hawkes.aiKey ideas
- The one idea — Nearly every modern model descends from one 2019 recipe — decoder-only transformer, next-token training — through five great branchings.
- How it is shown — The tree of light: one root (GPT-2) → scale (GPT-3) → instruction era → the open branch → the reasoning era → today's canopy, with the series' acts mapped onto the trunk.
- The trap to avoid — Treating each release as a new species — it's one family, and knowing the tree turns marketing into genealogy.
- What it sets up — The root was once called too dangerous.
Under the hood, today's trillion-parameter giants still run the same core recipe as a 2019 experiment: read text, predict the next token.
The one idea
Nearly every modern model descends from one 2019 recipe — decoder-only transformer, next-token training — through five great branchings.
Every model you've heard of — the chatbots, the coders, the reasoning giants, the trillion-parameter committees — descends from one line of experiments published in twenty-eighteen and nineteen. One family, five great branchings. Learn the tree once, and every future release becomes genealogy instead of noise. From the root, then. The root is GPT-2, twenty-nineteen — grown from the transformer paper two years before it. Inside the seed, the complete recipe this series taught: a decoder tower, episode sixty's next-word ritual, nothing else. Hold that. Everything since — everything — differs in size, data, and finishing school. The kind never changed. That's what makes it a family. Branching one: scale.
How it works — the demo
The tree of light: one root (GPT-2) → scale (GPT-3) → instruction era → the open branch → the reasoning era → today's canopy, with the series' acts mapped onto the trunk.
GPT-3, twenty-twenty — roughly a hundred times its parent's size, and strange fruit appeared that nobody planted; that shock gets its own episode. Branching two: manners. The instruction era — act three's entire finishing school shipping as InstructGPT, then ChatGPT in late twenty-twenty-two, the release that made the tree famous to everyone on Earth. Branching three: the open limb. In twenty-twenty-three a major lab's weights escaped, then were licensed on purpose — and the downloadable canopy erupted: Llama's line, then Mistral, Qwen, DeepSeek, the families from your act-five marketplace. Two canopies ever since, sealed and open, growing in visible dialogue — each forcing the other taller, faster. The recent branchings you've already studied. Four: reasoning — the think-before-answering crowns of act three's finale, arriving closed in late twenty-twenty-four and open within months. Five: committees — mixture-of-experts reshaping the trunk itself, this act's opening story. Today's canopy is all five branchings compounding: scaled, mannered, opened, thinking, and committee-built. The trap: reading each release as a new species.
The trap to avoid
Treating each release as a new species — it's one family, and knowing the tree turns marketing into genealogy.
Why it matters — and what’s next
The root was once called too dangerous.
It never is. The genealogist's questions cut every announcement to size: which branch does it sit on? What's actually new — and what's inherited wood wearing fresh leaves? Most "revolutionary" releases are one real graft on familiar limbs. Knowing the tree doesn't make you cynical. It makes you accurate. And the root deserves its own telling. Because the seed this whole canopy grew from — tiny by any modern measure, outclassed by what runs on your phone — was once considered too dangerous to publish. The panic, what it got right, what it got wrong, and why that cycle keeps repeating: next. How GPT-2 shocked everyone.
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
Every model you've heard of — the chatbots, the coders, the reasoning giants, the trillion-parameter committees — descends from one line of experiments published in twenty-eighteen and nineteen. One family, five great branchings. Learn the tree once, and every future release becomes genealogy instead of noise. From the root, then.
The root is GPT-2, twenty-nineteen — grown from the transformer paper two years before it. Inside the seed, the complete recipe this series taught: a decoder tower, episode sixty's next-word ritual, nothing else. Hold that. Everything since — everything — differs in size, data, and finishing school. The kind never changed. That's what makes it a family.
Branching one: scale. GPT-3, twenty-twenty — roughly a hundred times its parent's size, and strange fruit appeared that nobody planted; that shock gets its own episode. Branching two: manners. The instruction era — act three's entire finishing school shipping as InstructGPT, then ChatGPT in late twenty-twenty-two, the release that made the tree famous to everyone on Earth.
Branching three: the open limb. In twenty-twenty-three a major lab's weights escaped, then were licensed on purpose — and the downloadable canopy erupted: Llama's line, then Mistral, Qwen, DeepSeek, the families from your act-five marketplace. Two canopies ever since, sealed and open, growing in visible dialogue — each forcing the other taller, faster.
The recent branchings you've already studied. Four: reasoning — the think-before-answering crowns of act three's finale, arriving closed in late twenty-twenty-four and open within months. Five: committees — mixture-of-experts reshaping the trunk itself, this act's opening story. Today's canopy is all five branchings compounding: scaled, mannered, opened, thinking, and committee-built.
The trap: reading each release as a new species. It never is. The genealogist's questions cut every announcement to size: which branch does it sit on? What's actually new — and what's inherited wood wearing fresh leaves? Most "revolutionary" releases are one real graft on familiar limbs. Knowing the tree doesn't make you cynical. It makes you accurate.
And the root deserves its own telling. Because the seed this whole canopy grew from — tiny by any modern measure, outclassed by what runs on your phone — was once considered too dangerous to publish. The panic, what it got right, what it got wrong, and why that cycle keeps repeating: next. How GPT-2 shocked everyone.