ATRIUMsearch → argument graph
DefinitionVideo · 31:00 — 32:30

"Open weight" only means a model's trained numerical parameters are published; it does not disclose the training process or data and is a separate concept from open-source licensing.

CJ clarifies that 'open weight' is distinct from 'open source' — releasing weights only exposes the billions of trained numbers, not how the model was trained or what data was used. ✦ AI generated

CJ Reynolds · Syntax · 2026-07-20 · original ↗

starts at this moment · 31:00

Elicited by

Can you speak on this comment? Someone says open weight does not equal open source.

That at the end of the day results in a file that has billions of numbers inside of it. And so when we say open weight, they mean you can look at the weights. But the thing about neural networks is just looking at the weights doesn't tell you how it was trained or what they did special with it. You just can see the numbers. So that's what we say when we mean open weight.

verbatim transcript · starts at 31:00

Transcript · around this moment

30:40training a model is long and intense, right? They have to they have to get all of the training data. They have to run it on GPUs and they run >> they have to distill all of anthropics models. Exactly that. We're going to talk about that next. But yeah, so basically there's this whole big upfront process that they do >> that at the end of the day results in a

31:00file that has billions of numbers inside of it. And so when we say open weight, they mean you can look at the weights. But the thing about neural networks is just looking at the weights doesn't tell you how it was trained or what they did special with it. You just can see the numbers. So that's what we say when we mean o open weight. Um, and then on top

31:18of that, these models also have various licenses like yes, you can run this for profit or you you can only you you can run this model, but you uh can't charge people for it. So yeah, that's a good distinction to make. We can't see into how they trained this thing, but we can see the numbers that the training process produced. Um, so something to think about. >> And like I'm curious like like you

31:40obviously wouldn't be able to take this model and then start charging people or I'm not sure what the license actually is, but like what about like if you were to like slap this into your product that uses AI, you know, there's the rules around that as well. >> Exactly. You'd Yeah, I'm I'm not a lawyer. You'd have to look at their their terms of service and licenses for

31:58each of these models, but that that's the kind of thing. The other cool thing you can do with these openw weightight models is fine-tune them. Uh and so like there's a whole ecosystem of uh taking an openweight model and then further training it on specific types of data to get specific outputs. So >> that's the other cool thing about openw weight is like we could take a a

32:18specific model and then train it really really good on all of the the syntax podcast transcripts and then it would be like and this is as opposed to something like retrieval augmented generation where you just like paste in the transcripts. You could literally fine-tune the weights to know more about how Scott and Wes talk and what topics they talked about in what episode. Um, or if you work at an enterprise, you

32:42could take all of your internal documentation and and everything else and fine-tune a model on that. So instead of having to wait for API calls, you have a model that knows your company uh inside and out, that kind of thing. >> Yeah. >> Remember everybody remember that was what everybody wanted to do. They're like, I can't wait till I can train my own model. Um, and that doesn't seem to

Around this claim
This moment responds to
explains mechanismKimi makes no attempt to describe this as an 'open source' license, consistently using the term 'open weight' instead — which is honest.Author · Simon Willison's WeblogextendsOpen weight models are not the same as open source — you can see the weights (numbers) but not how the model was created or what data trained it.CJ · Syntaxprovides contextPre-training data filtering is an overlooked but promising approach to reducing misuse risk in open-weight models, and FAR.AI plans to open-source safe filters and fund $2M in grants for this work.Adam Gleave · The Cognitive Revolutionprovides contextNVIDIA publishes the training data, tools, recipes, and environments alongside its models — Nemotron is not just weights but the full reproducible stack — and this is what makes a model genuinely open.Bryan Catanzaro · ByteByteGo Newsletterprovides contextThe licensing ambiguity is the most concrete skeptical reaction: a flagged license prohibition covering the USA, EU, UK, and Korea would mean 'open weights' lacks OSI-style rights, imposes use-case and jurisdiction restrictions, and may forbid even downloading the model from the US.AINews · Latent Spaceprovides contextOpen-weight models are inherently decelerationist for the AI race, which is why it's surprising that self-described accelerationists are so enthusiastic about them.Dean Ball · Interconnectsprovides contextBecause U.S. open weight model makers must obey frontier labs' terms of service barring distillation, they end up worse than Chinese alternatives and distill Chinese models instead — the U.S. should legalize training-data collection as fair use and bar anti-distillation terms of service.Ben Thompson · Stratecheryprovides contextKimmy K3 is the first open-weight AI model whose benchmarks put it on par with frontier closed models like Opus and GPT.CJ Reynolds · Syntax