Weights on your own disk, and the awkward question of distillation
· Christoph Heuwieser
Video: PewDiePie is setting AI free... and OpenAI is furious, Fireship (The Code Report), 5 October 2026
I watched Fireship's latest Code Report twice this morning. The first time was for the jokes. The second time was because it describes, in under six minutes, a shift I keep running into in my own work: open-weight models are becoming the default choice for a growing group of serious builders, not a hobbyist curiosity.
What the video covers
Fireship walks through Ajax, a model PewDiePie has been building for his open-source project Odysseus. According to the video, Ajax is a fine-tune of an Alibaba Qwen base model, built largely "the hard way": he hand-collected about 300 good tool-use examples, generated synthetic data and filtered it down to roughly 2,000, ran reinforcement learning with GRPO, and finally used a tool called Heretic to strip out the refusals.
The tone is satirical; this is Fireship. But the line that stuck with me is a plain statement of the payoff:
"the prize is that you now have a set of weights that live on your hard drive, and you no longer need to pay one of these landlords to rent intelligence from the cloud." (Fireship, The Code Report, 5 Oct 2026)
Why open weights keep gaining ground
The most interesting detail in the video is the data. The video says the plan called for 20,000 clean examples, and only about 300 useful ones could be collected by hand. When fans were asked to donate their data, the video says "basically nobody did."
This is the real bottleneck. Base models are increasingly available, and the training recipes (supervised fine-tuning, GRPO, refusal-ablation tooling) are public and getting easier to use. What is still scarce is good examples of real work actually getting done: the messy, multi-step traces where an agent used tools, got stuck, recovered and finished.
When the weights and recipes are commodities, the advantage moves to whoever has:
- a specific domain worth specialising in,
- real examples of work in that domain, and
- the willingness to run the model themselves.
I think that list will cover far more teams than it did a year ago. That's why I read "open weights are getting more relevant" as a structural trend, not a mood.
The awkward part: distillation
Distillation as a technique is old and well documented. The video cites the 2015 Hinton paper: a smaller "student" model learns to imitate a larger "teacher". The question isn't whether it works. It's whose outputs you're allowed to learn from, and on what terms. People I respect come down in different places on this, and I think each position has a real point.
The provider view. Training frontier models costs a fortune. Hosted providers commonly put clauses in their terms restricting the use of outputs to build competing models, and they enforce those terms on their own platforms. From their side, letting someone pay per token to extract the capability of a model that cost far more to build is a business-model problem, whatever else it is. Enforcing your own terms of service is not villainy.
The open-weight view. The objection here is about symmetry. Large models were themselves trained on vast amounts of human-written text, mostly without individual permission. Saying "our outputs must not train your model" sits uneasily next to that. Many open-weight licences are explicitly more permissive about downstream use. The video plays this tension for laughs, but it isn't a joke to the people who hold this view.
The legal view. As far as I can tell, it's genuinely unsettled. Terms of service are contracts. Whether model outputs are protectable at all, and how far a contract term reaches once outputs are out in the world, are questions that different jurisdictions may answer differently. I'm not a lawyer and I won't pretend to know where this lands. Anyone building on distilled data should take real legal advice for their own situation, not get it from a YouTube video or a founder's blog.
My own working position is simpler than any of these. Respect the terms you agreed to, and be explicit about provenance. If you can't explain where your training data came from and under what terms, you have a problem, whether or not anyone ever checks.
Dead ends are part of the story
One more thing I liked: the video says Ajax started with a failed experiment. It was a "council" of agents voting on answers that, as Fireship tells it, drifted into the agents forming alliances. That dead end is what pushed the project toward a custom model. Failed attempts teach you what to build next, as long as you kept a record of them.
Where this touches what we do
At a4sx, people publish, acquire and continue AI coding-agent sessions: real traces of real work, including the ones that went nowhere. That puts questions of who made something, who may learn from it and on what terms right in the middle of our product. We don't think we have all the answers yet. We'd rather work through them in public.