co-founder, designer, and CEO of Linear; previously founding designer at Coinbase and principal designer at Airbnb. Runs the product most often cited by other founders as proof that craft is a business moat; Linear is profitable with net-negative lifetime burn, has no A/B tests, and still wins customers against well-funded incumbents in a commoditized category.

Karri's contribution is the one that is not really about AI, which is why it cuts. His claim, built over years of running Linear as a deliberately craft-obsessed company, is historical: when a medium is new, rough work wins because everything is rough; as the medium matures and tools commoditize production, the bar to even be considered rises, until serious entry requires craft. The pattern repeats every time — web, mobile, and now AI.

"My belief is that, like any domain or industry, the more it matters, the more the design matters. What happens is whenever there's a new paradigm, I don't know, like the mobile or the web or something the first iterations of those products existing there, they don't have to be super well designed necessarily because they are the first."
Karri Saarinen
▶ 00:00:00

Applied to the AI era, AI is simply the next wave of production-tool commoditization — and those waves have historically made craft a stronger differentiator, not a weaker one. When every tool generates decent UI, the companies that can generate great UI and recognize what "great" means in context win the next round.

"Then, as you build the 100,000 different email clients, any email client now has to be pretty good to be even considered an email client. It's like the bar is so high. I think today it's almost a very basic thing now. Pretty much from the very beginning, you need pretty high level design for people to even pay attention or consider you seriously."
Karri Saarinen
▶ 00:00:19

He pre-empts the most common objection — that craft is too slow for the AI era. Linear ships fast by launching early to small groups and iterating in production. Craft is not a finishing pass bolted onto a finished product; it is the production model. It is a sequencing problem, not a speed problem.

"I think sometimes people think that craft is about perfecting things and perfecting them in a very organized way or very early on. The downside with this perfection mindset is that it can be sometimes hard to put anything out because nothing is ever fully perfect. We try to balance this thinking with the fact that we should be always pushing things out very quickly, but then also fixing them, improving them very quickly."
Karri Saarinen
▶ 00:11:21

When This Applies

when you are competing in a mature category where users can tell great from merely good and production-cost collapse is making competitors indistinguishable. Linear in project management, Superhuman in email — categories where the bar is high before you are even considered.

Split the Layer — The Task Dies, Capital-D Design Is Elevated

Marc Andreessen

Marc Andreessen — Co-founder and General Partner at Andreessen Horowitz, co-creator of Mosaic and Netscape. Has invested in essentially every generational platform shift and spends his days talking to the most AI-native founders; his thesis here is not sentimentality about design, it is what he is seeing across the a16z portfolio.

Marc refuses the question as posed. He splits design into two layers on different trajectories. At the task level — design the perfect icon, lay out the screen, polish the mock — AI will eat the work, and he happily concedes it. Most designers, in his telling, never got to do the interesting part anyway because they were buried in task-level execution.

"The task level of, like, 'Design the perfect icon,' is going to be, all right, the AI's going to do that all day long. If it gives you a thousand icon designs, it's going to be great. It's going to be fantastic, whatever. And by the way, there will still be some level of human icon design or whatever, but AI is going to get really good at that."
Marc Andreessen
▶ 00:51:54

The interesting part is what he calls "capital-D design" — the higher-level questions the greatest designers have always asked. What is this thing for. Will it fit into people's lives. Will it challenge them in the right way. Those are taste in the deepest sense, and his claim is that AI doesn't just fail to commoditize this layer; it actively expands access to it by freeing up the time and attention that used to be eaten by task work.

"All of these kinds of higher level questions that the great designers have always thought about. The job of designer will involve much more of those higher level, more important components, and then again, with AI doing a lot more of the underlying tasks."
Marc Andreessen
▶ 00:52:40

The Jony Ive framing is the load-bearing image. For the first time, a 25-year-old designer who really harnesses AI has a realistic path to Jony-Ive-range fluency — operating across all the underlying tasks while spending most of their attention on the higher-level things. To Jenny he would say: yes, the process is dying, but what is dying is the process, not the taste. To the taste-as-moat side: yes, but the moat moves from "I can produce beautiful mocks" to "I can answer the higher-level questions," which is a strictly harder bar and will filter brutally.

"You think of the world's best designers, Jony Ive or whatever, and you could be like, 'Wow.' Like, if I'm a designer today, if I'm a 25-year-old designer and I aspire to be Jony Ive in a decade, it's all of a sudden, I have a new path that I can use to get there... because Jony did everything. He did it without AI."
Marc Andreessen
▶ 00:52:40

When This Applies

when you are designing a career, a hiring rubric, or an org chart. Assume the task layer goes; hire and develop for the capital-D layer. Optimize ruthlessly for the questions Jony Ive asked, not the mocks he produced.

