LLMs are general. Your library isn't.
Sean Goedecke argues that LLMs reward expertise. He's right — and it means the things you've already saved are worth more than you think.
Sean Goedecke published an essay with the argument in the title: LLMs reward expertise.
The setup is that everyone now has the same models. Same weights, same training data, same chat box, roughly the same monthly fee. So it's tempting to conclude that this is the great leveler — that the distance between someone who knows a field and someone who doesn't has finally been closed by a subscription.
Goedecke thinks that's backwards. His evidence is a transcript.
What the transcript shows
The transcript is Terence Tao working through a mathematical problem with ChatGPT. What's striking about it isn't the model's output. It's Tao.
His messages are short. He doesn't take the first suggestion. When a proposed direction looks more complicated than it ought to be, he says so and turns the conversation somewhere else. He treats the model less like an oracle and more like a capable colleague who needs managing.
None of that is a prompting trick. You cannot copy Tao's messages and get Tao's results, because the messages aren't doing the work — the judgment behind them is. He can tell when an answer is good. That single capability is what converts a general model into a useful one.
Goedecke makes the same point about software, and his phrasing is the part worth keeping: what helps you get more out of a model is "a good theory of your codebase," not a better grasp of software design in the abstract. Specific beats general. The concrete thing you actually know beats the principle you could recite.
His conclusion: the human is the bottleneck, not the model. The information is already in there. Getting it out is the skill.
We think that's right. We also think it has a second half.
Generality is an average
Here's the thing about a general model. General means averaged.
A frontier model has read more about your field than you will in your life. It has also read everything wrong that was ever written about your field, weighted more or less by how often it got written down. It holds no opinion about which of those things is true. Ask it a question cold and you get the center of mass — the consensus position of everyone who has ever published on the subject.
As a default, that's extraordinary. In the specific case where you actually know something, it's the wrong answer.
Because the consensus is what you already believed before you started reading. The entire reason you worked through forty articles on the topic is that you were hunting for the six that were better than the consensus. That's what reading is for.
And the model cannot tell you which six those were. Not because it hasn't read them — it has, along with the other thirty-four — but because nothing in its training run recorded that you found them convincing. It has the content. It's missing the weighting.
Your library is the weighting
Which is exactly what a saved library is. Not a pile of documents. A record of judgment.
Every save is a small expert act. You decided this one was worth keeping and the other thirty tabs weren't. You did that a few hundred times, mostly without thinking about it. Any individual decision is trivial. In aggregate they're a fairly precise map of what you believe, what you think matters, and what you've already considered and rejected.
That map exists nowhere in the model. It can't. It was built from everyone's material, and the averaging is the point — it's what makes the thing work at all for the millions of people who aren't you.
So when we say connecting your library makes your AI better, we mean something narrower and more defensible than the usual version of that claim.
It doesn't gain knowledge. It gains your priors.
It stops answering as the average of everyone who ever wrote on the subject and starts answering from the material you already decided was worth your time. That is most of the practical distance between a generalist and a specialist. The specialist isn't running better hardware. They know which sources to trust.
The bottleneck is bandwidth
Goedecke says the human is the bottleneck. Look closer at the mechanism and it's narrower than that: the bottleneck is bandwidth.
You know things. A little of it is in your head; most of it is in material you read once and half-remember. In a chat window you can transmit maybe a few hundred words of that before the effort exceeds the payoff and you let the model fill in the rest. Everything you fail to transmit gets filled in from the average.
That's why the first answer is always so generic. Not because the model is limited. Because the channel was.
A connected library changes the economics of that channel. You stop typing your context and start pointing at it. Ask for something "based on my Takt library" and the assistant pulls the relevant excerpts itself — the six articles, not the forty — and works from those. The expertise you spent months accumulating stops being locked in a folder you never reopen.
Two honest limits
The strong version of this claim is false, so let's mark where it stops.
It doesn't manufacture expertise you don't have. If you saved things you didn't understand, connecting them transmits your confusion at higher fidelity. Curation is only signal when it reflects real judgment. A library of things you meant to read is a library of nothing.
It doesn't do the steering. Tao's advantage wasn't only that he knew which sources were good. It was that he could tell, in the moment, when the model was heading somewhere unproductive, and had the standing to say so. A library lowers the cost of supplying context. It does not lower the cost of thinking, and anyone selling you that is selling you something else.
What it does is stop you beginning every conversation from zero.
The part that compounds
Expertise takes years, which is what makes Goedecke's essay slightly bleak if you're not already an expert. You can't shortcut it and you can't buy it.
A library isn't like that. It starts paying out in the first week, it compounds every time you save something, and unlike your memory it doesn't decay. The article you read in March is exactly as available in August as it was the day you saved it — which is more than can be said for the version in your head.
That's a strange kind of asset. Most things that compound require you to keep doing something hard. This one requires you to keep doing something you were already doing, and then makes it retrievable.
The part that cares
The model's generality is its greatest asset and its only real limitation. It has read nearly everything and cares about none of it. It has no stake in which answer is right, no memory of what worked for you last time, no sense of which of the forty articles was the one that changed your mind.
You're the part that cares. Your library is where the caring got written down.
Goedecke is right that the human is the bottleneck. But a fair amount of what makes you good at your work isn't in your head at all — it's in the trail of things you decided were worth keeping. That's not locked away any more. It's context now.
If you want to connect your library, we wrote the setup instructions in the previous post — Claude, ChatGPT, Cursor, Codex, and Grok, about a minute each. Or start a library and give it something to work with.