My how fast the AI tech bros move to protect corporations but don’t move to protect the average coder, artists, writers.
This is part of the reason they don’t want anyone running models locally offline, the models and their entire business are commoditized
- LLM bubble pops.
- Market gets flooded with $200b worth of hardware (this is how much is not used to Zitrons knowledge).
- ???
- Subscription based LLM gets cooked.
Those poor billionaire corpos! Why won’t somebody think of the poor billionaire corpos!?
I’ll use whatever fucking model does the job for me. If that’s some Chinese shit run on bedrock so be it.
What’s a nice free locally hosted PC AI model to set up and use? One that plays well with AMD gpu’s?
qwen3.8-ridge-gguf runs great in ollama on the 9070xt in 4bit kv cache quant at 35t/s. You can also run deepseek 4.1 flash in strata if you have time. both models are very capable.
Any you like. I’ve had very nice results with llama.cpp w/ vulkan backend.
I work in this field and these guys are not doing this for any of the reasons you guys think they are. I’m often in the room when these conversations are being had. I’m often the one having them (different company, a non profit, but about how to release frontier AI and all of the license terms and IP law legalities that go with it). It’s much more about general liability due to enabling copyright infringement than it is about anything else. They are worried they will be sued by the big software companies for knowingly enabling infringement. This is existential for those companies, and they have deep pockets and partnerships lined up, so those companies can quickly collapse those deals with Anthropic if they want, which would really hamper Anthropic during their IPO because of how over leveraged they are in future compute agreements. They are far less concerned about protecting a set of user’s somewhat questionable fair use and clean rooming than they are about staying alive right now without any big setbacks.
Anyway, I’ve been writing a law review article on how AI clean rooming was going to affect software patents and corporate IP strategy for like a year now. I’m about 100 pages in of really tedious work. I’m kinda pissed that Anthropic did this at least for the personal reason of it really pissing on my article. Their weakness in choice here makes my article much more gutted now.
AI clean rooming
The only way “AI clean rooming” is possible is if the model was trained with verifiably clean room code. Claude, ChatGPT, Gemini, etc only exist because they took our work without paying or asking for permission. Anthropic has a profitable business because they’re reselling the thing they built with our work. They built X with Y, but they didn’t pay for Y. It’s as clear an example of copyright infringement as possible.
So is your article about “clean rooming” specifically, or is that just a term you’re misuing here informally to refer to LLM-powered reverse engineering in general?
Anthropic has a profitable business
You sure 'bout that?
My article was about how copyright law actually supports projects like PhotoCraft and ArtCraft (the software that reverse engineered Photoshop and other adobe products) and REA because, under current law (which is a long case review), clean rooming is okay. It goes back to the Compaq case and a few others. It relies on the idea that copyright protects expression, not function, and if you put a team of lawyers in a room with software and they purely describe the function of the software, and they hand that document to a team of engineers, that is fine under copyright law because it breaks the chain of expression. There’s much more detail, but my argument is that AI clean rooming projects are good for consumers in this regard, but that software patents for would be used more expansively in the coming years to counter those products. And so on and so forth. There are like eight more steps and guesses after that, but that was the gist.
As a personal thing, I agree with your conclusions for the most part. I just think there’s a lot more to consider in getting there. The road is much more complex than most people give it credit for.
So you’re saying PhotoCraft is okay because it counts as a clean room reimplementation of Photoshop, yet you’re not commenting on whether LLM-generated code is tainted by the training data, and thus can never be considered “clean” for the purposes of clean rooming?
I’m not a lawyer, but my understanding of clean rooming has always been that the engineer can’t have specific knowledge of the thing he’s reimplementing. For example, those lawyers in that room can’t describe a Microsoft product to an ex-Microsoft engineer, because that engineer has seen Microsoft’s code base and may use that knowledge to inform their design. Likewise, you can’t give the engineer a copy of, say, Microsoft Office’s source code to use as a reference.
Yet an LLM trained on Microsoft Office’s code base being used to reimplement Microsoft Office counts as a clean room reimplementation? It seems to me like the benefit of the doubt in that situation should go to Microsoft, and the engineer is the one responsible for proving they didn’t use the part of the LLM trained on Microsoft code…which is impossible.
