hawkes.aiKey ideas
- The one idea — "Open" labels three different offers — endpoint access, downloadable weights, and full-recipe openness — and only the third is open source in the software sense.
- How it is shown — Three doors opened in turn: the API counter, the weights warehouse (with license fine print), the rare full workshop (weights + code + data docs).
- The trap to avoid — Hearing "open source" and assuming software-style rights — most "open" models fail the actual open-source definition.
- What it sets up — Why isn't having the numbers having the source?
Most 'open source' AI models aren't open source — there's an official definition now, and they fail it. Here's what you're actually getting.
The one idea
"Open" labels three different offers — endpoint access, downloadable weights, and full-recipe openness — and only the third is open source in the software sense.
The word "open" does three different jobs in AI, and confusing them has real consequences — legal, security, strategic. Three doors: the counter, the warehouse, and the workshop. Only one is open the way software people mean it. Let's walk all three. Door one: the counter. API-only — you touch an endpoint, never the model. Requests in, answers out; the machinery stays behind glass — metered, revocable, sometimes changed beneath you. Most frontier-capability models live here, with honest virtues: zero hosting, instant improvements. But nothing is yours — not the weights, not the guarantee it exists tomorrow. Door two: the warehouse. Open weights — the crates from episode one hundred, genuinely downloadable. A real offer: run locally, fine-tune, keep forever; no one revokes what's on your disk.
How it works — the demo
Three doors opened in turn: the API counter, the weights warehouse (with license fine print), the rare full workshop (weights + code + data docs).
It powers everything this series has done at kitchen-table scale. Two shadows attach: a license wraps the crate — lightly or tightly — and the recipe that made it is not on these shelves. Door three: the workshop. Weights plus training code plus documented data — everything needed to rebuild and genuinely audit. A handful of research models ship this way, precious to science. But it's rare, because data disclosure is legally and competitively radioactive. There's now even a formal open-source-AI definition — created precisely because most models wearing the label fail it. Why do labels blur? "Open" is warm — generosity, community, trust — and warmth drapes over any door. A counter publishing a paper becomes "open research"; a warehouse becomes "open source" in a headline. Your x-ray, three questions no halo survives: Can I download it? Could I rebuild it?
The trap to avoid
Hearing "open source" and assuming software-style rights — most "open" models fail the actual open-source definition.
Why it matters — and what’s next
Why isn't having the numbers having the source?
What binds me if I ship? Ask all three, every time. The trap: importing software instincts. In software, "open source" carries specific rights — inspect, modify, redistribute, guaranteed. Assume the same about a model, and you've inherited obligations unread and dependencies unpriced. The word is doing sales work. Your job — before anything ships — is to make it do legal work instead. The next two episodes are exactly that. Next, door two gets taken seriously — the door this series has used all along, and the strangest. You can download every number, own them outright, run them forever — and still not know how they were made, what they were fed, or what's buried in them. Having the artifact without the recipe: next, why open weights isn't open source.
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
The word "open" does three different jobs in AI, and confusing them has real consequences — legal, security, strategic. Three doors: the counter, the warehouse, and the workshop. Only one is open the way software people mean it. Let's walk all three.
Door one: the counter. API-only — you touch an endpoint, never the model. Requests in, answers out; the machinery stays behind glass — metered, revocable, sometimes changed beneath you. Most frontier-capability models live here, with honest virtues: zero hosting, instant improvements. But nothing is yours — not the weights, not the guarantee it exists tomorrow.
Door two: the warehouse. Open weights — the crates from episode one hundred, genuinely downloadable. A real offer: run locally, fine-tune, keep forever; no one revokes what's on your disk. It powers everything this series has done at kitchen-table scale. Two shadows attach: a license wraps the crate — lightly or tightly — and the recipe that made it is not on these shelves.
Door three: the workshop. Weights plus training code plus documented data — everything needed to rebuild and genuinely audit. A handful of research models ship this way, precious to science. But it's rare, because data disclosure is legally and competitively radioactive. There's now even a formal open-source-AI definition — created precisely because most models wearing the label fail it.
Why do labels blur? "Open" is warm — generosity, community, trust — and warmth drapes over any door. A counter publishing a paper becomes "open research"; a warehouse becomes "open source" in a headline. Your x-ray, three questions no halo survives: Can I download it? Could I rebuild it? What binds me if I ship? Ask all three, every time.
The trap: importing software instincts. In software, "open source" carries specific rights — inspect, modify, redistribute, guaranteed. Assume the same about a model, and you've inherited obligations unread and dependencies unpriced. The word is doing sales work. Your job — before anything ships — is to make it do legal work instead. The next two episodes are exactly that.
Next, door two gets taken seriously — the door this series has used all along, and the strangest. You can download every number, own them outright, run them forever — and still not know how they were made, what they were fed, or what's buried in them. Having the artifact without the recipe: next, why open weights isn't open source.