ChatGPT and Anthropic are now under restriction. I think it is logical to think that HuggingFace will be next. Is it time to create Piratebay for models? It is my opinion that only reliable access to LLM is the machine on your desk and the models in your storage. It is more prudent to prepare for environment where internet access can be taken away anytime.
I think this post deserves more engagement. I think some people think it won’t happen. But it’s likely it will to at least some extent, bare minimum we might be gated from downloading new models unless their deemed “safe” by the white house or huggingface may be required to do KYC.
I disagree, maybe farm for engagement on X, Reddit, Zhihu, or some other social platform. I hope this space remains the one technical forum on the internet for DGX Spark discussions, and not devolve into another social media forum.
I would like to clarify my intention. Maybe I should be more specific in my original post. I was seeking advice on a strategy for models storage. I think people might have heard of 3-2-1 storage strategy. Due to the model size, some medium like optical storage might not be suitable due to its limitation on per disk storage size. Other options are tapes, hard drive and SSD. The tricky part in my opinion is your selection of remote location storage. Government regulation has become a potential issue. It is my opinion that online sources are no longer trustworthy.
US government gave Anthropic 90 minuets to shutdown Fable. This is give us a reference on how long you might have to react in the future.
Right now my solution is hard drive due to its cost and portability. I have M-DISC but it is just too slow and storage capacity is small.
Why is this cluttering up a DGX Spark forum again? Is this DGX Spark specific? Seems well suited to Stack Overflow.
I’m downloading to 1 NAS which mirrors to an offsite NAS; good enough for me for now.
Nothing a VPN won’t stop, they’ll probably move operations to another country if it actually gets to that point. I would be surprised if the US government tries anything here because it’s seemingly impossible to actually restrict access to open source models.
I agree today, but I’ll bet within 5 years we have something similar to China’s great firewall, but AI powered… to protect us.
Maybe I’m just not very smart, but to me these restrictions come off as desperation from the companies since they’re losing tens of billions every year (OpenAI lost around $38.5 BILLION in 2025), have hit the wall hard with scaling, and have no idea how to move forward to something substantially better, let alone something that makes profit.
The CEO of Amazon went to the Trump administration shortly before the Fable 5 decision and told them the models are dangerous because they’re so powerful. Are we to believe that the CEO of a company that’s invested tens of billions into Anthropic was doing this to hurt his company’s investment? Or is it simpler to assume it’s more fakery from a company with a known history of fakery?
I think these companies know the writing is on the wall, the AI bubble will be popping soon, and that they can’t release a model that will live up to their hype, so the next best course of action for them is to pretend the models are too powerful to release. Is this not obvious to anyone else? I don’t think anyone will care about random open source models, since this incentive doesn’t exist for them.
I completely agree. The definition of fable is “a succinct fictional story, in prose or verse, that features animals, legendary creatures, plants, inanimate objects, or forces of nature that are anthropomorphised, and that illustrates or leads to a particular moral lesson (a “moral”), which may at the end be added explicitly as a concise maxim or saying.” which when you think about it sounds a lot like a joke pointed at us consumers. It almost feels like they’re making fun of us and they know they’re boned.
What fever dream is this thread ?! Go to Reddit when you want to philosophy about the Gov blocking your local AI 😂
There is Model Scope, Chinese hugging face. In general worry is open source tech will be illegitimate unless complying with spying rules, gov backdoor, age ver. People worry github will be restricted.
I think the situation is more fluid and government action seem more arbitrary when compare with the past. China is a wild card, their government is even less transparent than western countries.
2.5" portable hard drive is my storage of choice. For remote site storage, I just carried them over. Put them in ziplock bag with desiccant packets(make sure storage area will not go above 90 C). High value data(proprietary data) go onto M-DISC. Storage strategy is still 3-2-1.
If you are building a tool that you depends on, backup strategy should be planned to minimize interruption and improve chances of recovery.
I normally avoid speculation, but the likelihood that this is purely IPO-driven seems high to me. And history shows: try to take down HuggingFace, you end up with a decentralized HuggingFace. Every takedown of open infrastructure just makes it harder to take down next time. Not a fortune teller, just pattern matching.
Yeah there’s a ton of lying going on with these companies, like all of the “AI layoffs” when the reality is no AI is replacing jobs, just those companies are doing badly financially.
I just want to point out that companies will remove models that they generated. Some one just posted in reddit that Microsoft took down FastContext model from HuggingFace and Github. Microsoft Vibevoice is gone too. You need to backup that data that you depends on.
Someone actually get the LLM only torrent search site working now:
Boy, I like the idea conceptually (Go Decentralization!), but it’s also a real can of worms. I’d say better make sure someone is archiving all the major model pages on HF using something like Internet Archive so that the hashes can be as easily verified to the source as possible, make sure to only download safetensors model files and not .bin/.pt/.pte/.pth/.ckpt/.onnx and so on, and in general don’t use that other than a last resort and only if one knows what they’re doing. That’s my two cents. Now that the devs can use Fable 5 again maybe they can use it to implement features like these 😁
They are actually voted for a law (did not pass) that would criminalize importing (downloading) and USING Chinese AI models, up to 20 years in prison. This is just the start :(