AI evaluation has been an engineering problem. Setor Zilevu, speaking at Figma's Config 2026, argues it no longer belongs there alone. As AI capabilities commoditize, the teams pulling ahead are the ones measuring quality against actual human needs, not just benchmark scores.
The core claim is structural: designers and researchers should own the evaluation loop, not inherit results from it. Zilevu frames evals not as a QA step but as a design artifact, something built from human-centered definitions of what good output actually means. That framing shift is the argument worth sitting with.
The talk is worth watching in full for how it operationalizes that shift. What does a designer-led eval process look like in practice? How do you encode 'good' into something repeatable? Those answers are in the details Zilevu walks through, and they have direct implications for how AI product teams structure their workflows going forward.
[WATCH ON YOUTUBE →]