Re:I do not agree with Github's use of copyrighted code as t...
by dathinab on Saturday July 03, 2021 @04:23PM
(#27724588)
Fair use for training and "independent creation" are one think a AI "remembering and mostly verbatim copying code over" an another.
Many of the current Machine Learning application try to teach AI to understand the concepts behind their training data and use that to do whatever they are trained to do.
But most (all?) fail to properly reach the goal in any more complicated cases, at least the kinds of models which are used for things like Copilot (GPT-3?).
Instead what this models learn can be described as a combination of some abstract understanding and verbatim snippets of input data of varying size.
As such while they sometimes generate "new" things based on "understanding" they also sometimes just copy things they have seen before!! (Like in the Quake code example where it even copied over some of the not-so "proper" comments expressing programmers frustration).
It's like a human not understanding programming or english or Latin letters but has a photographic memory and tries to somehow create something which seems to make sense by recombining existing verbatim snippets, sometimes while tweaking them.
I.e. if the snippets are small enough and tweaked enough it's covered by fair use and similar, BUT the person doing it doesn't know about this, so if a large remembered snippet matches verbatim it will put it in effectively copying code of a size which likely doesn't fall under fair use.
Also this is a well known problem, at least it was when I covered topics including ML ~5 years ago. I.e. good examples included extracting whole sequences of paragraphs of a book out of such a network or (more brilliantly) extracting thinks like peoples contact data based on their names or credit card information (in case of systems trained on mails).
So that Copilot is basically guaranteed to sometimes copy non super smalls snippets of code and potential comments in a way not-really appropriate wrt. copyright should have been a well know fact for the ML specialist in charge of this project.
Many of the current Machine Learning application try to teach AI to understand the concepts behind their training data and use that to do whatever they are trained to do.
But most (all?) fail to properly reach the goal in any more complicated cases, at least the kinds of models which are used for things like Copilot (GPT-3?).
Instead what this models learn can be described as a combination of some abstract understanding and verbatim snippets of input data of varying size.
As such while they sometimes generate "new" things based on "understanding" they also sometimes just copy things they have seen before!! (Like in the Quake code example where it even copied over some of the not-so "proper" comments expressing programmers frustration).
It's like a human not understanding programming or english or Latin letters but has a photographic memory and tries to somehow create something which seems to make sense by recombining existing verbatim snippets, sometimes while tweaking them.
I.e. if the snippets are small enough and tweaked enough it's covered by fair use and similar, BUT the person doing it doesn't know about this, so if a large remembered snippet matches verbatim it will put it in effectively copying code of a size which likely doesn't fall under fair use.
Also this is a well known problem, at least it was when I covered topics including ML ~5 years ago. I.e. good examples included extracting whole sequences of paragraphs of a book out of such a network or (more brilliantly) extracting thinks like peoples contact data based on their names or credit card information (in case of systems trained on mails).
So that Copilot is basically guaranteed to sometimes copy non super smalls snippets of code and potential comments in a way not-really appropriate wrt. copyright should have been a well know fact for the ML specialist in charge of this project.