I don’t care for Sam Altman and his general untrustworthy behavior. But DeepSeek is perhaps more untrustworthy. Models from American companies at least aren’t surprising us with government driven misinformation, and even though safety can also be censorship, the companies that make these models at least openly talk about their safety programs. DeepSeek is implementing a censorship and propaganda program without admitting it at all, and once they become good at doing it in less obvious ways, it can become very damaging and corrupt the political process of other societies, because users will trust the tools they use are neutral.
I think DeepSeek’s strategy to announce a misleading low cost (just the final training run that optimizes a base model that in turn is possibly based on OpenAI) is also purposeful. After all, High Flyer, the parent company of DeepSeek, is a hedge fund - and I bet they took out big short positions on Nvidia before their recent announcements. The Chinese government, of course, benefits from a misleading number being announced broadly, causing doubt among investors who would otherwise continue to prop up American technology startups. Not to mention the big fall in American markets as a result.
I do think there’s also a big difference between scraping the Internet for training data, which might just be fair use, and training off other LLMs or obtaining their assets in some other way. The latter feels like the kind of copying and industrial espionage that used to get China ridiculed in the 2000s and 2010s. Note that DeepSeek has never detailed their training data, even at a high level. This is true even in their previous papers, where they were very vague about the pre training process, which feels suspicious.
> Models from American companies at least aren’t surprising us with government driven misinformation, and even though safety can also be censorship
Being a citizen of a western nation, I'm inclined to agree with the general sentiment here, but how can you definitively say this? You, or I, don't know with any certainty what interference the US government has played with domestic LLMs, or what lies they have fabricated and cultivated, that are now part of those LLMs' collective knowledge. We can see the perceived censorship with deepseek more clearly, but that isn't evidence that we're in any safer territory.
> There are loads of examples on the internet of LLMs pushing (foreign) government narratives e.g. on Israel-Palestine
There isn’t even a single example of that. If an LLM is taking a certain position because it has learned from articles on that topic, that’s different from it being manipulated on purpose to answer differently on that topic. You’re confusing an LLM simply reflecting the complexity out there in the world on some topics (showing up in training data), with government forced censorship and propaganda in DeepSeek.
Fine, whatever. It's actually much more concerning if the overall information landscape has been so curated by censors that a naively-trained LLM comes "pre-censored", as you are asserting. This issue is so "complex" when it comes to one side, and "morally clear" when it comes to the other. Classic doublespeak.
That's far more dystopian than a post-hoc "guardrailed" model (that you can run locally without guardrails).
> I don’t care for Sam Altman and his general untrustworthy behavior. But DeepSeek is perhaps more untrustworthy. Models from American companies at least aren’t surprising us with government driven misinformation, and even though safety can also be censorship, the companies that make these models at least openly talk about their safety programs. DeepSeek is implementing a censorship and propaganda program without admitting it at all, and once they become good at doing it in less obvious ways, it can become very damaging and corrupt the political process of other societies, because users will trust the tools they use are neutral.
These arguments always remind me of the arguments against Huawei because they _might_ be spying on western countries. On the other hand we had the US government working hand in hand with US corporations in proven spying operations against western allies for political and economic gain. So why should we choose an American supplier over a Chinese one?
> I think DeepSeek’s strategy to announce a misleading low cost (just the final training run that optimizes a base model that in turn is possibly based on OpenAI) is also purposeful. After all, High Flyer, the parent company of DeepSeek, is a hedge fund - and I bet they took out big short positions on Nvidia before their recent announcements. The Chinese government, of course, benefits from a misleading number being announced broadly, causing doubt among investors who would otherwise continue to prop up American technology startups. Not to mention the big fall in American markets as a result.
Why should I care about the stock value of US corporations?
> I do think there’s also a big difference between scraping the Internet for training data, which might just be fair use, and training off other LLMs or obtaining their assets in some other way.
So if training of copyrighted work scrapped of the Internet is fair use, how would the training of the LLMs not be fair use as well? You can't have it both ways.
> Models from American companies at least aren’t surprising us with government driven misinformation
Is corporate misinformation so much better? Recall about Tienanmen Square might be more honest but if LLMs had been available over the past 50 years, I would expect many popular models would have cheerfully told us company towns are a great place to live, cigarettes are healthy, industrial pollution has no impact on your health, and anthropogenic climate change isn't real.
Especially after the recent behaviour of Meta, Twitter, and Amazon in open support of Trump and Republican interests, I'll be shocked if we don't start seeing that reflected in their LLMs over the next few years.
I think DeepSeek’s strategy to announce a misleading low cost (just the final training run that optimizes a base model that in turn is possibly based on OpenAI) is also purposeful. After all, High Flyer, the parent company of DeepSeek, is a hedge fund - and I bet they took out big short positions on Nvidia before their recent announcements. The Chinese government, of course, benefits from a misleading number being announced broadly, causing doubt among investors who would otherwise continue to prop up American technology startups. Not to mention the big fall in American markets as a result.
I do think there’s also a big difference between scraping the Internet for training data, which might just be fair use, and training off other LLMs or obtaining their assets in some other way. The latter feels like the kind of copying and industrial espionage that used to get China ridiculed in the 2000s and 2010s. Note that DeepSeek has never detailed their training data, even at a high level. This is true even in their previous papers, where they were very vague about the pre training process, which feels suspicious.