Sean Goedecke wrote a post in August with a claim in the title: LLMs reward expertise. It did 1,416 points, and the argument is worth more attention than the score suggests.
The claim runs against the received story, which says these tools democratise, so the junior can now do what the senior does and the gap narrows. Goedecke argues that the tools multiply what you already have, and multiplying a small number gives you a small number.
The evidence he uses
His main exhibit is a transcript of Terence Tao (one of the strongest living mathematicians) working with a model on a counterexample to the Jacobian Conjecture.
Watch how Tao works. His messages are short and blunt. He pushes back on suggestions. He ignores the model's advice about where to go next, makes his own leaps, and uses the model for the parts he chooses.
Goedecke's point is that the key to the technique is understanding the mathematics, so you can pull the relevant idea out of a multi-paragraph response.
A prompting workshop cannot teach that skill, because the thing being exercised is knowing which paragraph matters, and that knowledge belongs to the subject.
Why this inverts most AI training
Look at what companies are buying right now.
What is being bought | What the evidence suggests is needed |
|---|---|
Prompt engineering workshops | Deeper domain knowledge |
Tool access for everyone | The judgement to evaluate an answer |
Usage mandates and quotas | The confidence to say "no, simpler than that" |
Measured by adoption | Measured by whether output survives review |
If Goedecke is right, the left column is a category error. You are training people in how to operate the amplifier while the thing being amplified stays where it was.
The uncomfortable implication
This argument has an elitist reading.
If the tools reward existing expertise, they widen the gaps. The senior engineer with a good theory of the codebase gets faster. The junior without one gets output they cannot evaluate, which looks like productivity and accumulates as debt, because somebody eventually has to hold a model of that code in their head and it will not be them.
Juniors should still use the tools. The years of learning still have to happen, though, and treating the tools as a shortcut past them produces engineers who ship confidently and cannot debug.
What I would change tomorrow
If you run learning at a company that just bought everyone a licence, the highest-value thing you can do is unglamorous. Keep teaching the fundamentals, and change what you assess.
Assess how the person defends the artefact, since the artefact itself is free now. Put them in a room, ask why this approach and not the obvious alternative, and ask what breaks if the input doubles. Four minutes of those questions show whether they hold a model of the code in their head.
Then keep doing the boring thing: reading, practising, being wrong in front of someone who can correct you.
The post is here. Read the Tao transcript he links, and watch how often the human says no.
One email when we publish. Research, product decisions, and what teams report back.
