Taste as Moat

Everything scales except the judgment

Capital can be raised. Models improve and get cheaper on their own schedule. The one input that refuses to budge is the judgment of which business problem is worth solving, and Alex Karp has compressed the whole competitive question into it. "It's taste plus money. And there is no AI, you can't scale the taste of what is the business problem you want to have to solve."¹ [5:01] Taste, in his formulation, is that judgment plus the ability to configure AI correctly against it. Everything else in the stack, he argues, commodifies on contact; taste is where the advantage now lives.

He applies it at three layers. Problem selection first, deciding which problems are worth attacking at all: a report on Chinese GDP is easy and disposable, while encoding how a company actually underwrites risk, drills for oil, or manages a supply chain is hard and defensible. Deployment configuration second: what stays on-prem, what goes into the ontology, which data gets protected and which gets deliberately exposed to bring more data back. Casting third: who gets placed where, and how the work is organized.¹ [17:07]

The ontology executes; the taste decides

Palantir is the worked example, and the tell is what the ontology does not hold. It executes decisions taste has already made.¹ [17:16] The structural expression is a three-code typology Karp gave in public for the first time here. Primitives: infrastructure-grade hardcoded logic that encodes deep domain knowledge and would take millions of technical hours to rebuild ("how do you build a steel beam?"). FDE code: written by forward-deployed engineers and managed on the platform as product. Free code: the LLM-generated dashboards and one-off analyses that are "magical" and almost right, and the source of the "one more dashboard" addiction (the comparison he used onstage was porn).¹ [10:00-11:11] The durable moat sits in the primitives. The magic, and the trap, both live in the free code.

Everything can be copied except the twenty-five years

The strongest evidence is the copy test. Competitors can copy the look of Palantir's code, the forward-deployed model, even parts of the ontology, and still not copy the twenty-five years of judgment embedded in the people and the deployments.¹ [16:04] The charisma asymmetry is the second tell. The frontier LLM companies are "super charismatic with investors" and, in Karp's telling, "super not charismatic with enterprises," because a magical demo is not the same as knowing which enterprise problem is worth pointing it at.¹ [7:27] And the failure mode when taste is absent already has a name in his vocabulary, tokenmaxxing in the pejorative sense, the tool-shaped objects that manufacture the sensation of work without touching a real problem. "It feels productive to have every email classified."¹ [4:46]

The moat is only legible in the win

The framework's own limit is that Karp treats taste as something to select for, not something to teach. The people who have it get found, and more of it cannot be manufactured by adding headcount, which is precisely why it stays scarce.¹ [16:04] That makes it a hiring filter more than a transferable playbook, and it makes the claim awkward to falsify, because taste is easy to assert once the outcome is already known. Karp half-concedes the problem when he insists the credibility of having taste is also required, and that a lot of places "don't really understand how unpopular they are in enterprise."¹ [17:39] The circularity is real. Everyone who wins claims they had taste all along, and the framework offers no instrument for catching them.