hawkes.aiKey ideas
- The one idea — Text is built token-by-token, left to right (autoregression); images are refined all-at-once, iteratively (diffusion) — opposite philosophies matched to opposite media, now starting to cross over.
- How it is shown — Split screen: left-to-right token generation vs whole-canvas refinement; each medium's structure shown fitting its engine; the boundary blurring at the edges.
- The trap to avoid — Treating the divide as a permanent law — the engines are converging (diffusion text, autoregressive images), so match by structure, not by dogma.
- What it sets up — You can carve an image — now, how do you STEER the carving?
Your chatbot can never revise a word it already typed. An image model revises everything, every step, and commits to nothing until the end. Opposite philosophies — here's why.
The one idea
Text is built token-by-token, left to right (autoregression); images are refined all-at-once, iteratively (diffusion) — opposite philosophies matched to opposite media, now starting to cross over.
Put the two engines side by side and they're mirror opposites. Text: one token at a time, left to right, each word locked before the next — sequential and committed. Images: the whole canvas at once, refined over many passes, nothing final until the end — parallel and revisable. Autoregression — the text engine from episode fifty-six — bets on sequence. Language is sequential: meaning unfolds in order. Its strength is coherent flow. Its bind is commitment: once a token is placed, it's fixed — the model can't revise, only forge forward. A one-way writer for a one-way medium. Diffusion bets the opposite way. An image has no reading order — every region is present at once and depends on every other, lighting and composition global. So diffusion refines the whole thing together, free to revise any region until the final step. Its strength is revisable global coherence.
How it works — the demo
Split screen: left-to-right token generation vs whole-canvas refinement; each medium's structure shown fitting its engine; the boundary blurring at the edges.
Its cost is the many passes. The deeper principle: the structure of what you make picks the engine. Sequential things — text, audio — lean autoregressive. Simultaneous things — images, spatial data — lean diffusion. Not fashion — it fits how the thing is structured. And video is both sequential and spatial. Hold that. And here's the twist: the engines are converging. Researchers now build diffusion language models that refine whole passages at once, and autoregressive image models that generate pictures token by token. The clean divide is softening. Hold the mapping as a strong default, not dogma — and watch the frontier crossing over. The trap: carving 'text is autoregression, images are diffusion' into stone.
The trap to avoid
Treating the divide as a permanent law — the engines are converging (diffusion text, autoregressive images), so match by structure, not by dogma.
Why it matters — and what’s next
You can carve an image — now, how do you STEER the carving?
It's the current best fit, driven by structure and cost — not a law, and already shifting. The durable skill isn't the pairing; it's reasoning from the medium. Sequential or simultaneous? What does revising cost? Those predict which engine fits. But notice what diffusion gives you so far: a beautiful image — of whatever the denoising happens to produce. Random. What you want is control: type 'a red bicycle at sunset' and get that. How does a string of words reach into a noise-carving process and steer it? Next group. Conditioning: how text controls the image.
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
Put the two engines side by side and they're mirror opposites. Text: one token at a time, left to right, each word locked before the next — sequential and committed. Images: the whole canvas at once, refined over many passes, nothing final until the end — parallel and revisable.
Autoregression — the text engine from episode fifty-six — bets on sequence. Language is sequential: meaning unfolds in order. Its strength is coherent flow. Its bind is commitment: once a token is placed, it's fixed — the model can't revise, only forge forward. A one-way writer for a one-way medium.
Diffusion bets the opposite way. An image has no reading order — every region is present at once and depends on every other, lighting and composition global. So diffusion refines the whole thing together, free to revise any region until the final step. Its strength is revisable global coherence. Its cost is the many passes.
The deeper principle: the structure of what you make picks the engine. Sequential things — text, audio — lean autoregressive. Simultaneous things — images, spatial data — lean diffusion. Not fashion — it fits how the thing is structured. And video is both sequential and spatial. Hold that.
And here's the twist: the engines are converging. Researchers now build diffusion language models that refine whole passages at once, and autoregressive image models that generate pictures token by token. The clean divide is softening. Hold the mapping as a strong default, not dogma — and watch the frontier crossing over.
The trap: carving 'text is autoregression, images are diffusion' into stone. It's the current best fit, driven by structure and cost — not a law, and already shifting. The durable skill isn't the pairing; it's reasoning from the medium. Sequential or simultaneous? What does revising cost? Those predict which engine fits.
But notice what diffusion gives you so far: a beautiful image — of whatever the denoising happens to produce. Random. What you want is control: type 'a red bicycle at sunset' and get that. How does a string of words reach into a noise-carving process and steer it? Next group. Conditioning: how text controls the image.