semiquaver

Good on them for being good sports, but after reading this I would never touch a piece of software built by anyone involved with this company. The sheer hubris of their claims combined with the specific reality of their broken garbage software (with the contradicted marketing claims and hallucinated documentation) is literally breathtaking.

It must have been amusing to see all these defect tickets coming in, only for them to frantically shovel each one to increasingly overworked LLM agents to apply pig-lipstick while the humans in charge oh-so-slowly start to realize their software quality is not quite what their glazing agents had been telling them and that what they have on their hands is basically mongo-quality slop. What a world we live in, huh?

You’re confusing two very different things. One is Oracle being careful with the provenance of IP that they let allow into their copyrighted codebase. That’s prudent in the face of real legal uncertainty over the issue. The other is

  > technically such a model should abide by the licence of the inputs, hence all models would probably be GPL

so whether a the weights file of a machine learning model itself can “be” GPL as a result of including GPL content in its training corpus.

That question is a category error because machine learning weights are not copyrightable at all, being mere lists of numbers and facts, the result of a statistical process and not a human creative hand, very much ineligible for copyright protection. They would be protected by trade secret law only, like the coke recipe.

So the GPL, which operates entirely within the domain of copyright law, cannot possibly virally spread from a licensed source code file to a machine learning model which trained on it. That would be a cross species jump. Certainly there could be other legal consequences but in no universe is someone going to jump out and yell gotcha and force OpenAI to personally hand deliver the newly GPLed weights for their frontier model to RMS.

Yeah, it does matter. This:

  > I think technically such a model should abide by the licence of the inputs

is vibes baloney.

  > If you're downloading it for commercial purposes (clearly the case) you're violating the license.

No. Copyright doesn’t work that way, no matter how much you might wish it did. And it’s a very important thing for the web that it doesn’t.

Legally speaking, it is extremely well established that downloading is not the type of copying that “copyright” concerns itself with. And the GPL and similar licenses are predicated on copyright laws.

Once you make the content available in some other way, then copyright law might have something to say.

People in tech used to make fun of “you wouldn’t download a car”. It’s sad that so many have reversed their principles the moment a convenient villain shows up.

Not a lawyer, are you?

Ok, so go yell at Microsoft about that incident instead of speculating about a different party doing a different thing.

Did someone do that in this case? I don’t see any indication let alone evidence that happened. And I don’t think it’s likely that it will.

Let me refer you to my earlier distinction between things that have happened and things you fear.

Who said we control the connecting device? This is a website. Websites are usually pretty arms-length things, and that’s the case with us. They’re our customers but we can’t install certificates on their computers and enforce mTLS. Can you do that with any of your customers?

Of course not. But there’s no indication whatsoever that any license has been violated. Downloading a file from a free access server cannot violate license terms.

Fear that something will happen is not the same thing as evidence it has happened.

This is a website about starting businesses.

Let me tell you, at the business I work at, some of our customers work at a company that uses a commercial VDI service for all workstations couldn’t access our service because the VDI egress was in GCP and we had an AWS WAF rule banning datacenter IPs.

Obviously, we removed the rule because it was doing nothing useful but plenty of harm. Real bots come from residential IPs.

So what? You made your content available and someone downloaded it. Why is this even notable?

Your “oh well, can’t trust anything” reaction is an insane reaction to what’s being implied.

To be clear here, the accusation is that the Chinese communist party or Alibaba itself intentionally engineered malicious secret loyalties into Qwen3.8-Flash-Next and made sure that evidence of them stays out of CoT. To compare that with “ordinary” stochastic model quality issues is bizarre and shocking. No idea if this accusation Is true, but if it is, this is an absolutely watershed moment.

There is no reason this needs to be a he-said-she-said thing. The model is open weights. Cohere should immediately publish their code to deterministically reproduce this on a given checkpoint of this model. Then anyone can verify the claim for themselves.

Presumably the cost of replacing these based on the higher than average defect rate is already baked into the price.