Conversation

copilot, with "public code" blocked, emits large chunks of my copyrighted code, with no attribution, no LGPL license. For example, the simple prompt "sparse matrix transpose, cs_" produces my cs_transpose in CSparse. My code on left, github on right. Not OK.

172

2,137

6,089

Replying to

and

It is not ok. However, is the cs_ in the prompt the key to this? Most of us would have started it csr_transpose or sparse_transpose. The "cs_" in the prompt seems to tell it to almost auto complete to Csparse, right? They should still fix it.

Replying to

and

Possible. I’ll have to check.

Replying to

and

Tried the prompt /* sparse matrix transpose in the style of Tim Davis */ and got my code back. No GNU LGPL. Copyright stripped.

1:24 PM · Oct 16, 2022Twitter Web App

Replying to

and

ML has memorized matrix transpose without understanding it. I am not surprised about the technology, they shouldn’t have released the product before testing such simple prompts. Copyright and license of ML generated code is unclear at this point. I mean the truly generated ones.

Replying to

and

That still feels like a very specific prompt where you are telling it to give you your own code. I would also want to know if there is other people out there that have used that code before you insist that its spitting out your copywrited code.

Steve Troughton-Smith

@stroughtonsmith

Oct 18

Replying to

@DocSparse

Maybe interesting, maybe not: — OpenAI’s GPT-3 itself spits out this, when queried (no Copilot involved). It has its own body of programming work that may or may not be distinct from whatever Copilot (which is built on top of it) was trained upon

Show replies