Do you think AI "things" like Midjourney or ChatGPT will have or are already having some kind of "piracy" around them?

Of course, I'm not in favor of this "AI slop" that we're having in this century (although I admit that it has some good legitimate uses but greed always speaks louder) but I wonder if it will suffer some kind of piracy, if it is already suffering or people simple are not interested in "pirated AI"

39 comments

I'm not sure what you're asking, but it seems you're not aware of the huge AI model field where various AI models are already being publicly shared and adjusted? It doesn't need piracy to see or have alternatives.
The key to hosted services like ChatGPT is that they offer an API, a service, they never distribute the AI software/model.
Other kinds of AI gets distributed and will be pirated like any software.
Considering piracy "around" them, there's an intransparent issue of models being trained on pirated content. But I assume that's not what you were asking.

I'm pretty sure those things are trained on content which was obtained without paying royalties to the creators, hence by definition pirated content - so that would count as "piracy around them".
On the opposite side, as far as I know the things created with Generative AI so far can't be copyrighted, hence by definition can't be pirated as they've always belonged to the Public Domain.
As for the engines themselves, there are good fully open source options out there which can be locally installed (if you have enough memory in your graphics card) and there seem to be thriving communities around it (at least it looks like it from what bit I dipped into that stuff so far). I'm not sure if it's at all possible to pirate the closed source engines since I expect those things are designed to be deployed to very specific server farm architectures.
- Training is transformative use. Sluicing data through a pile of linear algebra, to mechanically distill the essence of words like "fantasy," is not what copyright protects against.
- There are quite a few options for running your own LLM. Ollama makes it fairly easy to run (with a big selection of models - there's also Hugging Face with even more models to suit various use cases) and OpenWebUI makes it easy to operate.
  Some self-hosting experience doesn't hurt, but it's pretty straightforward to configure if you follow along with Networkchuck in this video.
  
  Any that are easier to set up on a phone? I tried something before but had trouble despite having enough RAM.

I mean they stole people's actual work already, so they're the bad kind of pirates.
- How is what they're doing different from, say, an IPTV provider?
  
  I'm talking about the data sets LLMs use, just so we're on the same page.
- Just like people steal movies from the high seas? I hope this is sarcasm.
  
  Nope, there's a difference there in that they aren't taking something from ordinary people who need the money in order to survive. Actors, producers, directors etc have already been paid and besides, hollywood etc aren't exactly using that money to give back to society in any meaningful way most of the time.

Incredibly weird that this thread was up for two days without anyone posting a link to the actual answer to OP's question, which is g4f.

Not pirated. But my country, Spain, released an open AI model completely for free. Everything is open. The training data the models and everything. It's supposedly ethically trained with open data(I have not personally dig in the training data but it's there published).
It's focused on spanish and regional languages of spain. But I think it can also do things in English.
Not piracy per se, as it's completely legal. But there's something you don't depend on any bussiness to run.
- First i hear about it, any links?
  
  Of course.
  http://alia.gob.es/eng/

That makes no sense. Define pirated AI first.
- Yeah the whole of generated AI feels like legal piracy (that they charge for) based on how they train their data

Weight leaks for semi-open models have been fairly common in the past. Meta's LLaMa1.0 model was originally closed source, but the weights were leaked and spread pretty rapidly (effectively laundered through finetunes and merges), leading to Meta embracing quasi-open source post-hoc. Similarly, most of the anime-style Stable Diffusion 1.5 models were based on NovelAI's custom finetune, and the weights were similarly laundered and became ubiquitous.
Those incidents were both in 2023. Aside from some of the biggest players (OpenAI, Google, Anthropic, and I guess Apple kinda), open weight releases (usually not open source) have been become the norm (even for frontier models like DeepSeek-V3, Qwen 2.5 and Llama 3.1), so piracy in that case is moot (although it's easy to assume that use non-compliant with licenses is also ubiquitous). Leakage of currently closed frontier models would be interesting from an academic and journalistic perspective, for being able to dig into the architecture and assess things like safety and regurgitation outside of the online service shell, but those frontier models would require so much compute that they'd be unusable by individual actors.

There are groups that give access to pirated AI. When I was a student, i used them to make projects. As for how they get access to it? They usually jailbreak websites that provide free trials and automate the account creation process. The higher quality ones scam big companies for startup credits. Then there are also some leaked keys.
Anyways thats what i would call "pirated AI". (Not the locally run AI)

Meta's model was pirated in a sense, someone leaked it early last year I think, but Llama isn't that impressive, and after using it on whatsapp seems like nothing got better.

Not sure it it counts in any way as piracy per say, but there is at least jail broken bing's copilot AI (Sydney version) using SydneyQT from Juzeon on github.

Its already free, you cant pirate cloud services but stable diffusion is free, deepspeek is free, you just need the hardware to run it

Well... sufficient local processing power would enable personalized "creators" which are pre-trained to provide certain content (e.g. A game). Those thingies will definitely be pirates, hacked and modded.
As they currently already are...

Well yeah, self-instruct is how a lot of these models are trained. Bootstrap training data off a larger model and fine-tune a pre-existing model of that.
It's similar but different.

There already is. You can download copies of AI that are similar or better than ChatGPT from hugging face. I run different models locally to create my own useless AI slop without paying for anything.
- Which model would you say is better than GPT-4? All I tried are cool but are not quite on GPT-4 level.
  
  The very newly released Deepseek R1 "reasoning model" from China beats OpenAI's o1 model on multiple areas, it seems – and you can even see all the steps of the pre-answering "thinking" that's hidden from the user in o1. It's a huge model, but it (and the paper about it) will probably positively impact future "open source" models in general, now the "thinking" cat's outta the bag. Though, it can't think about Tiananmen Square or Taiwan's autonomy – but many derivative models will probably be modified to effectively remove such Chinese censorship.
- Are you referring to ollama?
  
  No because that is just an API that can run LLMs locally. GPT4All is an all in one solution that can run the .gguf file. Same with kobold ai.

You can just run Automatic1111 locally if you want to generate images. I don't know what the text equivalent is though, but I'm sure there's one out there.
There's no real need for pirate ai when better free alternatives exist.
- There’s no real need for pirate ai when better free alternatives exist.
  There's plenty of open-source models, but they very much aren't better, I'm afraid to say. Even if you have a powerful workstation GPU and can afford to run the serious 70B opensource models at low quantization, you'll still get results significantly worse than the cutting-edge cloud models. Both because the most advanced models are proprietary, and because they are big and would require hundreds of gigabytes of VRAM to run, which you can trivially rent from a cloud service but can't easily get in your own PC.
  The same goes for image generation - compare results from proprietary services like midjourney to the ones you can get with local models like SD3.5. I've seen some clever hacks in image generation workflows - for example, using image segmentation to detect a generated image's face and hands and then a secondary model to do a second pass over these regions to make sure they are fine. But AFAIK, these are hacks that modern proprietary models don't need, because they have gotten over those problems and just do faces and hands correctly the first time.
  This isn't to say that running transformers locally is always a bad idea; you can get great results this way - but people saying it's better than the nonfree ones is mostly cope.

39 comments