The silence when someone asks how AI actually works

Yuri spoke to a recruiter who interviews AI engineers every week. The description was blunt: everybody builds agents, everybody wires up systems that search for information before answering. Almost nobody can explain why any of it works. Ask why text gets cut into pieces before it is stored for search, why the size of those pieces changes the quality of what comes back, or what happens inside the model before the first word shows up on screen, and the room goes quiet.

Yuri recognised himself in that room. He has used these tools long enough to trust them, which is not the same thing as being able to explain the engine underneath. This time he decided not to let it slide. He set aside a whole session, took every question the recruiter described, and went looking not for the textbook answer, but for the mechanism.

The pattern he found is simple. The answer backed by real experience always has the same shape: it names the mechanism and says what it costs in production, with real users and real deadlines. The answer from someone who has only used the tool describes what the tool does. The answer from someone who can speak for the whole system describes what happens when the piece breaks, and what that breakage costs.

Take the size of the text chunk. Most people explain that large documents need to be split before they go into storage. End of answer. But there is a reason the size matters: a chunk that is too big mixes several topics into a single numerical summary. That summary ends up matching nothing very well when someone searches for a specific fact. In published tests, shrinking the chunk from over a thousand fragments of text to just over sixty improved search accuracy by ten to fifteen percentage points, specifically on documents packed with names and figures. Citing a number like that shows someone tested several configurations before picking one. Saying they use chunks of five hundred because that is the standard is just repeating a number someone else measured, in a different context, for a different problem.

The same goes for why an answer takes a moment to start appearing. There is a phase where the model processes the entire request at once, before writing a single word. That phase is limited by how much computing power is available right then. Then comes the phase of writing word by word, which is limited by how fast the model can read its own memory. These are different bottlenecks with different fixes. One published benchmark showed the initial wait climbing from under half a second to just over two seconds simply by increasing the size of the request. Understanding that split explains why very long requests make a system feel stuck at the start, even if it writes quickly once it gets going. Not understanding it makes people call the system slow and switch models without fixing the root cause.

Then there is the reason search by meaning works at all, perhaps the most overlooked piece. The system does not store words. It stores a set of numbers representing the meaning of that text, in a space where similar concepts sit close together. That is why typing "I forgot my password" surfaces an article titled "how to reset your login", with not a single word in common. It is also why switching the model that generates those numbers forces a full reprocessing of everything already stored. The old numbers live in one space, the new ones in another. Comparing them is like comparing metres with kilograms, making any transition expensive and slow.

Yuri came out of that session with a clear conclusion. Tools change their names every year. Models go out of date in months. The reason behind each decision, why this chunk size, why this delay, why this search finds what it finds, does not age at the same pace. That reason is what survives the next update. It separates someone who used the tool from someone who can speak for the whole system.

This explains why recruiters ask the way they ask. They are not checking whether a candidate operates some popular open-source library well. They are checking whether, with that library removed, the person could still put the same thing together using raw instructions and explain why each piece is there. It is a test of mechanism, not of tooling. The tooling will have a new name next year.