Frontend development is where AI hits a wall. A piece from nerdy.dev argues that LLMs struggle badly with UI work. Visual consistency and state management trip them up. And they lack the aesthetic judgment that good interfaces demand. Good taste is becoming the only real moat left. The Hacker News discussion added a useful insight: your perception of AI's coding ability says more about your own skill level than the tool's actual capability. Experts spot the flaws because they understand what good output looks like.
This isn't a frontend-specific problem. AI writes decent boilerplate and handles rapid prototyping well. It falls apart on production complexity everywhere. Frontend just exposes the gap faster because you can literally see when something's wrong. A broken layout is obvious. A poorly structured API endpoint isn't.
The industry's fix is to give AI eyes. Anthropic's Computer Use feature for Claude 3.5 Sonnet and Cognition AI's Devin now render code in browsers, screenshot it, and visually debug their own output. Instead of guessing at CSS probabilistically, these agents spot misalignments and layout issues through vision models. It's a feedback loop that pure text generation can't provide.
None of this replaces human developers. The HN consensus from people shipping real code is straightforward. AI accelerates prototyping. Turning that prototype into something production-ready still demands someone with actual skills. Imbue's testing swarm approach demonstrates one path forward, but the tools are getting better at seeing. They still can't judge what looks right.