Ras wrote: ↑Mon Jul 13, 2026 1:03 pm
chrisw wrote: ↑Mon Jul 13, 2026 11:57 amIf an LLM, asked to code an idea in the context of the chess engine eco system (100 sources or more?) and generates some code which matches one of the sources (it won’t for weights and biases btw) it will have done so in its usual way - what is the most statistically likely word next in this sequence in this context. That’s not copying.
If it reproduces verbatim code passages, or verbatim translations from another programming language, this
is copying. See the legal case I provided as example. Obviously, ChatGPT did not store the lyrics as a text file. But it did encode it sufficiently in the weights to reproduce it. Which isn't even astonishing because the models get larger and larger, and the risk of "overtraining" has been a known NN phenomenon for decades.
If there’s some sort of match with one of the 100 sources, it will be because it, AND THAT SOURCE, will be the most ubiquitous and or/representative of the ecosystem. Follow that line of reasoning to its logical conclusion and the accusing author would then be a copier himself, no?
Ideas and concepts are not copyrightable. Only a specific implementation - if the creative value is sufficient (e.g. some quicksort implementation will hardly count IMO). But since Andrew made the claim, the proof of burden is on him. The general claim that by the way how LLMs work, there can be copyright violations, that is correct - but the mere possibility is not enough. Let's see whether Andrew is willing to share some specific analysis, and whether it has merit.
Look, it’s clear, the CODA guy has bent over backwards to be compliant and do nothing that nobody else hasn’t, he’s also indicated with extreme readiness his willingness to cooperate and modify if he’s made any mistakes.
That doesn't change the fact clear attribution is a condition of pretty much all OSS licences. The problem with AI coding is that you can't even really know whether you have a problem. The provenance tracking gets lost through the LLM. You don't even know what exactly you need to comply with. This isn't specific to CODA. It's even more of a problem for companies who develop proprietary software and can accidentally pull in e.g. GPL code that no SBOM auditing is even aware of.
Right now, the legal debate largely focuses on the AI providers themselves, see the linked lawsuit. But there's a whole next rabbit hole waiting with the consumers of AI if they use the AI output.
AG has basically dropped the case with respect to the thirty or so GPL engines, CODA is a GPL release and it does attribute and it is honest about it.
AG continues on with two engines that are AGPL as opposed to GPL, the difference being requirement to provide source code if your engine is playable over the cloud (which CODA is via Lichess). Looks like CODA has not done that, has acknowledged it, said it's an oversight and will fix it. It's not exactly an earth shattering breach of a licence, and surely not enough for claims of theft, deceit and dishonour.
So where is his gripe? What has CODA done wrong? (getting out of wholesale copying charges is standard practice in computer chess, you just prefix the GPL announcement). What has CODA done that is different? Well, it's got a lot of Elo and authors who've spent time copying ideas, tuning and testing feels usurped. Well, okay, such is life, a process has arrived which automates human work. But it's neither theft, deceit nor dishonour.