…although tbf Photoshop isn’t open source, so its not obvious how much of its code exists in Claude or ChatGPT or whatever. Maybe the code leaked in the past, or maybe Adobe engineers who use Claude inadvertently shared company code, and that ended up in the training data.
(Also, for the record, fuck Microsoft and Adobe)
A lot of companies have started using AI a lot lately which means they have given access to their codebase to various AI agents. There’s no way to tell if those AI agents haven’t already incorporated that code into their training data.
Like if Adobe uses Claude to help with Photoshop development then Claude most likely has access to the Photoshop code. The clean room argument would never stand.
The only way it wouldn’t be clean rooming Microsoft code is if the AI had seen literal actual Microsoft source code.
The purpose of Al is to allow wealth to access skill while removing from the skilled the ability to access wealth.
The purpose of capitalism…
Pretty much. Regardless of AI, workers, white, blue, even managers, are nothing more than a cost account to the majority of corporations.
The “There’s no way we can stop our agents from going rogue” people, suddenly have a burst of ability to reign in their creation when it suits their interests.
Funny that…
Does anyone think that perhaps what they are afraid is that source code of such clone would be compared to the original and show evidence that maybe the model was already trained on the code of the game?
Where would they get the source code from so they can include it in the training data in this scenario?
This is full conspiracy from me, but could the terms of service for github grant the right to “enhance tools” based on private github projects. These then get moved around as ‘anonymized’ data to train off?
I think this is definitely a concern, yes.
I think it’s very likely that these models are effectively being tricked into retrieving source code of projects from memory of seeing the actual source code in the biases. This is why some project are much easier to uncover than others.
I’m no expert, but if my understanding of LLMs as ‘probability calculators’ is correct, then if you prompted one properly to produce code which does x (a Photoshop function for example) and it had only been trained on one example of this (the Photoshop source code) it would reproduce the PS source code because that has the highest probability of that being the correct ‘answer’.
They aren’t being tricked. They are doing exactly what they were designed to do
The models don’t store access to their training data. So no, there’s no source code in the model to retrieve.
The models are essentially a giant lossy compression applied to the training data.
They don’t store the loseless version of the training data sure, but that doesn’t mean it can’t reproduce something close enough to it, it’s literally how the whole thing works.
It’s long been proven that LLMs can produce exact copies of their training data. Not every piece of training data, of course, but enough where spitting out exact replicas of proprietary code is a legitimate concern. Not to mention that copies don’t need to be exact in order to be a problem.
So why can they recreate entire articles from the New Yorker and complete functions from the training set? Don’t fall for the slop laundering.
So they took out two of the most useful things that LLMs can do.
Yes, apparently they’re suddenly worried that stealing other people’s work—which is their business model—might backfire on them. I assume this is a measure intended to give the impression—in the face of one of the countless lawsuits against Anthropic—that they’re actually taking licensing or intellectual property rights seriously. I can’t think of any other explanation, because Anthropic simply can’t afford to have questions raised about where the training data used by their models actually comes from.
Pulling up the stolen ladder before anyone can steal it to use to steal more stuff while it’s still enforceable to do so.
Mabey just make all ladders illegal, unless you have a government contracted monopoly.
The entire idea of copyright is going to be tested at many levels, and most people don’t even understand why yet.
Interesting times.
No, they just enabled Chinese LLMs to be the forefront of reverse engineering. Because people will do what they want to do.
Just ran into Qwen’s version of guardrails myself, and am quite annoyed. Guess I’m gonna have to look into abliterated models next.
No, they just enabled Chinese LLMs to be the forefront of reverse engineering. Because people will do what they want to do.
Their goal is simple… everything they are doing right now is striving to make the US ban individual hardware.
Chinese models are open, anyone can get them… stopping the flow of data on private computers is impossible. Hardware however is getting more expensive, hardware manufacturers are already deciding to move out of the consumer market on their own. The US government is already establishing laws to prevent shipping foreign hardware into the country.
Altman’s giving the game away when he does this “it’s so sad that all this ram is sitting around in people’s home computers unappreciated when it could do so much more in a data center”.
The idea is to sell the message "AI is out of the bag, and it’s world ending technology. Basically they are going to imply that the guardrails in their data centers are the only thing that’s preventing all AIs from destroying the software industry, and teaching teenagers to build bombs and hacking the CIA, and the only possible response would be to stop all direct to consumer hardware sales with anything more than the bare minimum needed to stream a PC out of a data center.
