How to Research Technical Topics With AI

Most people ask a chatbot how something works and stop there. The author of a new post on their personal dev blog did something slower and more deliberate: they used AI throughout the process of learning how large language models actually work, but refused to let any single answer stand in for understanding. The result was an article called How LLMs Actually Work, and the process behind it is more interesting than the piece itself.

The first move was narrowing the question. "How do LLMs work" is the kind of prompt that produces a confident, sprawling, ultimately useless tour through tokenization, training, hardware, and alignment all at once. Instead they picked one slice: what happens between a typed prompt and the model predicting its next token. That gave them a start point, an end point, and a natural order to follow — tokenization, then embeddings, then positional encoding, then attention, then the generation loop. Everything else, including training, got set aside as a separate problem.

From there the AI became a research assistant, not an oracle. They'd ask what a concept depended on, what terminology they were missing, which papers or docs or code were worth opening — and then they'd actually open them. A phrase like "how does the model know word order" led them to positional encodings, RoPE, and relative position methods, terms that made ordinary search far more productive. But the model's job stopped at pointing; verifying the claim against the original source was always on them.

The most useful trick, by their own account, was asking AI to attack their explanations rather than rewrite them. With multi-head attention, the easy mental picture — each head gets a slice of the token vector — is technically wrong; each head actually uses its own learned projection matrices. Asking "what's wrong with my explanation" instead of "explain this to me" forced them to locate and repair the actual misunderstanding instead of just swapping in someone else's phrasing.

Only after they could reconstruct the whole mechanism on paper, without the chat open, did they let AI touch the prose. At that point the model got their notes, their ordering, and their intended reader — not the messy research conversation itself, which they point out is full of dead ends and half-right guesses that shouldn't survive into a published explanation. A separate fact-check pass at the end pulled every claim back out of the draft and checked it against the original sources one more time.