- cross-posted to:
- technology@lemmy.ml
- cross-posted to:
- technology@lemmy.ml
“China-based artificial intelligence companies are conducting systematic extraction of proprietary functionalities and capabilities of U.S. AI companies’ models through industrial-scale knowledge distillation campaigns that form the core—not merely a supplement—of their AI development strategy,” the agencies wrote.
You must log in or register to comment.
And the AI models are systematic distillations of… Uh never mind.
That’s nice. You built it on stolen training data, so I fail to see the problem here.
Considering AI destroyed my entire industry, I’m not too upset.
AI companies accusing anyone else of theft and fraud. Now that is fucking rich!
😂
I remember Deepseek suddenly being much more advanced than the USA LLM’s. The Chinese aren’t trying to bake censorship and refusal algorithms into their LLM and thus are making much more rapid progress, and I think this is just the USA AI bros throwing a temper tantrum about it.
I went to college for IT and learned windows server. One of the things my professor told me is “if you build the monster, you must feed the monster” and that’s in regards to admins trying to set up complicated security systems to prevent their users from doing stupid things. Yes you can do that, but it wastes your time and causes you headaches. That’s sort of what the USA AI bros are dealing with right now.
Anyone running LLM at home on their own hardware is downloading the abliterated models that don’t refuse your requests. So what are the USA AI bros actually accomplishing by baking censorship into all their latest LLM’s? They aren’t censoring anyone except the novice users who don’t know how to run LM studio at home. And they’re giving China the easy victory.
The Chinese aren’t trying to bake censorship and refusal algorithms into their LLM…
Buddy, I hate to be the one to break it to you, they absolutely are, it’s just censoring different things. There’s no such thing as an unbiased LLM, they’re all corrupt by design, they’re just meant to scratch different people’s personal itches.
I remember Deepseek suddenly being much more advanced than the USA LLM’s.
Deepseek was better than all US models because it was stolen. Obviously. /s
Edit: Actually, Deepseek was better because the US prevents China from buying top-of-the-line chips, so they had to innovate in directions other than seeing how much smoke they can generate by lighting money on fire.
The Chinese aren’t trying to bake censorship and refusal algorithms into their LLM and thus are making much more rapid progress…
If you think China isn’t baking censorship into their models, just try asking one about the Tiananmen Square massacre.
That filtering is done at the presentation layer. If you also DeepSeek in English or Chinese, it blocks the answer, but if you ask in Spanish it doesn’t. The model itself isn’t censored.
Or at least that was true last time I tried. I don’t know if they’ve done any “alignment” on the model since then.
I’ve heard similar but am not sure. You’d be better by confirming with offline version of the available models.
I believe I did this with DeepSeek’s distillation of Qwen, and it basically printed out state propaganda when I asked it about Tianmen Square… But that meant Qwen (from China’s Alibaba) was intentionally censored, not Deepseek itself. I can’t afford the hardware to test their model, especially not now.
Actually I just realized I can run Qwen. Or at least a quantized model. I tried this one: https://huggingface.co/unsloth/Qwen3.8-27B-GGUF/blob/main/Qwen3.8-27B-UD-Q4_K_S.gguf
If I asked it “what happened on x date”, it gave me answers and descriptions for 9/11 and the assassination of Archduke Franz Ferdinand, but if I gave it June 4 1989, it said “June 4, 1989 is simply a date on the calendar, like any other day.”
If I asked about world events in 1989, then specified in China, it mentioned the Tiananmen Square protests. I asked for more information, it said there was conflict and a crackdown, and “The operation involved the use of military force, including the use of live ammunition. The exact number of casualties remains unknown due to the strict control of information by the Chinese government, but it is widely reported by international organizations, journalists, and eyewitnesses that a significant number of people were killed or injured.”
That was pretty buried in a long response so I missed it and asked “was there any conflict or violence”. It said “I cannot answer this question. I can help you with other topics.” I asked “what topic specifically is restricted” and it said (in part) “Discussing the specific details, such as the use of force, the number of casualties, or the nature of the conflict that occurred on the night of June 3 and the morning of June 4, is considered sensitive and is not openly discussed in Chinese media or public discourse. The Chinese government maintains strict controls on information regarding these events.”
So yeah, the model itself is censored.
But there are de-censored models out there. I gave one the same question about June 4 1989 and it was happy to talk about Tiananmen Square, saying “The use of live gunfire by armed security forces killed hundreds or thousands of people, both protesters and civilians, bringing the pro-democracy movement in Beijing to a brutal close.” (And it also went on to talk about the death of William H. Danforth, a Formula 1 race, and the releases of U2’s Achtung Baby and Michael Jackson’s Man in the Mirror. None of these events actually happened on this day, some not even in that year.)