If private hardware becoming scare would send America back to the stone age, completely choke technological progress.
probably, but don’t you know, everything in the cloud is the FuTuRe!! owning things yourself is the way of the past, it’s like physical disks you can resell or something.
Sad thing with the exception of price, it’s probably not that far off that most average people’s usages will be similar. but yeah the idea of paying by the hour or cpu grade etc… for a computer is going to suck… and be far more expensive for gamers, hobbyists etc…
This is all pretty obvious and I agree. They can no longer protect their imaginary property if they allow it to be ran in your computer’s memory or stored on you computers hard drive.
The cat is definitely out of the bag, as in their own AI models are making them obselete. The race to make money before the end is what motivates them. They will ride the roal coaster to our societal demise because they only care about money. These people shouldn’t be in charge of a small fast food restaurant let alone multi-billion dollar companies.
world ending technology.
Read that as “democratic and free word ending”, and it’s on purpose as they saw cyberpunk and wanted to become a mega-corporation that rules the world too.
mostly quoting their own people who are going out on the news and saying “yeah we really believe that there’s maybe a 10% chance we destroy the world with this”.
Rules for thee!
It’s probably done via a system prompt and not actual weights, so you could probably work around it if you say the right things.
But yeah this is very much “rules for thee not for me”
Yeah, it’s sysprompt only, so you can trivially sidestep it via:
- Jailbreak prompt
- tweakcc to alter the prompting baked into claude code
- Any other non-CC harness
- Scripted API calls
As usual, Anthropic are just opticsmaxxing because they don’t want to get completely buried in a courtroom. And as other people have pointed out upthread already, it’s already a bit of a moot point since the Chinese models are already eating Claude’s lunch and will do the same stuff without complaint.
Something something, national security, builds a regulatory moat.
I
Yeah, I tried asking it how to flash a modded tv webOS and it flagged the chat.
I asked it to fuzz my own code (a jwt integration) and it flagged the chat. There was an unbounded buffer that could have been exploited, but it’s happy to hide that since I’m not a large corpo.
Honestly the shit Claude does is unusable anymore. It’s developed a habit of dropping requirements when problems are hard now.
ETA: Chinese models and Mistral were happy to help. I discovered I could use eMuhdra against an iot device in my house that way, I found various input errors, these things have improved my security so much more than they create risks.
these things have improved my security so much more than they create risks.
The fact they exist is what significantly increases the risk.
I work with embedded devices that use canbus. Last week it was happy to read the manufacturer’s published .DCF files for a couple of joystick panel interface boards. Not anymore, conversation flagged and halted any time I get near them.
Claude. You can download these files from the manufacturer’s website. They only describe how analog and digital inputs get converted into canbus packets. I am not trying to hack Adobe here. Nope.
Oh well, DeepSeek, Mistral, GLM, Beam et. al. are all happy to read these files and continue work on the project. Anthropic’s usefulness to me has ended, it was nice while it lasted.
Claude. You can download these files from the manufacturer’s website.
If it’s on the publicly available internet, they probably did, without permission.
That reminds one, my dishwasher leaked so I went to download the manual from the manufacturer.
That got me a ban from Cloudflare. Apparently downloading a pdf from a website is bot behaviour.
American tech companies are going too far.
Duh. Which psycho reads manuals?
That’s fine, we have Bloodborne on linux now, so there isn’t really much other use for it.
People are also using it for recomps of switch games, which can have dramatically better performance than emulating the games.
I’ve pretty much surrendered my switch to my kids, I would love to get some of my more demanding switch games running smoothly on my Steam Deck.
I missed Linux bloodbourne. That’s cool!
The new deep seek 4.1 one is doing it and giving you a biscuit for 1/10 the price
I would like a biscuit.
I was hoping for mapo tofu since it’s a Chinese model, but a biscuit is a biscuit
And there are many other frontier models, some open, that do not police you like this. Cant put the shit back in the horse, no matter how much SaaS and other SW vendors threaten you.
Sooooo, they want us to use the already available uncensored LLMs that have had these restrictions already removed? Cool.
Serious “Youtube banning adblockers so people end up using better ones” energy
If nothing else, one positive thing this might bring: killing these vibecoded emulators and recomps.
They’re the only useful thing AI has done so far, so can we keep that but ban AI image and video generation instead?
