hawkes.aiKey ideas
- The one idea — Three lineages turned downloadable models from toys into near-frontier tools — and compressed the open-vs-closed gap from years to months.
- How it is shown — The ridge-line: open capability climbing beneath the sealed peaks — Llama's lineage, Qwen's breadth, DeepSeek's efficiency shocks — the gap narrowing on a time axis.
- The trap to avoid — Two stale reflexes — "open = far behind" and "open = private by default"; both outdated, both differently.
- What it sets up — And then open went TRILLION-scale.
Downloadable AI used to trail the closed labs by years. Now it's months — and the models you can own keep closing in.
The one idea
Three lineages turned downloadable models from toys into near-frontier tools — and compressed the open-vs-closed gap from years to months.
Below the sealed peaks of the closed frontier runs a ridge-line you can actually walk: the open frontier — models you download, own, and school yourself. Three lineages built it, each in a different style, and together they compressed the open-versus-closed gap from years to months. The three routes up the ridge. Route one: the Llama line. Born as episode one-thirty-six's accident — the leak — then licensed on purpose, generation after generation, up to four-hundred-billion-class dense giants. Its deepest gift wasn't any single model. It was the ecosystem: the tools, the fine-tune culture, the community. Llama made open-weights a movement. Route two: Qwen — breadth as strategy. Not one flagship but a laddered network: sizes from pocket to giant, specialist spurs for code, math, and vision, licensing mostly at the genuinely-free end of episode one-thirty-four's shelf.
How it works — the demo
The ridge-line: open capability climbing beneath the sealed peaks — Llama's lineage, Qwen's breadth, DeepSeek's efficiency shocks — the gap narrowing on a time axis.
Whatever your hardware and use case, there's a rung. For builders, coverage like that is its own frontier. Route three: DeepSeek — fewer releases, bigger shocks. The efficiency inventions you've studied debuted here: the latent cache of episode one-twenty-eight, the committee refinements of one-twenty-four, and open reasoning taken within months of the sealed peaks — episode eighty-two's earthquake. Each arrival didn't climb the ridge. It moved it. Measure the gap across eras: years behind, then quarters, then — by the reasoning era — months. The caveat: the sealed peaks keep rising too, and the ceiling stays behind glass. But the strategic fact stands: for a widening class of real jobs, the walkable ridge is now high enough. That reframes build-versus-buy.
The trap to avoid
Two stale reflexes — "open = far behind" and "open = private by default"; both outdated, both differently.
Why it matters — and what’s next
And then open went TRILLION-scale.
Two stale reflexes to retire. "Open means far behind" — the measurements say months, and for many jobs it doesn't bind. And "open means private": a downloaded model on rented cloud machines is exactly as private as that arrangement. Privacy comes from deployment, not the license. Ownership makes privacy possible, not automatic. And the ridge kept rising — into territory nobody predicted. The largest downloadable models on Earth are now trillion-scale committee giants, and they come overwhelmingly from labs the early family tree barely featured. The open frontier's newest chapter, and its geography: next. GLM, Kimi, and the trillion-parameter open era.
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
Below the sealed peaks of the closed frontier runs a ridge-line you can actually walk: the open frontier — models you download, own, and school yourself. Three lineages built it, each in a different style, and together they compressed the open-versus-closed gap from years to months. The three routes up the ridge.
Route one: the Llama line. Born as episode one-thirty-six's accident — the leak — then licensed on purpose, generation after generation, up to four-hundred-billion-class dense giants. Its deepest gift wasn't any single model. It was the ecosystem: the tools, the fine-tune culture, the community. Llama made open-weights a movement.
Route two: Qwen — breadth as strategy. Not one flagship but a laddered network: sizes from pocket to giant, specialist spurs for code, math, and vision, licensing mostly at the genuinely-free end of episode one-thirty-four's shelf. Whatever your hardware and use case, there's a rung. For builders, coverage like that is its own frontier.
Route three: DeepSeek — fewer releases, bigger shocks. The efficiency inventions you've studied debuted here: the latent cache of episode one-twenty-eight, the committee refinements of one-twenty-four, and open reasoning taken within months of the sealed peaks — episode eighty-two's earthquake. Each arrival didn't climb the ridge. It moved it.
Measure the gap across eras: years behind, then quarters, then — by the reasoning era — months. The caveat: the sealed peaks keep rising too, and the ceiling stays behind glass. But the strategic fact stands: for a widening class of real jobs, the walkable ridge is now high enough. That reframes build-versus-buy.
Two stale reflexes to retire. "Open means far behind" — the measurements say months, and for many jobs it doesn't bind. And "open means private": a downloaded model on rented cloud machines is exactly as private as that arrangement. Privacy comes from deployment, not the license. Ownership makes privacy possible, not automatic.
And the ridge kept rising — into territory nobody predicted. The largest downloadable models on Earth are now trillion-scale committee giants, and they come overwhelmingly from labs the early family tree barely featured. The open frontier's newest chapter, and its geography: next. GLM, Kimi, and the trillion-parameter open era.