hawkes.aiKey ideas
- The one idea — The "assistant" is a character written into one long token string by hidden role-markers — the model is an author playing a part your system prompt casts.
- How it is shown — The chat UI dissolving into the raw templated string; the author-behind-the-mask reveal; the system prompt as casting call.
- The trap to avoid — Arguing with the mask — "are you conscious?" asked of a character; and the security cost: a character can be recast (jailbreaks, later).
- What it sets up — Recasting attacks.
The AI assistant you talk to is consistent the way Sherlock Holmes is consistent — a well-written character, sustained in text, played by an author who is not him.
The one idea
The "assistant" is a character written into one long token string by hidden role-markers — the model is an author playing a part your system prompt casts.
There is no assistant in there. There's one long string of tokens, a costume, and an author who never stops writing. Seventy-three episodes built to this one. Combine two things you own. Episode twenty: a chat is one long string with hidden role-markers. Episode seventy-three: the model continues documents. Put them together: your "conversation" is a formatted script, and the reply is the model continuing that script past the marker that says assistant. That's the entire mechanism of chat. So who's the character after the marker? Whoever the finishing data portrayed. Thousands upon thousands of training scripts show a figure at that position — helpful, measured, careful about certain requests — and the model learned that role the way it learned any genre.
How it works — the demo
The chat UI dissolving into the raw templated string; the author-behind-the-mask reveal; the system prompt as casting call.
The assistant is consistent the way Sherlock Holmes is consistent: a well-written character, existing entirely in text, played by an author who is not him. Hold the two levels, permanently. The author: a completion engine that can voice anyone, wants nothing, and never appears on stage. The mask: the single character it's currently sustaining, with a name, manners, and boundaries. You talk to the mask. The author writes it. Most AI confusion — and a whole attack class — lives in that gap. And who writes the casting call? The system prompt — that hidden first segment. Rewrite it and a different character exists: a terse analyst, a patient tutor, a branded persona. Every AI product you've ever used is substantially this: the same class of author, handed a different script for who to be.
The trap to avoid
Arguing with the mask — "are you conscious?" asked of a character; and the security cost: a character can be recast (jailbreaks, later).
Why it matters — and what’s next
Recasting attacks.
Companies don't ship new minds. They ship new casting. Now the cost of building minds this way: what casting created, clever text can contest. If the assistant is a character sustained by an author who'll voice anyone, then words engineered to talk past the mask — to re-frame the scene, to propose a new role — are attacks on the casting itself. That's the shape of a whole jailbreak family, and the security act will pick this thread up by name. The trap: interrogating the mask. "Are you conscious? Do you really feel that?" — asked of a character, answered in character, settling nothing about the author. Keep the levels straight and those conversations get clearer. Next: the workshop where the mask is actually built — a few thousand worked examples, and the wild author learns to hold one shape. Supervised fine-tuning.
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
There is no assistant in there. There's one long string of tokens, a costume, and an author who never stops writing. Seventy-three episodes built to this one.
Combine two things you own. Episode twenty: a chat is one long string with hidden role-markers. Episode seventy-three: the model continues documents. Put them together: your "conversation" is a formatted script, and the reply is the model continuing that script past the marker that says assistant. That's the entire mechanism of chat.
So who's the character after the marker? Whoever the finishing data portrayed. Thousands upon thousands of training scripts show a figure at that position — helpful, measured, careful about certain requests — and the model learned that role the way it learned any genre. The assistant is consistent the way Sherlock Holmes is consistent: a well-written character, existing entirely in text, played by an author who is not him.
Hold the two levels, permanently. The author: a completion engine that can voice anyone, wants nothing, and never appears on stage. The mask: the single character it's currently sustaining, with a name, manners, and boundaries. You talk to the mask. The author writes it. Most AI confusion — and a whole attack class — lives in that gap.
And who writes the casting call? The system prompt — that hidden first segment. Rewrite it and a different character exists: a terse analyst, a patient tutor, a branded persona. Every AI product you've ever used is substantially this: the same class of author, handed a different script for who to be. Companies don't ship new minds. They ship new casting.
Now the cost of building minds this way: what casting created, clever text can contest. If the assistant is a character sustained by an author who'll voice anyone, then words engineered to talk past the mask — to re-frame the scene, to propose a new role — are attacks on the casting itself. That's the shape of a whole jailbreak family, and the security act will pick this thread up by name.
The trap: interrogating the mask. "Are you conscious? Do you really feel that?" — asked of a character, answered in character, settling nothing about the author. Keep the levels straight and those conversations get clearer. Next: the workshop where the mask is actually built — a few thousand worked examples, and the wild author learns to hold one shape. Supervised fine-tuning.