[pegasus]

Regardless of the motivations behind the paper, the fact is that they are right. It's a topsy-turvy world where America closes up and China opens, and I hope it gets corrected soon.

And besides, it seems to me like the ethical choice. Given how these models are, in some very real sense, mechanical plagiators, built on the generosity of creators past and present, some of them now in danger of being replaced by the machine. I think the least the labs can do is open these models up. These and other such considerations were the reason OpenAI started with that name. Of course, it was questionable that those ideals would survive the encounter with generational wealth. Just look up what the founders of Google were saying about advertising when they were two students tinkering at an as of yet unproven tech. Same thing for OpenAI, self-interest speaks that much louder when there's real money on the table.

It just boggles the mind that people now make excuses for their all-too-predictable about-turn.

[throw0101d]

> I think the least the labs […]

Perhaps worth not calling them "labs". Are they not (for-profit) companies?

[astro1234]

Of course but I think labs is a good term. Like a lot of terms like this it comes from history. The people working on AI at these places come predominantly from academia and predominantly do research. Of course now these companies are far more than just a “lab” and I can imagine their headcount may be more non-research folks at this point but most of the results are indeed coming from the “lab”-y areas

[swingandamiss]

> It's a topsy-turvy world where America closes up and China opens

There's nothing "open" about China. Google, meta, openai, etc all blocked. Go visit and see how it goes when you try to access your gmail or open facebook. Try to use chatgpt. Try to get citizenship and see how that goes. China blocks many western companies with their great firewall and force internal similar products. This is smart, China wants to prioritize their own.

[smallmancontrov]

Your examples are the typical case, which is why OpenAI being closed and Kimi/DeepSeek/Qwen being open is topsy-turvy.

[plufz]

Im guessing parent meant more open in this particular case, i.e open weight models vs closed cloud hosted. Obviously china is not an open country.

[klik99]

He means specifically open vs closed AI model development, and how that’s topsy-turvy for the reasons you mention.

[chasd00]

well let's be clear, there's nothing open about an open-weight model.

To me, "open these models up. " must mean provide all the data and supporting documentation required to reproduce the model. That would be "open". Postgres is open because you can download all the data and supporting documentation and reproduce the binary yourself. However, being able to only download a postgres binary would make it no longer open.

Additional training on top of an open-weight model sounds analogous to writing mods for minecraft. You may change some behavior but that doesn't make minecraft "open".

[true_religion]

It's open, with a small 'o', and no moral judgements attached to it.

1. You have the weights, so you can run the model yourself on your own hardware. 2. You have the weights, so you can do post-training and shift those weights for your own purposes. It's not the same as training the model, but for many people its fine as what we want is a quantization, or a fine tune, or to create hybrid models.

What they don't give you are the training data, and reproduction instructions but... the toolchain to create the software has never been a part of 'Open Source'.

Even though it feels like a huge loophole, it's technically open source if you deliver the source code, without having a compiler that's available so you force them to recreate the toolchain from scratch.

To me, though, a better analogy is to research science where you'll be happy when they give you the full result set they compiled even if you don't get the raw data which may have IP or privacy concerns, or their often poorly documented lab notes so you can actually reproduce.

What you want to do is run your own experiment, and get your own results... not duplicate theirs directly. Even if reproduction is your aim in science, being unable to reproduce without copious notes sometimes points out that the original experimental process must have been flawed.

LLMs have the same issue. The creation process isn't entirely well documented, and the raw data can't be released since although the company have the right to use certain sources, they can't transfer those rights to others.

[cyberax]

The problem with that analogy is that compiling Postgres takes a few minutes. Training a model takes hundreds of thousands of dollars (at least) and specialized hardware.

Having a standardized training set is valuable, though.

[jfrbfbreudh]

Weights are the cake and source is the recipe. Except you aren’t going to spend millions working through the recipe to arrive at the same exact cake anyway, so why do you care?

[pegasus]

Sure, I'm all for opening the whole enchilada. But at least Chinese AI development is more open than the US one. And not just on the weights front, but AFAIK also more open when it comes to the techniques used and lessons learned in the process (DeepSeek at least is). Which is in a sense even more valuable and laudable.

[fragmede]

Here here! Open weights is the term because it's close to open source but everyone thinks everyone else is an idiot and saying the model is downloadable but you can't recreate it is too difficult for our feeble minds to understand. There are actually open source models out there though, with datasets and training code.

[StableAlkyne]

I think it has more to do with the fact that there has to be some term to describe the concept of "model which has weights that are openly accessible"

Like, that's just a logical thing to call it. I don't believe anyone is making a judgement on the intelligence of the reader to call it "open weight" when it refers to weights that are openly available.

"Open source" would be a more appropriate term to describe a model which also includes the training source.

[pegasus]

Indeed.