hawkes.aiKey ideas
- The one idea — Nearly all model compute — layers, attention, everything — is multiplying big grids of numbers.
- How it is shown — Rows meeting columns: the dot product from EP 026 revealed as one cell of a giant grid; the tower's every stage recast as grids.
- The trap to avoid — Expecting exotic math at the bottom — the exotic thing is the scale and arrangement, not the operation.
- What it sets up — Build a machine for the one operation.
At the bottom of every AI there's no exotic genius math. It's multiply and add, the arithmetic you know, arranged at a scale that breaks intuition.
The one idea
Nearly all model compute — layers, attention, everything — is multiplying big grids of numbers.
The chips, the datacenters, the market frenzy — the entire boom runs on one operation. Rows meet columns, multiply, add up. Matrix multiplication. That's the show. Everything else is how much, how fast, and how cleverly arranged. And you already know it. One cell of the grid: a row against a column, multiply the pairs, add them up. That's the dot product — episode twenty-six, the neuron's whole job, the ritual that's followed us all series. A matrix multiplication is just dot products by the million, marching in formation. Now recast the tower you know. A layer of neurons — episode twenty-nine — is one matrix multiply.
How it works — the demo
Rows meeting columns: the dot product from EP 026 revealed as one cell of a giant grid; the tower's every stage recast as grids.
Attention, scoring every word against every word — episode thirty-six — matrix multiplies. The vaults where facts live: matrix multiplies. Peel the costumes off every stage and the same geometry looks back. It's grids all the way down. Here's the scale. Rule of thumb: generating one token costs about two operations per weight. A seventy-billion-parameter model spends roughly a hundred and forty billion operations — on every single token it utters. Say a sentence, that's trillions. The boom's absurd arithmetic bill, and the reason the last episode's interns exist at all. And when one operation is ninety-nine percent of the bill, silicon reorganizes around it. Modern GPUs grew dedicated organs — tensor cores — that swallow small grids whole, matrix multiplication executed as a single hardware act.
The trap to avoid
Expecting exotic math at the bottom — the exotic thing is the scale and arrangement, not the operation.
Why it matters — and what’s next
Build a machine for the one operation.
Each generation, those organs claim more of the chip. A modern GPU is a matmul machine with accessories. The trap: expecting exotic math at the bottom. People peer under the hood, find rows-times-columns, and feel cheated. Don't be. The operation being simple is the profound part — everything this series has shown you, the understanding, the reasoning, the taste, emerges from arithmetic a teenager can check, arranged and scaled past all intuition. So the machine's job description is one line: move numbers to the multipliers, multiply, move the results back. So the machine itself is three things — the multipliers, the storage, and the road between. Watch the loop closely: the moving already looks slower than the math. Next: inside a GPU.
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
The chips, the datacenters, the market frenzy — the entire boom runs on one operation. Rows meet columns, multiply, add up. Matrix multiplication. That's the show. Everything else is how much, how fast, and how cleverly arranged.
And you already know it. One cell of the grid: a row against a column, multiply the pairs, add them up. That's the dot product — episode twenty-six, the neuron's whole job, the ritual that's followed us all series. A matrix multiplication is just dot products by the million, marching in formation.
Now recast the tower you know. A layer of neurons — episode twenty-nine — is one matrix multiply. Attention, scoring every word against every word — episode thirty-six — matrix multiplies. The vaults where facts live: matrix multiplies. Peel the costumes off every stage and the same geometry looks back. It's grids all the way down.
Here's the scale. Rule of thumb: generating one token costs about two operations per weight. A seventy-billion-parameter model spends roughly a hundred and forty billion operations — on every single token it utters. Say a sentence, that's trillions. The boom's absurd arithmetic bill, and the reason the last episode's interns exist at all.
And when one operation is ninety-nine percent of the bill, silicon reorganizes around it. Modern GPUs grew dedicated organs — tensor cores — that swallow small grids whole, matrix multiplication executed as a single hardware act. Each generation, those organs claim more of the chip. A modern GPU is a matmul machine with accessories.
The trap: expecting exotic math at the bottom. People peer under the hood, find rows-times-columns, and feel cheated. Don't be. The operation being simple is the profound part — everything this series has shown you, the understanding, the reasoning, the taste, emerges from arithmetic a teenager can check, arranged and scaled past all intuition.
So the machine's job description is one line: move numbers to the multipliers, multiply, move the results back. So the machine itself is three things — the multipliers, the storage, and the road between. Watch the loop closely: the moving already looks slower than the math. Next: inside a GPU.