[habagav]

You could do “clean room engineering” approach where the reversing agent generates a specification from its findings, and then have a separate agent reimplement the code without seeing the original binaries/code.

You’d just have to make sure the specification doesn’t include actual source snippets (the AI will try this if you don’t specify). Pseudo code would be sufficient I guess where necessary.

[alberto-m]

Unless you develop your AI agent from scratch or you clone a never-released game, it would be extremely easy for the rightholders to claim that both agents have most certainly ingested the binary during their training phase, since it's well known that the hyperscalers have pirated everything that could be pirated to train their LLMs. Which is why malus.sh is a parody, not a real service.

One should be honest about what one builds. The F-15 project does that: the aim is the reconstruction of the original game, down to the opcodes; on the other hands it requires the user to provide the original game assets.

[skerit]

> it would be extremely easy for the rightholders to claim that both agents have most certainly ingested the binary during their training phase

Ingested the binary?

[gmerc]

Reality would disagree. If the models are large enough few legal consequences are to be expected because governments don’t want to lose out on AI investments. And in the US the industry owns the executive, legislative and well, thoughts and prayers.

This creates an interesting problem via the normative power of the present - only people willing to accept the erosion of copyright can partake in the AI economy and since the AI economy has become essentially all growth, money and power rapidly concentrates on these people.

There won’t be a legal reckoning because the regulatory capture is complete - these models are “national security level” too big to fail now.

[justinclift]

> ingested the binary during their training phase

That's extremely unlikely. The pirating seems to have been of eBooks, which were first converted to text, losing graphics and such too.

That's my understanding anyway, from reading a few reports. :)

[luckydata]

Just no, this isn't a thing at all.

[habagav]

If they try to claim that then they need proof, right? Good luck getting that.

[skerit]

Yeah, that approach makes the most sense.