American companies are spending enormous sums to develop high-performing AI models. Distillation attacks are attempting to maliciously extract them — and nobody is doing much to stop it.
Nobody is doing much to stop the American AI companies crawling the web to scrap tons of licensed content to illegally use in their training, either.
Oh no, wouldn’t anyone think of the billion dollar companies? The Chinese are stealing the models that they have spent so much effort on getting all the training data. What a shame.
If they open source it then that’s a win. Closed source models don’t help anyone because when the company goes bust they need to be reinvented again. People like to talk about the advancements capitalist industry has made but if they never publish any of it because of “tRaDE sEcrETs” they might as well have never done it because the next person will have to reinvent it when they go bust or kill it for money.
This article should be titled undertaxed american corporations waste trillions on ai.
Whoever wrote this article didn’t even bother to do the most basic of research.
DeepSeek fully admitted they started with ChatGPT outputs to train its model. And then they released it as an open-source model, so that everybody else can “steal” their work. On the image/video front, the general public has created every possible variation on top of every model you can think of. On top of that, any model that has ever been released with full weights has been spun into whatever variation or VRAM size you want.
The ugly truth that the American companies want to hide is the fact that they are spending trillions of dollars on an oligopoly that they can’t keep long-term. They hope that they can just keep spending more money to add more billions of parameters to their models, and keep technologically competitive with the secondary open-source models. But, they’ve already ran into diminishing returns over a year ago, and the global compute sector physically cannot keep up with demand for another cycle of even more diminishing returns.
The other factor is that realistic miniaturization of models is already here. Some of the smaller sizes aren’t as effective as the 250GB models they use on cloud-based services, but you can still do a lot with a 16GB or 24GB video card, using models of those sizes. Optimization and LLM quantization is getting better and better each year. The AI bubble burst is going to force a cascade shift into a new era of localization. Everybody is sick to fucking death of renting and subscribing to everything. Us pirates already do so on the media front, and soon localization of LLMs is going to become way more popular.
The question isn’t “Can people steal the tech?”. It’s “how long will people notice that it’s already happening?”
I would reckon that China is perfectly satisfied to let us be the sole host of the thing that is rapidly destroying our economy and trust in all media from the inside out.
China is 38%, and growing, of the world’s investment in LLMs
Most of that goes toward implementation (data centers) and chip manufacturing. China is making money on compute services and maintaining capability parity on software the good old fashioned pirate way merely to prevent a technology gap with the US, as is their way.
That is not “allowing us to be the sole host of the thing that is destroying our economy and trust in all media from inside out”. That is keeping parity with it. China is also having major issues with fabricated media from AI. The Chinese government has also latched on to AI, as many others, to manipulate media, and many other police state things. Their economy is heavily, heavily, invested in the success of llms. When this bubble bursts, it will be bad for every major economy on earth, as they are all disproportionately invested in this.
Can China just steal what America just stole from everyone else?
Models getting better does give extra information for making newer models better too. China publishes far more advanced research than US models “steal”, and they open source exceptionally strong/fast models that US can also steal from.
Calling it a “Distillation Attack” is wild. Get fucked Anthropic.
Some of the terms that have been coined to describe stuff related to AI are just so funny.
“Prompt injection attack”, also known as… asking nicely for the chatbot to do a thing.
Yeah, because American LLMs are so immensely useful that people are throwing money at them.
they did. its called deepseek
yadayada, more moronic ‘China baad’ propaganda.
lol. love when people are panicking about this when it means that these things are basically interchangeable anyways. Didn’t someone at Google write a memo that was like “we’re kinda fucked b/c you can re-create this stuff with enough resources” like 2 years ago?
Didn’t someone at Google write a memo that was like “we’re kinda fucked b/c you can re-create this stuff with enough resources” like 2 years ago?
Basically, yes. They were specifically decrying the amount of open-sourcing they and their American competitors were doing, because capitalism, of course. Around this time, we had examples like StabilityAI’s StableDiffusion and Meta’s LLaMA as open-source models. And around this time, everybody else started closing their models, despite the fact that the research kept on going out in the open. StabilityAI kept their models open, mostly because they had no choice, but the attitude shifted towards profitability.
So, China took the open-source mantle, and these open/closed lines are being drawn strictly around national divisions as this American vs. China slant. Which is mostly a diversion of the real battle.
I truly wish they could steal it away.
