@pyromuffin Hello! I mention this later in the thread but I want to be very clear I have no criticism of your post, and was not trying to criticize your post. I apologize if it came across otherwise.

The post just made me Think About Something, is all.

@mick Is that a decompilation? Or is it a art program designed to look like adobe photoshop in screenshots?

@nerpulus i have yet to see any evidence decompilation or even "reverse engineering" as that article calls it took place. i think he might have just made some computer programs that kind of look like adobe photoshop. i haven't really investigated deeply though due to not caring what ai coders are doing

@lanodan "oh right… there are forms of intellectual property other than copyright…"

@discatte LOL

@mcc Not a lawyer, but having kept an eye on recent legal developments, my layperson’s understanding is:
- So far, courts largely think that *training a general-purpose model* on a wide variety of unrelated material can be fair use, because it’s sufficiently transformative and the resulting product fills a very different market niche than any particular item in the training set
- However, a recent case held that training *on a competitor’s IP* in order to produce a product that competes directly with theirs is *not* fair use (this was Thomson Reuters vs Ross Intelligence, regarding an AI search engine trained on Westlaw headnotes)
- neither of those has gone to SCOTUS and neither is blackletter law at this point (few things about fair use jurisprudence are)
- but on the decomp front, things are a bit more clear-cut: the LLM is simply a tool being used to process disassembly data to produce a result that directly competes with the original product, so it seems like it’d be pretty obvious that the decompiled output is a non-fair-use derivative work of the input machine code, and thus the property of the original IP holder
- pure black-box *reimplementations* (produced without giving the agent direct access to the original product’s machine code) are a lot more of an open question! my reflex is that they’re probably fine, as long as they don’t run afoul of trademark, trade dress, or patent protections; from a pure copyright perspective
- Google vs Oracle might be relevant precedent, amusingly enough? though “reimplementation of APIs declarations is fair use” doesn’t necessarily transfer directly to GUI apps
- the whole “have one agent look at the original code and describe its functionality in English to another, which reimplements it” thing is *probably* too cute, but tbh who knows; feels like a situation that’d depend on how well the two sides argued their cases

@mcc this is why I grind my teeth at … certain commentators' position that we should not support copyright litigation against AI vendors because it "sets precedents" that can be used against "the commons". it betrays a shocking naïveté about how precedents are used. if we can score a victory that is an unalloyed good, sufficient precedents already exist and are already applied unfairly that no *new* anti-AI precedent is going to be used against legitimate fair use where existing law wouldn't be

This is a circuitous way of saying. Occasionally, I see someone saying "I think it's great the government is allowing AI training/regurgitation because that weakens copyright, and I don't like copyright" and you fool. You stooge. You instrumentalized doe-eyed naif. AI training/regurgitation is copyright working normally. Corporate enclosure is what copyright is *for*. Upward transfer of wealth is what copyright is *for*.

@mcc I would absolutely love to see that happen to someone else in court