Stop Outsourcing to AI. Start Learning With It.

The conversation around LLMs is almost entirely about delegation. What can the model do for you? It writes your code, drafts your reports, summarizes your meetings. But a small group is asking the opposite question: what if you used AI to learn the things you don’t know, instead of handing them off? That inversion is worth a closer look.

A quick note of honesty first. There was very little fresh community chatter on this topic over the past 30 days. So this piece leans less on the week’s headlines and more on a trend that has been quietly building, plus my own read on where it goes.

The LLM as a “skip button”

The most seductive feature of an LLM is that it lets you skip. Don’t know regex? It writes the pattern. Don’t know statistics? It runs the analysis. Measured purely on efficiency, it feels like magic.

The catch is what skipping leaves behind. Whatever you skipped, you still don’t know. Hit a similar problem next week and you’re back at the prompt, asking again. You’ve turned your own work into a permanent outsourcing contract.

One camp shrugs at this. The knowledge is obsolete, they say — we stopped doing mental arithmetic after the calculator, so why memorize what the machine handles? There’s a real point in there. But there’s also a trap. When you offload the doing, you often offload the ability to judge whether the output is right.

Lathe: a small counterexample

One project pushed back against this current. It surfaced in developer circles under the name “Lathe,” and the idea behind it is simple. Instead of spitting out an answer, it walks the user toward understanding the thing themselves.

The name is not an accident. A lathe is a machine tool — you press rough material against it and shave it down into shape. That’s the whole point. You don’t receive a finished answer. You take a crude understanding and grind it into something precise, a little at a time.

This is a meaningful shift. The AI moves from being the one who answers to the one who teaches. Same model, same weights — but change the user’s intent and the output changes completely.

Why this matters now

Three reasons.

First, verification. AI output is often confidently wrong — we call it hallucination. If you know the fundamentals, you catch the bad answer. If you don’t, you ship it. The gap between the person learning and the person delegating opens up right here.

Second, transfer. What you delegate ends when the task ends. What you actually learn carries forward — to the next problem, into a different field. Learn it once and it compounds for the rest of your career.

Third, leverage. Keep skipping the things you don’t understand and the core of your own work ends up in the grip of an external tool. It’s comfortable. It’s also a dangerous kind of dependence.

How to use an LLM to learn

You don’t need a special tool. You need a different prompt.

Instead of “write this for me,” try “explain step by step why this works.” Once you have the answer, flip it around: “Check whether I actually understood this correctly.” Make the model grade you.

When an unfamiliar concept shows up, resist grabbing the answer. Ask it to “explain this to a beginner with an analogy.” The same LLM becomes a tireless private tutor — one that works around the clock and never gets annoyed when you ask the same question for the hundredth time.

The principle is one line. Don’t consume answers. Produce understanding.

The takeaway

The real divide in the AI era isn’t whether you use AI. It’s whether you use it to do the work or to learn the work. The first buys you convenience. The second buys you capability. Run that forward a year and the two arrive in completely different places.

Think about the last question you typed into an AI. Was it a question that handed off the work — or one that pulled you an inch closer to understanding it? That small difference might be the thing that separates you from yourself, twelve months from now.