Concept Over Craft

The Warhol precedent

Eric Glyman frames generative AI through Andy Warhol.¹ The pre-modern art world, epitomized by the Mona Lisa, held that an artist had to give a lifetime to a craft; Da Vinci's years of learning to paint were themselves the art. Warhol, following Marcel Duchamp roughly a hundred years earlier, inverted this. He came from commercial printing and illustration and used silk-screen, the medium of advertising, to make works that were, in Glyman's words, just as good as the original, repeatedly. He ran what he called the Factory, a standardized daily process where anyone, even a security guard, could contribute, deliberately stripping out how a piece was made so that focus went entirely to what was made: the concept. Glyman's read: "Maybe he's the original generative artist." The objection that a studio member printed a work and Warhol only signed it, so it is not really a Warhol, dissolves against the market: Warhols sell for hundreds of millions of dollars and carry an unmistakable style. Taste persisted even when craft was outsourced.

The AI generalization

Generative AI is the same move at civilization scale. By 2024 it was already possible to generate art indistinguishable from the Mona Lisa, and Glyman expects Hollywood-level video and songs to follow, produced by "just thinking of it." As making collapses toward zero cost and skill, what remains scarce is judgment: "The thing that's most in scarcity is going to be great taste, finding people who know what great looks like." The defensible inputs become concept, what is worth making, and taste, the ability to recognize what is great. This sits inside Glyman's broader view that AI tools are becoming as good as the best-run company on the planet and available in minutes, a great equalizing force that democratizes both creation and business-building, even as it remains hard to build a genuinely unique business worth creating. Democratized execution raises the premium on judgment rather than removing the need for it.

The same conclusion appears from a different angle in Alex Karp's remark that no one can scale the taste required to choose which problem to solve, a claim about enterprise problem-selection that parallels Glyman's claim about creative and product concept. Brian Chesky reaches for the same Da Vinci reference and a similar democratization argument, that AI gives everyone a paintbrush and canvas, from the angle of creative energy and identity rather than taste-scarcity specifically. All three converge on the same structure: craft commoditizes, taste does not.

Corroboration: the Midjourney campaign

Glyman offers a first-person production example from Ramp's own marketing, describing what he says was the first scaled out-of-home advertising campaign primarily powered by Midjourney.² Individual shots whose lighting, props, and emotional tone would previously have cost tens of thousands of dollars each, adding up to a multi-hundred-thousand-dollar campaign, were instead produced over the course of a day for about as much as it costs to add one software seat. His conclusion echoes the Warhol logic directly: "the cost of production of artistic and great work is probably as low as it's ever been in human history," which pushes the scarce inputs toward taste, meaning judgment about what is a striking image and what will connect emotionally, and toward curiosity as a hiring criterion.

The eleven-day Super Bowl ad: compression as the real product

A later instance, in late 2025 roughly eleven days before a Super Bowl broadcast, shifts the emphasis from cost to cycle time.³ The organizational precondition came first: roughly half a year earlier, much of Ramp's marketing had been moved to report to Chief Technology Officer Karim Atiyeh, on the explicit view that marketing was becoming an increasingly technical function, since understanding an audience and what it cares about at scale is a data and infrastructure problem, and machines could now generate serviceable video and audio.

When someone asked, eleven days out, whether Ramp should run a Super Bowl ad, Saquon Barkley agreed to appear but could film for only one hour on a Saturday. Concepting an ad normally consumes months of scripts, and it is difficult to know in advance how a script will look visually. Instead the team generated hundreds of variants in Midjourney, used Runway for motion and ElevenLabs for audio, and watched candidates rather than reading scripts, compressing what would have been months of work and research into days. The result, in Glyman's account, was the biggest month Ramp had ever had, notable because the company runs on a roughly month-long sales cycle. The refinement this adds to the concept is that when craft is commoditized, the gain is not only that images become cheaper; it is that a concept can be evaluated in its finished form before anyone commits to it. Taste stops being a prediction made from a script and becomes a selection made from among hundreds of realized options.

Open questions

If AI also becomes capable at problem-selection and concept generation itself, through better recommendation, market analysis, or taste models, it is unclear whether the taste moat erodes; Glyman and Karp are both betting it does not. Taste is also hard to credential and easy to claim, which means recognizing genuine taste may itself require taste, an unresolved circularity in the framework.