hawkes.aiKey ideas
- The one idea — A pretrained-only model completes text — it won't answer, refuse, or stop being whoever the text implies.
- How it is shown — A question answered with more questions; a letter continued past its ending into the reply and the next letter.
- The trap to avoid — Judging base models as broken assistants — they're a different animal, and prompting one is fishing, not asking.
- What it sets up — So who is the assistant you talk to?
The raw model under your assistant will be anyone the text implies — pirate, villain, sincerely, no hesitation. Refusing isn't a thing documents do. Meet the base model.
The one idea
A pretrained-only model completes text — it won't answer, refuse, or stop being whoever the text implies.
Ask the freshly-trained model the capital of France, and it may reply with four more geography questions. Not broken — doing its job perfectly. Raw models don't chat. They continue. Meet the weirdo the furnace actually produces. Why more questions? Because in the training pile, a lone question usually lives on a quiz sheet — and the likeliest next text after one quiz question is another. The machine answered the only question it was ever asked: what comes next? Flawlessly. Your question wasn't a question to it. It was a genre. The letter test makes it vivid. Feed it half a letter: it finishes beautifully — then writes the reply, then your response to the reply, then a postscript from someone new.
How it works — the demo
A question answered with more questions; a letter continued past its ending into the reply and the next letter.
It has no concept of "my turn," because turns don't exist in raw text. There is no it speaking. There is only the document, growing. And it will be anyone. Start a pirate's journal, it's the pirate. A villain's monologue — it's the villain, sincerely, with no hesitation and no floor, because refusing isn't a thing documents do. The base model isn't an assistant with opinions. It's an author with none, wearing whatever the text hands it. So how did anyone use these things? By fishing. Write a page where three questions get answered, leave the fourth blank — and the likeliest continuation is your answer. Few-shot prompting: building a document whose natural next text is the thing you need. The entire craft of early prompt engineering was this — trap-building for a genre machine.
The trap to avoid
Judging base models as broken assistants — they're a different animal, and prompting one is fishing, not asking.
Why it matters — and what’s next
So who is the assistant you talk to?
And the weirdo isn't history. Model families ship both versions today — base and instruct, same weights-mass, different finishing. Researchers probe the raw ones to study the mind before the manners. Writers prize their wildness. Builders buy them as blanks for custom finishing. The weirdo is a product line. The trap: grading a base model as a broken assistant. It's not broken — it's a different animal, and everything assistant-like was added later. Which sharpens the strangest question in this act: the helpful character you talk to every day — the one with manners, boundaries, a name. If the furnace doesn't produce it… who is it? Next: the assistant doesn't exist.
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 2:45 of narration
Ask the freshly-trained model the capital of France, and it may reply with four more geography questions. Not broken — doing its job perfectly. Raw models don't chat. They continue. Meet the weirdo the furnace actually produces.
Why more questions? Because in the training pile, a lone question usually lives on a quiz sheet — and the likeliest next text after one quiz question is another. The machine answered the only question it was ever asked: what comes next? Flawlessly. Your question wasn't a question to it. It was a genre.
The letter test makes it vivid. Feed it half a letter: it finishes beautifully — then writes the reply, then your response to the reply, then a postscript from someone new. It has no concept of "my turn," because turns don't exist in raw text. There is no it speaking. There is only the document, growing.
And it will be anyone. Start a pirate's journal, it's the pirate. A villain's monologue — it's the villain, sincerely, with no hesitation and no floor, because refusing isn't a thing documents do. The base model isn't an assistant with opinions. It's an author with none, wearing whatever the text hands it.
So how did anyone use these things? By fishing. Write a page where three questions get answered, leave the fourth blank — and the likeliest continuation is your answer. Few-shot prompting: building a document whose natural next text is the thing you need. The entire craft of early prompt engineering was this — trap-building for a genre machine.
And the weirdo isn't history. Model families ship both versions today — base and instruct, same weights-mass, different finishing. Researchers probe the raw ones to study the mind before the manners. Writers prize their wildness. Builders buy them as blanks for custom finishing. The weirdo is a product line.
The trap: grading a base model as a broken assistant. It's not broken — it's a different animal, and everything assistant-like was added later. Which sharpens the strangest question in this act: the helpful character you talk to every day — the one with manners, boundaries, a name. If the furnace doesn't produce it… who is it? Next: the assistant doesn't exist.