hawkes.aiKey ideas
- The one idea — GPUs do thousands of identical multiply-adds in parallel; CPUs do a few things fast — and AI's work is the interns' kind.
- How it is shown — A genius solving four problems in sequence vs a thousand interns solving four thousand at once.
- The trap to avoid — Treating a GPU as a "faster CPU" — single-file work runs slower on it; the win only exists when work splits into identical pieces.
- What it sets up — What exactly is the work being split?
A GPU is not a faster CPU. Give it normal single-file work and it's actually slower. So why does all of AI run on them?
The one idea
GPUs do thousands of identical multiply-adds in parallel; CPUs do a few things fast — and AI's work is the interns' kind.
Your CPU is a genius — a few cores that do anything, one thing after another. Your GPU is a thousand interns, each doing one tiny sum. For AI, the interns win. Every time. Here's why. Look at the work. Episode twenty-six: a neuron is multiply-and-add. A model is billions of them — and within a layer, none waits on another. This is not one hard problem; it's an ocean of identical easy ones. A genius does them in sequence. A crowd does them at once.
How it works — the demo
A genius solving four problems in sequence vs a thousand interns solving four thousand at once.
The interns weren't hired for this. Graphics chips evolved to color millions of pixels with the same small math, sixty times a second — same operation, oceans of data. Then neural networks arrived needing exactly that shape of work. The great coincidence of modern computing: the gaming chip was accidentally an AI chip. Under the hood, two philosophies. The CPU spends its silicon on decision machinery — predicting branches, caching cleverly — so one task finishes soonest. The GPU strips all that out and spends everything on arithmetic, so a thousand tasks finish together. Latency versus throughput. Not faster versus slower — two bets about what work looks like. The fuse lit in twenty-twelve. AlexNet — the vision model that woke the field — trained on two consumer gaming cards.
The trap to avoid
Treating a GPU as a "faster CPU" — single-file work runs slower on it; the win only exists when work splits into identical pieces.
Why it matters — and what’s next
What exactly is the work being split?
Work that would have taken CPU months ran in days, and it won so hard the whole field switched. A decade on, the intern-maker became one of the most valuable companies on Earth. The trap: thinking a GPU is a faster CPU. Hand the interns one long chain of steps — each depending on the last — and they're slower than the genius; a lone GPU core is feeble. The win exists only when work splits into identical pieces. That's why your laptop carries both: genius for the operating system, interns for the matrix. So what exactly is the work being split? Zoom out from the tiles and a shape appears: the multiply-adds aren't scattered — they're arranged, rows against columns, the same geometric object repeated everywhere you look. The entire boom runs on one operation. Next: it's all just matrix multiplication.
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
Your CPU is a genius — a few cores that do anything, one thing after another. Your GPU is a thousand interns, each doing one tiny sum. For AI, the interns win. Every time. Here's why.
Look at the work. Episode twenty-six: a neuron is multiply-and-add. A model is billions of them — and within a layer, none waits on another. This is not one hard problem; it's an ocean of identical easy ones. A genius does them in sequence. A crowd does them at once.
The interns weren't hired for this. Graphics chips evolved to color millions of pixels with the same small math, sixty times a second — same operation, oceans of data. Then neural networks arrived needing exactly that shape of work. The great coincidence of modern computing: the gaming chip was accidentally an AI chip.
Under the hood, two philosophies. The CPU spends its silicon on decision machinery — predicting branches, caching cleverly — so one task finishes soonest. The GPU strips all that out and spends everything on arithmetic, so a thousand tasks finish together. Latency versus throughput. Not faster versus slower — two bets about what work looks like.
The fuse lit in twenty-twelve. AlexNet — the vision model that woke the field — trained on two consumer gaming cards. Work that would have taken CPU months ran in days, and it won so hard the whole field switched. A decade on, the intern-maker became one of the most valuable companies on Earth.
The trap: thinking a GPU is a faster CPU. Hand the interns one long chain of steps — each depending on the last — and they're slower than the genius; a lone GPU core is feeble. The win exists only when work splits into identical pieces. That's why your laptop carries both: genius for the operating system, interns for the matrix.
So what exactly is the work being split? Zoom out from the tiles and a shape appears: the multiply-adds aren't scattered — they're arranged, rows against columns, the same geometric object repeated everywhere you look. The entire boom runs on one operation. Next: it's all just matrix multiplication.