[mschuster91]

> So you release it as open weights which is a win-win. Global adoption of the model and you get to give the American AI companies a kick in the nuts because you know they will never release open weights apart from highly quantised crippled shit.

And on top of that, it's a perfect opportunity to include poisoned training data or excluding it. You know, omitting anything about Tiananmen Square, China's genocides against Uyghurs and Tibetans, or including texts propagandizing for the "reunification" (aka, annexation) of Taiwan.

And everyone who builds something like an interactive chatbot based on such "open weights" models now has a subtle chance of the answer being ideologically poisoned by the CCP.

We need actual open source, not "open weights" scam.

[seanmcdirmid]

How does this work for RAG? Do they make it so the model doesn’t have that fact in their weights or do they make it not talk about it when it is included in context.

Ironically, Chinese models have the most uncensored versions available for download. Fairly sure they own the porn market.

[nostrebored]

It’s in the weights. Context needs to be attended to to create a response, and the weights dictate what response is decoded. If you include retrieved context that has an American perspective, I imagine the think trace has some reconciliation about how they must be incorrect.

[notnullorvoid]

I wouldn't be worried so much about those examples. One could take the open weights and fine tune them to either fix the poisoning or omission of obvious topics.

It's the subtle topics that we should be concerned about, and double so with closed models where even if oddities are identified they are harder to research further and impossible to fix.

[traceroute66]

> You know, omitting anything about Tiananmen Square, China's genocides against Uyghurs and Tibetans, or including texts propagandizing for the "reunification" (aka, annexation) of Taiwan.

I am not Chinese and I'm not defending the Chinese, but I see this argument come up a lot.

The hard reality is that what you say is simply not going to affect 99.9999999999% of users.

Is it realistically going to affect anyone using an LLM in coding ? No.

Is it realistically going to affect anyone using an LLM in $anything_else_not_politically_sensitive ? No.

Does anyone seriously use LLMs for researching politically sensitive matters ? No.

The US does not exactly have an entirely pristine history either. Shall we discuss the post-9-11 related infrastructure of Guantanamo Bay ? Or the "Detention and Interrogation Program" that included a network of clandestine extrajudicial detention centres, officially known as "black sites"[1]?

Or maybe you would like to discuss the US supply of weapons for use in Gaza ?

[1]https://en.wikipedia.org/wiki/CIA_black_sites

[leereeves]

> The US does not exactly have an entirely pristine history either. Shall we discuss the post-9-11 related infrastructure of Guantanamo Bay ? Or the "Detention and Interrogation Program" that included a network of clandestine extrajudicial detention centres, officially known as "black sites"[1]?

Linking a US website discussing the topic doesn't exactly support your point.

[traceroute66]

> Linking a US website discussing the topic doesn't exactly support your point.

It supports my point precisely. Recall I also said "Does anyone seriously use LLMs for researching politically sensitive matters ? No.".

Just as there is plenty of information out there on the US's less than perfect history, there is also plenty of information out there on the various Chinese politically sensitive matters. You do not need a Chinese LLM to find out about it, all you need is a search engine.

The point is you have an open-weights LLM that is very good for a vast number of non-political uses, such as coding.

The point is that you can use the open-weights model instead of paying through the nose for a US model where they harvest your data unless you have an "enterprise" zero-data retention "trust me dude" clause that you have no viable way of verifying – and which incidentally is still subject to the good old "law, or court or administrative order" contract clauses, so it may not be as much of a zero-data retention as you think it is.