hawkes.aiKey ideas
- The one idea — The full pipeline on one map — tokenizer → embeddings → stacked blocks (attention + MLP) → head → logits → sampler → loop — and this series climbs it station by station.
- How it is shown — The factory line traversed end to end in one continuous move.
- The trap to avoid — Hunting for "the intelligence" in one box — no single station is the mind; the loop around all of them is.
- What it sets up — Each station named with its act number — the map doubles as the series index.
Zoom into any part of an AI and you find plain arithmetic. There is no box labeled intelligence — the mind is what the whole loop does at speed.
The one idea
The full pipeline on one map — tokenizer → embeddings → stacked blocks (attention + MLP) → head → logits → sampler → loop — and this series climbs it station by station.
This is the whole machine, on one screen. Every episode of this series lives somewhere on this line. Station one: the tokenizer. Your text becomes chunks — tokens — each swapped for a number. It never sees letters again, and that blindness alone explains a whole family of famous failures. Act One. Station two: embeddings. Every token becomes a point in a space where meaning is geography — similar things sit near each other, and you can literally do math on words. Still Act One. Station three, the heart: a tower of identical floors — transformer blocks.
How it works — the demo
The factory line traversed end to end in one continuous move.
Each floor does two things. Attention: every token looks at every other and decides what matters. Then a bank of learned circuits where knowledge lives. Stack that floor dozens of times and understanding accumulates. Acts Two and Three. Station four: the head — one score for every token in the vocabulary. The raw scores are the logits; you'll hear that word constantly from here on. Station five: the sampler rolls the weighted dice and picks. Station six is just a conveyor: glue the pick onto the input and run the line again. That circle — not any single station — is what's generating your answer.
The trap to avoid
Hunting for "the intelligence" in one box — no single station is the mind; the loop around all of them is.
Why it matters — and what’s next
Each station named with its act number — the map doubles as the series index.
The trap: hunting for the box labeled "intelligence." Zoom into any station and you find arithmetic — embarrassingly plain. The mind isn't in a box. It's what the whole circle does at speed, with the right numbers inside. And the map keeps going. How the numbers got good: training. What the line runs on, and why it costs billions: hardware. What we bolt around the loop to make agents, memory, and things we can trust: the whole back half — ending inside two systems we built ourselves. Screenshot this map — it's the index of everything coming. Six stations, one loop. Next: why the smartest people you know look at this same machine and walk away telling opposite stories.
This is one short episode in AI: Zero → Frontier, a step-by-step climb through how AI actually works. Each episode builds only on the ones before it.
Full transcript 2:30 of narration
This is the whole machine, on one screen. Every episode of this series lives somewhere on this line.
Station one: the tokenizer. Your text becomes chunks — tokens — each swapped for a number. It never sees letters again, and that blindness alone explains a whole family of famous failures. Act One.
Station two: embeddings. Every token becomes a point in a space where meaning is geography — similar things sit near each other, and you can literally do math on words. Still Act One.
Station three, the heart: a tower of identical floors — transformer blocks. Each floor does two things. Attention: every token looks at every other and decides what matters. Then a bank of learned circuits where knowledge lives. Stack that floor dozens of times and understanding accumulates. Acts Two and Three.
Station four: the head — one score for every token in the vocabulary. The raw scores are the logits; you'll hear that word constantly from here on. Station five: the sampler rolls the weighted dice and picks.
Station six is just a conveyor: glue the pick onto the input and run the line again. That circle — not any single station — is what's generating your answer.
The trap: hunting for the box labeled "intelligence." Zoom into any station and you find arithmetic — embarrassingly plain. The mind isn't in a box. It's what the whole circle does at speed, with the right numbers inside.
And the map keeps going. How the numbers got good: training. What the line runs on, and why it costs billions: hardware. What we bolt around the loop to make agents, memory, and things we can trust: the whole back half — ending inside two systems we built ourselves.
Screenshot this map — it's the index of everything coming. Six stations, one loop. Next: why the smartest people you know look at this same machine and walk away telling opposite stories.