Despite the breathtaking pace of improvement in AI, there’s still plenty of skepticism around the potential impact of AI on our economy and technological progress. This fugue explores why sentiments like “yeah, but what can an LLM do besides writing code?” and “AI is improving in domains with verifiable rewards, and my discipline (experimental chemistry, biology, physics, whatever is your favorite) does not have well-defined rewards, so it’s not going to change much” are deeply, fundamentally, maybe even epistemically misguided.

All three voices return to the same theme: people systematically underrate software because they confuse writing code with a narrow technical activity rather than with a general instrument for changing systems, a method for reducing the time between hypothesis and verification, and a fertile ground for articulating epistemic guardrails whose validity is domain-transcendent.

The first voice is the most concrete and it shows how even an ugly nested for loop can have million-dollar implications for lifetime earnings.

The second voice broadens out and argues that software engineering is not limited to actually writing code, but is instead a craft that forces habits of decomposition, guardrails, explicitness, and iteration that are weirdly rare outside CS and increasingly useful everywhere. Practically, it demonstrates why you should still study programming even if AI will write all the code.

The third voice is the most general. It argues that success in many domains is downstream of the number of iterations you can afford, and if your domain doesn’t have clearly articulated verifiable rewards, then it’s not just an obstacle to AI progress; it very likely is a blocker for human progress too.