Hacker News Reader

Comment by brezelgoring | Hacker News Reader

Steve Ballmer once called Linux and the GPL License a cancer because to copy a portion of code from a copyleft project, minimal as it may be, would make the whole project require a copyleft license.

If Github Co-Pilot includes GPL code then produced works should have GPL too, right? It is known that it produces verbatim copies of sections of code, so the 'derivative' explanation doesn't hold water.

Alternatives may be to copy CC0-only (which can't be guaranteed, also who'd believe you lol) or target license-only - as in, if my project is MIT, go with MIT projects to source from, if I'm GPL, include GPL and so on.

View on HN

>It is known that it produces verbatim copies of sections of code, so the 'derivative' explanation doesn't hold water.

For reference it's been shown to cough up code from Quake verbatim. This from John Carmack(?) also includes his profanity laden comments:

https://twitter.com/mitsuhiko/status/1410886329924194309

To be fair, that's probably one of the most copied pieces of code already.

And it was (apparently) copied from something or someone at Xerox

> It is known that it produces verbatim copies of sections of code

This happens only rarely, like under 1% of the time. It happens mostly for well replicated code and not so much for code that only appears once. It can be filtered out with search and bloom filters of ngram hashes.

But the prompter can goad the model into copyright infringement by quoting the start of a copyrighted text verbatim, and asking for completion. The longer and more precise the prompt, the higher the chance of regurgitation. So, when it happens, we're often "asking for it".

Both regurgitation and hallucination seem to be LM problems we can tackle. They are complementary - in one we don't want the model to replicate the training data exactly (be creative), in the other we don't want the model to invent facts out of thin air (be factual). Both can be tackled by using search for reference testing.

> I don't see regurgitation as a long term problem, it's just waiting for attention, probably wasn't top priority

The fact that Microsoft is wary of providing it with things like the Windows source code says all we need to know about how much it can be trusted.

it's not just trained on GPL code. it's trained on code of incompatible licenses, which means that the code is produces is potentially unlicensable in general. There's of course also the question of attribution, which many licenses require

I wonder who's decision was it to train the bots in 'code that is accessible to our scraper' and not in 'code we can sell derivatives products of'. I can tell the second group is quite small, so I understand the incentive, at least.

Maybe they chose the Uber strategy of 'What we are doing is bordering on illegal but by the time the bell rings we'll be valuable enough to write the law ourselves'.

Microsoft chose to not use all code that is accessible to their scraper. All proprietary closed repositories on github that large companies use are excluded. I can easily image all the lawsuit that would had happened if paying customers that kept company secrets hosted at github had been used and leaked by copilot.

The decision to only train it on "open" repositories was one of self preservation.

But the question realistically is, is it the developers responsibility because they are the ones creating the program.

It's not like copilot made you use the code, copilot didn't commit or release the app with the code.

Copilot didn't breach the copyright, you did. It's a tool. You used it. You released it.

Maybe there's a product which you can include to see what code of yours violates copyright in the future?

> If Github Co-Pilot includes GPL code then produced works should have GPL too, right

No.

Reply on news.ycombinator.com