Hacker News Reader

Comment by kingstnap | Hacker News Reader

If you think of LLMs as programs. The weights and inference code are very much a binary.

While the training code and data are the true source. Since if you want to robustly modify the LLM that's actually what you need.

But since "compilation" (training) is extremely compute intensive this isn't something accessible to anyone without an entire datacenter.

Anyway semantics aside having the binary is still infinitely better than dealing with an api as far as privacy and control go.

View on HN

That’s not how it works though. Two training runs on the same data don’t produce the same weights. And if you want to modify the AI, you do so by fine tuning the weights not rerunning training. In every respect that matters, the weights are both the binary and the source code together.

By this standard there is no closed source software on PC. You can always modify the binary directly or hook your functions in.

Wouldn’t any software that is a client and relies on a proprietary server to function (like most modern AAA games) be “closed source” under this definition? Not that I agree with this definition of open source btw, just saying that there is precedent.

Two builds of the same source code don’t always produce exactly the same binary either, but we don’t use that as an excuse to distribute object files and call it open-source.

This is the same as saying every binary is open source because you can look at and change the machine code.

Open source means you reveal how you created this binary.

But training is not deterministic. The same training run on the same input can result in very different models. And even if you ignore the intrinsically stochastic part of training, the effort of creating a model is ad-hoc and involving humans. It is a crafted output, not a compiler output. The pre-training is the closest to being mechanistic, but these models are so many layers of work on top of the pretraining, and those higher layers have people in the loop. They run experiments, tweak things, run experiments again.

If I want to remove censorship from your open-weights model, how do I do that?

The process is called Ablation, there are many ablated models available to download

https://en.wikipedia.org/wiki/Ablation_(artificial_intellige...

> ...how do I do that?

Put in the work. This is akin to asking how to remove Rust from a Rust project; just because something is legally available to you doesn't mean you wont need to apply dome elbow grease, depending how deep the changes you want are, ablation, fine-tuning, or distillation are tools you can use to remove "censorship"

I agree. But models in difference to compiled binaries, are useful as just weights and can be further refined and post-trained, at least.

I don't know LLM theory well enough to say if there's some secret sauce they can hold back that makes training ineffective. Less effective I'm sure, we don't have access to their smart training schemes, but post-training should always be possible IIUC.

At the risk of taking the analogy too far, I would treat refining like modifying a dynamic library. You can technically modify behavior, but only in a very coarse way.

post-training is like writing a wrapper around the binary. It is closer to building on top of than truly modifying, in that you can tailor things to your needs slightly but cannot make fundamental changes to the underlying thing.

Are you assuming single-layer LORA fine-tuning on top? Because with open-weights you can do full back-propagation training to mold the model into whatever shape you want.

For a stretched analogy, I think it is more like LEGO sets. Someone hands you a 10,000 piece masterpiece, and a box of unused LEGO parts. Hackers on HN object that the LEGO part manufacturing process is not included, you can't make your own parts, etc. But it's LEGO. You can pull apart the model, see how it is constructed, add your own refinements and features, or even redo it from the ground up. In a practical sense having knowledge about the factory making the parts doesn't really matter here.

The weights + the architecture are already 100% of the code, the transformer is just a mathematical expression + helper programs whose sources are provided. The transformer itself is not even a stateful program, so a it is no more a binary than Piet or Tromp's BLC are. It's merely incomprehensible. Training isn't compilation either, since training a model is closer to program induction and the data are samples defining the solution space.

Reply on news.ycombinator.com