Localhost Competence
Boy is it beautiful.
“Mastery is the best goal because the rich can’t buy it, the impatient can’t rush it, the privileged can’t inherit it, and nobody can steal it. You can only earn it through hard work. Mastery is the ultimate status.” - Derek Sivers
I recently wrote an academic paper on computational linguistics. It had a clear thesis, rigorous math, a meticulous experimental framework, and lovely LaTeX formatting.
It was also almost certainly a big pile of steaming crap.
But boy was it beautiful.
A good friend and I have an ongoing conversation about the Sapir-Whorf hypothesis. This is the idea that human thought is shaped by language, that speakers of different languages literally think differently. The way this first came up is lost in the foggy mist of ancient friendships, but we’ve been going back and forth on this for years and always seem to find something new to say.
I had the blurry lines of an idea. Mechanistic interpretability researchers have been measuring geometric structures inside LLMs as manifolds that represent how concepts relate to each other. What if you could compare those structures across models trained on different languages? If a Mandarin-trained model and an English-trained model organize the same concepts differently, maybe you could actually measure linguistic relativity. Quantify it.
Naturally I turned to Claude and added more detailed math to my rudimentary version. This was all formalized into a cost metric so you can compare the geometric expense of moving from one concept to another inside a model. Then we designed an experimental framework using BERT models trained on single languages. I felt like I was earning the dopamine hit of discovery. It looked pretty real.
Except for the problems, which became more obvious in hindsight. The idea that model geometry maps to human cognition is not guaranteed. It’s baked in as an assumption, so at best I’m saying something about model geometry across languages and not how humans think. And the reasoning is circular. I’m trying to show that language encodes structure by examining models that learn structure from language.
Well, yeah.
The thing is, I have no expertise in this stuff. I couldn’t judge the merits of my own work. I didn’t have the resources to run the experiments, and even if I did, I couldn’t interpret the results.
I don’t actually know if the ideas are bad. That’s the point. I can’t tell.
LLMs are amplifying expertise and it feels amazing. But we’re using them to manufacture expertise at the same time. And we can’t always tell the difference.
Michael Crichton coined the term “Gell-Mann Amnesia”. He saw experts read a news article in their own field, spot every error and misunderstanding, then calmly turn the page and take the next article at face value. Expertise is narrow, but it builds false confidence. Your ability to spot errors in one field makes you believe you can spot them anywhere.
This localhost phenomenon is the inverse. I’m calling it Localhost Competence.
Gell-Mann is about falsely trusting your ability to evaluate. Localhost Competence is about falsely trusting your ability to produce.
When I use an LLM for a coding project, my actual professional expertise, I can evaluate, steer, and manage its output in ways a non-programmer cannot. I know when it’s wrong. I can dive in when it’s subtly wrong. Expertise provides mental guardrails that turn LLM output into genuine amplification.
But step outside that expertise and the guardrails vanish. The output looks just as polished. You just can’t tell what’s load-bearing and what’s decorative.
Experts aren’t going anywhere. If anything, they’ll end up on higher pedestals. They’ll also be inundated by people carrying beautiful, confident, and deeply flawed work produced in collaboration with machines that don’t know what they don’t know.
Isaac Newton, Grigori Perelman, Andrew Wiles all toiled in solitude and silence. They manifested intellectual fire from nothing. There’s more opportunity than ever to develop that kind of expertise, and more temptation than ever to skip it.
You’ve got an LLM in your back pocket now, which makes you think you can skip the solitude and get results on day one. You can produce things that look like expertise. But looking like it and having it are different problems, and only one of them compounds.
