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?
There S1 claims so, but, also, it’s financials cut off… suspiciously right as a bunch of pre payments for their services came in, and before they deferred payments for compute went out.
Except that it doesn’t? They claimed 33bn in revenue, but they had 104bn net loss. So no, they were not profitable. In their IPO they literally claim to not be profitable and that they don’t ensure they’ll be profitable “in the future”. Not the near future, just the future as a whole.
So they were claiming to have been net profitable on certain things in q2, which again, more to the point, is them trying to shine a turd and make it look like they’ve got a future.
I would counter that “AI clean rooming” is not clean rooming, because they are feeding the actual code to imitate directly. Even ignoring the training, it has, verbatim, the object code and is trying to make code that compiles to match the object code. It isn’t deriving the functions to a spec and the using the spec to go back to code.
I would further counter that the Compaq case dovetails with the idea that APIs are broadly not copyrightable in and of themselves. Compaq cloned a thin ABI layer, which was not considered substantial by itself, so the argument rested with whether the implementing backend necessarily infringed in resemblance or if resemblance derived from merely following a similar asked. If you clone an entire application down to the human presentation, I think no “clean room” argument would overcome that you cloned far more than some thin application to library interface.
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.
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.
I feel like the only real way to do AI clean rooming is to have the AI run the decompiler, clean up the output if necessary, and then write a spec document that incorporates no part of the code only a description of what it does and its constraints.
You then reimplement the code by hand, only reading the spec.
20 year old patents would be an improvement time-wise over 100+ years of copyright, but a horrible thing given how much more annoying it is to get them invalidated.
deleted by creator
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?
You sure 'bout that?
There S1 claims so, but, also, it’s financials cut off… suspiciously right as a bunch of pre payments for their services came in, and before they deferred payments for compute went out.
Except that it doesn’t? They claimed 33bn in revenue, but they had 104bn net loss. So no, they were not profitable. In their IPO they literally claim to not be profitable and that they don’t ensure they’ll be profitable “in the future”. Not the near future, just the future as a whole.
So they were claiming to have been net profitable on certain things in q2, which again, more to the point, is them trying to shine a turd and make it look like they’ve got a future.
I get your point but their argument is basically moving the goalpost. The important thing is if the company is net profitable, and they’re not.
I heard, big part of those losses are the debts paid in stock, so when their stock rises the debts rise and the difference is written up as losses.
Not an economist though, so I might be completely wrong
deleted by creator
I would counter that “AI clean rooming” is not clean rooming, because they are feeding the actual code to imitate directly. Even ignoring the training, it has, verbatim, the object code and is trying to make code that compiles to match the object code. It isn’t deriving the functions to a spec and the using the spec to go back to code.
I would further counter that the Compaq case dovetails with the idea that APIs are broadly not copyrightable in and of themselves. Compaq cloned a thin ABI layer, which was not considered substantial by itself, so the argument rested with whether the implementing backend necessarily infringed in resemblance or if resemblance derived from merely following a similar asked. If you clone an entire application down to the human presentation, I think no “clean room” argument would overcome that you cloned far more than some thin application to library interface.
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.
I feel like the only real way to do AI clean rooming is to have the AI run the decompiler, clean up the output if necessary, and then write a spec document that incorporates no part of the code only a description of what it does and its constraints.
You then reimplement the code by hand, only reading the spec.
20 year old patents would be an improvement time-wise over 100+ years of copyright, but a horrible thing given how much more annoying it is to get them invalidated.