hawkes.aiKey ideas
- The one idea — A model is one weights file plus a tiny loop of arithmetic — no database, no internet inside.
- How it is shown — Wi-Fi off, model still answering; the "AI" is a ~4 GB file being multiplied in a loop.
- The trap to avoid — "It's googling behind the scenes" — the offline demo kills it; so dies "it's a database of quotes."
- What it sets up — The weights file is *inert numbers* — a fact that returns at the very end of the series (the capstone treats the model as a sealed calculator behind an effector boundary).
People swear the AI is secretly googling their questions. Fine. Kill the Wi-Fi and ask it something. It keeps answering — because nothing was ever being fetched.
The one idea
A model is one weights file plus a tiny loop of arithmetic — no database, no internet inside.
This one file is the whole "AI." Watch it run with the Wi-Fi off. Here's what you download: one file, a few gigabytes. Inside: billions of numbers. Not sentences. Not facts. Not rules. Numbers — the weights you keep hearing about. The only other piece is a tiny program — small enough to read in an afternoon. Its job: take your tokens, multiply them through the weights, emit one next token. Repeat. That ring of arithmetic, spinning around the file, is the entire runtime. Now ask it things — with the internet gone.
How it works — the demo
Wi-Fi off, model still answering; the "AI" is a ~4 GB file being multiplied in a loop.
It writes. It codes. It explains. Slower than the big cloud models, and less brilliant — but unmistakably the same kind of mind. Nothing is being fetched, because there is nothing to fetch from. Hold this demo against the folklore. "It's secretly googling" — the Wi-Fi is off. "It's a database of quotes" — there's no database in the box. "Engineers wrote its rules" — nobody wrote rules. There are only weights, and the loop. So where does the knowing live? Smeared across the numbers.
The trap to avoid
"It's googling behind the scenes" — the offline demo kills it; so dies "it's a database of quotes."
Why it matters — and what’s next
The weights file is *inert numbers* — a fact that returns at the very end of the series (the capstone treats the model as a sealed calculator behind an effector boundary).
Paris-is-in-France isn't written in some cell — it's a pattern pressed into millions of weights during training, the way a face lives in no single pixel of a photograph. The trap: hearing "trained on the internet" and picturing an archive. The file is thousands of times smaller than what it read. It kept patterns, not pages — which is why it can explain brilliantly and misquote confidently, in the same breath. This is why open models change everything: a "brain" that is just a file can be copied, shared, run in a basement, fine-tuned in a weekend. And it means every claim about what AI "secretly does" has to survive one question: where, in a file of numbers and a loop, would that hide? Keep one phrase from today: the file is inert numbers. At the very end of this series, that boring fact becomes a safety strategy — the day we wall a mind off from the world and count exactly what can cross. For now: it's a file, and a loop. Everything else is engineering.
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 one file is the whole "AI." Watch it run with the Wi-Fi off.
Here's what you download: one file, a few gigabytes. Inside: billions of numbers. Not sentences. Not facts. Not rules. Numbers — the weights you keep hearing about.
The only other piece is a tiny program — small enough to read in an afternoon. Its job: take your tokens, multiply them through the weights, emit one next token. Repeat. That ring of arithmetic, spinning around the file, is the entire runtime.
Now ask it things — with the internet gone. It writes. It codes. It explains. Slower than the big cloud models, and less brilliant — but unmistakably the same kind of mind. Nothing is being fetched, because there is nothing to fetch from.
Hold this demo against the folklore. "It's secretly googling" — the Wi-Fi is off. "It's a database of quotes" — there's no database in the box. "Engineers wrote its rules" — nobody wrote rules. There are only weights, and the loop.
So where does the knowing live? Smeared across the numbers. Paris-is-in-France isn't written in some cell — it's a pattern pressed into millions of weights during training, the way a face lives in no single pixel of a photograph.
The trap: hearing "trained on the internet" and picturing an archive. The file is thousands of times smaller than what it read. It kept patterns, not pages — which is why it can explain brilliantly and misquote confidently, in the same breath.
This is why open models change everything: a "brain" that is just a file can be copied, shared, run in a basement, fine-tuned in a weekend. And it means every claim about what AI "secretly does" has to survive one question: where, in a file of numbers and a loop, would that hide?
Keep one phrase from today: the file is inert numbers. At the very end of this series, that boring fact becomes a safety strategy — the day we wall a mind off from the world and count exactly what can cross. For now: it's a file, and a loop. Everything else is engineering.