The internet's favorite lie about AI just took a hit
For two years the pitch has been: LLMs collapse the skill gap. Anyone can code now. Anyone can do design, legal, medicine — just ask the robot.
A post that hit 900+ points and 377 comments on Hacker News this week calls that out for what it is: half-true and dangerously incomplete. Sean Goedecke's argument, stripped of the hedging: the model isn't the bottleneck. You are. And the size of that bottleneck is set almost entirely by how much domain expertise you're bringing to the conversation.
This isn't a hot take. It's the most obvious thing in the world once you've watched two different people use the same model on the same problem and get wildly different outputs.
Exhibit A: Terence Tao doesn't talk to your ChatGPT
Goedecke's central example is a Terence Tao transcript — the guy solving a counterexample to the Jacobian Conjecture in a ChatGPT conversation. Tao's prompts are short. Almost terse. He doesn't explain what a Jacobian Conjecture is, doesn't ask the model to "think step by step," doesn't stack fifteen paragraphs of context.
He does something much harder to fake: he reads the output, immediately spots when it's unnecessarily complicated, and redirects with a single line that only makes sense if you already know where the proof should be going.
That's not a prompting technique. There's no template for it. Goedecke's line is the one worth stealing:
"The human is the bottleneck, not the model, because the difficult part is in communicating to the model exactly what kind of solution the human wants."
You cannot communicate what you don't understand. No amount of "act as a senior engineer" system prompt fixes that.
Exhibit B: the HN thread accidentally proved the thesis
The comments are where this gets good, because commenters tried to argue the opposite and kept failing.
One top comment describes running an experiment with a non-technical friend trying to build software with AI assistance. She got stuck — not because the model refused to write code, but because she didn't have the vocabulary to ask for what she wanted. She'd drift into vague feature discussions instead of scoping a deliverable, because scoping is a skill, not a UI affordance.
Another commenter, further down, dropped this on engineers who think AI makes juniors obsolete-proof:
junior engineers using AI to "fix" code without internalizing patterns, making the identical mistakes a year later.
That's the actual risk nobody's pricing in. It's not "AI takes your job." It's "AI lets you skip the rep that builds the judgment that was going to make you good," and you don't find out you skipped it until the bill comes due on a system you can't debug.
And then there's the number that should be pinned above every "vibe coding" tutorial:
20-50 lines of code for every 1 line kept.
That's not a knock on AI code generation. That's a description of what expert-level iteration actually looks like — rapid branching trial-and-error, discarding almost everything, keeping the 2% that's right. You can only run that loop fast if you already know what "right" looks like the moment you see it. A novice can't tell the difference between confident garbage and a real solution, so they either accept the first plausible-looking output or drown in options they can't rank.
What expert prompting actually looks like
Forget "prompt engineering" as a discipline. Compare these two prompts for the same bug:
Novice:
"My React app is slow, can you make it faster?"
Expert:
"This component re-renders on every keystroke in the search box
because `filteredResults` is recomputed inline and `onSearch` is
a new function reference each render, busting the child's memo.
Give me a fix that keeps the debounce behavior in useSearchDebounce
but doesn't require lifting state out of SearchPanel."
The novice prompt gets you a generic essay on useMemo and React.memo — technically correct, practically useless, because it doesn't know your constraints. The expert prompt gets you a targeted diff in one shot, because the human already did 90% of the diagnostic work and only outsourced the typing.
That gap doesn't close as models get smarter. It gets wider, because smarter models can act on more precise instructions — and precision is exactly the thing novices can't supply.
The uncomfortable takeaway
If you're an engineer who's been quietly panicking that AI is erasing the value of what you know: it's the opposite. Deep expertise is now a bigger competitive advantage than it was three years ago, because the payoff for having it compounds with every model upgrade instead of decaying.
If you're a team lead handing juniors a Copilot license and calling it mentorship: you're building people who can produce code they can't defend. That's not a training program, that's technical debt with a delay timer on it.
And if you're betting your product on "no-code, no-expertise, just prompt it" — Goedecke's post, and 377 comments of engineers agreeing with it, is your market research. The tool got more powerful. The floor didn't rise nearly as much as the ceiling did.
Go build expertise. The models are waiting to reward it.
Discussion thread: Hacker News, 900+ points · Original post: seangoedecke.com
Top comments (0)