Why AI Is Bad at Design

Why OpenAI is merging Codex and ChatGPT and the future of knowledge work | Andrew Ambrosino

Topic 5 of 17

Published Jun 28, 2026 · Clip 12:06–16:50 (4:44)

Andrew gives practical reasons frontier models lag at design: design is harder to grade than code — building a training loop for good vs bad design is tedious compared to 'does the code compile,' and human taste is part of the feedback signal. Labs historically invest in capabilities that accelerate AI research; correct code sits in that flywheel, design doesn't — though he expects models to get pretty good at design as those reasons fade. Murkier problems remain: design has a cultural/novelty component (a model that always outputs Linear's website isn't the challenge; software wants known patterns, design needs randomness), plus a deep abstraction layer between visual design and code semantics — a rebrand shouldn't mean updating 263 components one by one. The host wraps by noting human creativity and new paradigms like the Codex app's novel chat-that-codes shape remain human strengths 'for now.'