Programming with LLMs is genuinely useful and genuinely destabilizing. At Pydantic, we build tools for LLM-powered software reliability, yet we're experiencing a weird time. This isn't about AI replacing programmers—it's an honest account of developer fatigue.

When I learned to code, I felt I could shape the universe's fabric. That creative dopamine is what low-code tools promised but never delivered. Now the gap has narrowed, but the human reviewing and directing feels worse, not better. My colleague Douwe, who maintains Pydantic AI, described waking to thirty PRs pulled overnight by AI, facing the temptation to delegate review itself—"at that point, what am I still doing here?" I've spent two full days writing plans for an LLM, only to watch it invent components that don't exist. The models produce plausible code but lack coherent intent across complex changes, creating a "fatigue of supervision."

Simon Willison highlighted a Berkeley Haas study showing AI increases work intensity—the pull of "one more prompt." Marcelo joked about running five Claude sessions simultaneously. The parallelism is exhilarating but feral: we can start more things but can't thoughtfully finish more, because that requires the one unparallelizable resource—our brains. This is a "human reward function problem": LLM-assisted coding automated the satisfying work and replaced it with cognitive load of review. The feedback loop is broken.

It's also lonely. Natural moments of collaboration get replaced by another prompt. In teams without strong collaboration culture, this chills communication. And it's addictive: sometimes brilliant, sometimes garbage, a textbook Skinner Box. Switching between LLM-assisted and manual work is jarring, requiring maturity to permit the switch.