A software engineer's analysis of how large language models intersect with the labor market has crystallized into a sharp argument: the widespread belief that AI can replace other people's jobs but not your own is not just optimism — it is a cognitive blind spot rooted in a failure to appreciate the intellectual labor behind any given output.
The Paradox at the Center
The argument begins with an observation about how people think about AI and their own work. When asked whether their entire job could be done by an AI, most respondents fall into one of two camps. Some believe AI cannot do their job but can assist with it. Others believe AI will eventually take every job, including theirs. What binds these positions together is a shared, contradictory assumption about everyone else.
Most people believe that AI can do the job of their cross-disciplinary peers — the designers, product managers, data scientists, or engineers in adjacent teams. But they do not believe AI can do their own specific role. The result is a population where the majority simultaneously holds that their own work is unique and irreplaceable while assuming everyone else's work is automatable.
This is the paradox. How can the majority believe their work is special while everyone else's is not?
The Steinmetz Story as an Illustration
The post uses an old anecdote to make the point concrete. When Henry Ford's electrical engineers could not solve a problem with a gigantic generator, he called Charles Steinmetz from General Electric. Steinmetz arrived, asked for a notebook, pencil, and cot, and spent two days and nights listening to the generator and making computations. On the second night, he asked for a ladder, climbed up, and made a chalk mark on the generator's side. The engineers removed a plate at that mark and replaced sixteen windings from the field coil. The generator worked perfectly.
Ford was pleased until he received a $10,000 invoice. He asked for an itemized bill. Steinmetz responded: one dollar for making the chalk mark, and nine thousand nine hundred ninety-nine dollars for knowing where to make the mark.
The story, the author argues, captures the essence of the problem. If you do not understand or truly appreciate someone's craft, you end up reducing them to their output. But the output is not the job. For a software engineer, the visible result might be a dozen lines of code that took days to produce, while a junior engineer churns out hundreds of lines quickly — even before large language models entered the picture.
Management, Measurement, and the Output Trap
The post turns to how management interacts with this dynamic. Management — the people two or three layers above a given worker — often operates from a completely different daily reality. When they look at the people beneath them, the question they ask is whether those people could be replaced by AI, and the answer is almost always that someone is actively working on it.
Management has always been interested in measuring output, which necessarily reduces a worker to their volume of production. Large language models are exceptionally good at producing massive quantities of output. The danger is that this amplifies management's existing inclination to evaluate work by its visible results rather than the thinking behind it.
A CEO may be evaluated on stock performance, but how much of that value is driven by their individual work? The enthusiasm for reducing work to a number tends to dwindle the higher you climb the chain, the post notes, even though the same measurement impulse persists.
What LLMs Cannot Replicate
The core limitation of large language models, the argument goes, is that they are trained on output — the final answer, the finished code, the completed document. They never saw the workings of the mind behind it. The iteration, the problem-solving sleep, the testing of ideas before they manifested externally, the tacit reasoning that shaped a solution — none of that exists in training data because it never left the thinker's head.
The author introduces the phrase "Intellectual Empathy" to describe what is missing. Without it, people tend to reduce everyone to their output and ignore the actual intellectual labor involved. When an expert uses a large language model in their own domain, the model's limitations become immediately obvious because the expert can distinguish between the intellectual labor and the generated result. But when someone steps outside their area of expertise, the model's output suddenly feels easy — not because it actually is easy, but because the observer only understands the output, not the thinking it replaced.
In this way, large language models become a Dunning-Kruger amplification machine. Every idea, good or bad, gets produced at high volume, and without domain expertise serving as a filter, there is nothing to separate quality output from noise. Expertise gives a person taste, and taste takes time and experience to develop.
The Practical Implication
The argument does not dismiss large language models as tools. They are powerful when wielded by experts who understand what the output should look like and can evaluate whether it meets the standard. The risk lies in management and individuals confusing the ease of generating output with the ease of doing the work that produced it. Everyone's job requires intellectual labor, even when that labor is not obvious from the outside — and forgetting that fact has consequences for how teams, organizations, and industries decide what to automate and what to preserve.