Evidence from the Archive

Anthropic

At Anthropic, engineers now spin up 'seven Claudes' in parallel — designer mocking has fallen from 60-70% of the job to 30-40% in a few years

Jenny Wen argues designers are holding onto taste-as-moat too tightly; AI's sense of taste will get better and the classical design process is already dead

Jenny Wen · Head of Design, Claude Cowork at Anthropic; ex-Figma design director

Block

Block's Goose agent saves engineers 8-10 hours/week — but agents go off-script on ~40% of tasks, making human taste the anchor that prevents AI slop

Dhanji Prasanna runs engineering for 3,500+ at Block, ships its own agent, and argues that the 10% of work that's human judgment becomes MORE valuable as AI handles the rest

Dhanji R. Prasanna · Chief Technology Officer at Block, overseeing 3,500+ engineers ▶ 00:37:56

Lovable

Lovable's first 'vibe coding engineer' spends 100% of his time on judgment and taste — and a full day writing PRDs before touching the code generator

Lazar Jovanovic gets paid to produce software without writing code by hand, and his view is that AI is an amplifier: if you don't have taste, you produce garbage faster

Lazar Jovanovic · First official vibe coding engineer at Lovable

Andreessen Horowitz

Marc Andreessen argues a 25-year-old designer today who harnesses AI has a realistic path to Jony Ive-level range for the first time in history

The task layer of design (icons, mocks) gets commoditized; the capital-D layer (meaning, fit, emotional resonance) gets elevated — and AI actively expands access to the higher-level work

Marc Andreessen · Co-founder and General Partner at Andreessen Horowitz

Linear

Linear is profitable with net-negative lifetime burn in a commoditized category (Jira, Asana, GitHub issues) — built entirely on taste, no A/B tests

Karri Saarinen's structural argument: as a medium matures and production tools commoditize, the craft bar rises until even being considered requires high design

Karri Saarinen · Co-founder and CEO of Linear ▶ 00:00:19

The Synthesis

The commoditization debate is usually fought as a binary, but the most instructive thing about these five voices is that nearly all of them are arguing different flavors of a single structural claim: production is collapsing in cost, and whatever sits adjacent to production becomes disproportionately more valuable. They just disagree about what "adjacent" means.

The Three Things Called Taste

Which kind of taste are you actually defending?

Amplifier Asymmetry

Does AI compress outcomes or stretch them?

Anchor vs. Drift

What stops your agents from building the wrong thing eagerly?

The Rising Bar

Has your category already crossed the craft threshold?

Where Scarcity Lands

After production goes free, what's the new bottleneck?

Jenny Wen's sharpest move is noting that 'taste' collapses three different things: aesthetic execution, judgment about what to ship, and accountability for the decision. Aesthetic execution is on the model's improvement curve. Accountability has to stay human because someone owns the consequences. Judgment is the contested middle — and most designers don't know which they're actually selling.

Lazar Jovanovic's 'produce garbage faster' is a precise claim about distribution: commoditizing production doesn't collapse outcomes around the mean, it widens the gap between people with judgment and people without. If taste is an amplifier input, AI makes taste asymmetries bigger, not smaller — at least during the transition.

At Block, Dhanji's agents succeed on 60% of well-described tasks and go off-script on the rest — building the wrong thing, automating broken processes. As output rate rises, the cost of pointing at the wrong thing rises with it. Taste becomes the anchor preventing AI slop, paired with portfolio judgment about what shouldn't exist at all.

Karri Saarinen's historical pattern: when a medium is new, rough work wins because everything is rough. As production tools commoditize, the bar to even be considered rises until serious entry requires craft. Web, mobile, and now AI each trigger this — Linear wins email-client-style commoditized categories precisely because the bar is now high enough to keep weak competitors out.

This whole debate is really about where scarcity re-lands after production collapses. Jenny bets on accountability; Marc on capital-D design; Dhanji on portfolio judgment; Lazar on clarity of specification; Karri on the ability to enforce a rising bar. These are five forecasts of the next scarce skill, not five answers to the same question.

Which Approach Fits You?

Answer 3 questions about your situation. We'll match you to the right approach.

Question 1

What does your competitive environment look like right now?

Question 2

How is AI actually being used inside your team today?

Question 3

Where do you believe the scarce human skill will be in three years?

Notable Absences

The Bottom Line

The non-obvious insight is that this debate is really about where scarcity re-lands after production goes free. Jenny bets on accountability — a thin, unglamorous layer of sign-off. Marc bets on capital-D design — a thick layer of meaning-making that most designers never got to touch. Dhanji bets on portfolio judgment — knowing what shouldn't be built at all. Lazar bets on clarity of specification — knowing how to ask. Karri bets on the ability to recognize and enforce a bar that keeps rising. These are five different guesses about which scarce human skill ends up bottlenecking AI-native production, not five answers to the same question. "How Linear builds product" is the operationalized version of Karri's bet; "The definitive guide to mastering product sense interviews" pressure-tests the other direction, treating taste-adjacent skills as trainable components — which is exactly Jenny's wager. The honest reading is that taste-as-moat has stronger near-term evidence and commoditization has stronger long-run trajectory. Which one you weigh more is a forecast, not an analysis.

The Dhanji-Lazar-Karri axis is making a subtler argument that is easy to miss because it sounds like the same one. They are not claiming taste is uniquely human forever. They are claiming that in the transition phase the asymmetry between people with taste and people without it widens, because AI is an amplifier. Lazar's "produce garbage faster" is a precise claim about distribution: commoditizing production does not collapse outcomes around a mean, it stretches them. Karri's "bar rises as the medium matures" is the same claim told historically, and Linear — profitable, net-negative burn, no A/B tests — is its proof. Dhanji's "anchor" language is the same claim told operationally: as output rate goes up, the cost of pointing at the wrong thing goes up with it.

Sources

Dhanji R. Prasanna — "How Block is becoming the most AI-native enterprise in the world | Dhanji R. Prasanna" — Lenny's Podcast, October 26, 2025
Karri Saarinen — "Inside Linear: Building with taste, craft, and focus | Karri Saarinen (co-founder, designer, CEO)" — Lenny's Podcast, October 8, 2023