Yeah, that is a pretty shitty consequence, but it’s nothing new or different from the strandard clean room techniques the industry has been using for decades.
A model that was trained on specific code that gets put out exactly the same doesn’t seem clean room to me.
Even if you argue it was turned into weights and probability and may not have originated from that specific code, by having included it in the training set you poisoned the model with no longer being able to claim plausible deniability - being produced rather than copied.
The transformation is still a machine and algorithm, even if it goes through a ton of obfuscation, distribution, and probability.
I’m more miffed about GPL or copyleft code going to proprietary software and then being profited from without any attribution. It’s a poison pill.
Yes! This concern absolutely cuts both ways.
Yeah, that is a pretty shitty consequence, but it’s nothing new or different from the strandard clean room techniques the industry has been using for decades.
I see it quite different from clean room.
A model that was trained on specific code that gets put out exactly the same doesn’t seem clean room to me.
Even if you argue it was turned into weights and probability and may not have originated from that specific code, by having included it in the training set you poisoned the model with no longer being able to claim plausible deniability - being produced rather than copied.
The transformation is still a machine and algorithm, even if it goes through a ton of obfuscation, distribution, and probability.
I mean it’s clean room but without even having to pay highly skilled engineers to do all the labour. It’s pretty catastrophic to the field.
Yeah, fair enough.