How it works

You've probably used ChatGPT, or Claude, or something like them. Type a question, get an answer that reads like it was actually thought through. It's natural to wonder what's happening in the half-second between the two.

The honest answer is: the same four steps, over and over. Every one of those systems — GPT, Claude, Gemini, Llama, nearly everything since 2017 — is some size of the same idea, called a transformer. It's even in the name: GPT stands for Generative Pre-trained Transformer. One architecture, run at wildly different scales.

The four steps

Strip away the scale — the billions of parameters, the dozens of stacked layers, the unimaginable amount of text it was trained on — and a transformer is doing four things, in a fixed order, for every word it produces:

• Read. Turns words into positions on a map, where things that mean similar things sit close together.
• Look Around. Figures out which words relate to which other words, and blends them together based on how strongly they relate. This step has a real name in the research — attention— and it's the part that made transformers work at all. Before attention, models mostly read left to right and forgot the beginning by the time they reached the end. Attention lets every word check every other word directly, no matter how far apart they are.
• Think. Takes the blended idea from Look Around and refines it using a fixed rule — the same rule, every time, for every word.
• Answer. Lines up every possible answer and gives each one a confidence score. The most confident answer wins — but it's never the only option considered.

That loop runs once per word out — GPT and Claude just run it through many stacked layers at once, at a scale that takes a data center to hold, and repeat the whole thing for every word they generate. It's the same four steps as many times as it takes to answer you, not a different process underneath.

The part almost nothing else shows you

Most explanations of AI stop at an analogy — a brain, a very good autocomplete, a black box with sparks in it. YDABlocks doesn't, because the four steps above aren't a metaphor for what the Builder does. They're a description of it. The Builder is not a simulation of a transformer. It's a small one, actually running, at a size you can watch — the same math a real one runs, on a sentence you choose, with every step's real output shown on screen rather than hidden inside a model with billions of parameters.

Tap a block below to see exactly what it does — the same four, always in this order.

What you can do once you've built one

A build on its own already answers the question this page opened with. Past that, there's more: Training Pairs let you teach it your own examples instead of the ones it ships with; Fine-Tuneadjusts how much weight Look Around and Think each get in the blend, and shows you the before-and-after; and once you've built and taught something, Richard— your YDABlocks partner — can talk through why your build answered the way it did. He's not a tutor and not a general-purpose chatbot; he's trained on YDABlocks itself, to help you understand your own build.

None of this is required to get the point. The free build is the whole idea, not a trial of it — everything past that is for going deeper, on your own terms.