Maybe it weights the beginning of new projects as having a new license or copyright notice.
If the model has been updated, it would be interesting to correlate with an uptick in files being created with copyright notices up to that moment.
I remember `npm init` defaults to ISC, so it's surprising to see a named copyright. Maybe author's most-used languages favor repos with proprietary notices.
Wasn't even a new project, just a new file in the project, completely unrelated to the person mentioned in the comment or their projects (checked their github projects to see if there was anything remotely similar). While this is a blatant example since it's a "made by" comment, what happens when actual copyrighted code gets written?
Suppose copyright is a legal attempt at ownership or attribution of something like "I did this work and time is finite; therefore, copyright to me. All rights exercised thereby, etc."
Now let's say for any given length L, there exist programs P1..Pn that are generable from Copilot, given some description(s) D1..Dn.
But not even programs, because those have interfaces and implementations. Wasn't there something between Oracle and Google where copying the API was okay?
So let's say for some string of length L, there exist modules M1..Mn that are generable from Copilot, given some description(s) D1..Dn.
We say those modules are reachable, and modules are divisible so long as an alternate module Ma exists whose outputs are exactly the same, even if some portions of the string are substituted.
But Copilot could have generated those as well. Any alternate modules also exist in the generated space M1..Mn.
Now let's say instead of Copilot, we have a box with listings of modules M1..Mn of any length L. Assume retrieving all modules of a given length is instant.
Let's say we pull out a bunch of these modules and compare them to the copyrighted code. Let's say they are exactly equal.
Copyright has therefore been violated, right?
So we establish even a copyrighted work is an instance of a collection of characters that came from a box with infinite memory.
Then we may need a service that scans Copilot output and adds text like, "We recognize this could be copyrighted, which in this universe represents actions requiring legal compliance; yet our only duty is to inform, not take action."
Because is it not true that engineers must both 1) reduce stakeholder damage and 2) inform, not persuade?
So whether or not a copyrighted work is reused without permission, user assumes all responsibility, even if that law changes ten or a hundred years in the future.
Then it's not necessarily Copilot's requirement to eschew copyrighted output. It merely generates the most probable next token, given some existing collection generated. (Or something like that?)
Computation has increased to cover most, if not all, permutations ever possible.
The artist executes the definition of a `line()` method in a drawing package. The description D is opaque! It later takes the form of interviews, with curators and historians presenting views on some hidden data table.
We can also observe as well: suppose description D contains the exact copyrighted code to generate, and Copilot is asked to echo the output.
Since Copilot "generated" it, should we now accuse Copilot of violating copyright?
(But we assume the operator acted in good faith and did not knowingly include copyrighted code in the input.)
Okay, now let's turn to training. Suppose copyrighted works were used in Copilot's training. If we accept Copilot is an instance of our infinity box, no one owns the algorithm to the English language.
It follows that any string generator modeling probability space is merely pulling values out (in finite time).
Now we may say, "We can't ignore time. Our universe proceeds by time." So let's say copyright is important, by virtue of time.
Then only humans can violate copyright, not software. If the law extends to generated instances, it may overreach, attempting to control the babel box.
If the law defines "publishing to public spaces" a copyrighted work, that's another matter.
I say these things typing into a tiny text area. My source control is copy-paste. Of course, I reserve the right to change my mind, contradict myself, and basically be totally wrong.
There is a GZ on Github who tends to write the following in the header [1], which is different to what copilot generated:
/*
* Copyright (C) 2019 GZ
*/
It is unlikely that copilot copy and pasted a comment block verbatim. In my experience copilot might do something that resembles this if the code is very common across many projects and there is probably only one logical way to implement it. eg. a function to calculate the area of a circle.
Edit: I replaced the person's name with "GZ" because I didn't feel they needed be spammed on search engines because a LLM randomly produced their name.
Do you think copilot's developers would have this attitude if google's LLM spit out their code? How about if it spit out the training weights of GPT-4 into a csv?
I know this doesn’t apply to now concerns. But hasn’t everyone already pretty much vacuumed up our code…?
And in a few years, we’ll kinda be like the guys who optimized assembly code (leading to optimizing compilers), printed BASIC out in printed publications that people had to type in (generations of self-taught hackers), and open-source toolchains (allowed everything else to be built).
Our copyright on a `sqlalchemy` monkey-patch in `heythatsmine/product_model.py` should be about as protected as something Dijkstra scribbled in the margins… if we are to be consistent in our attitudes about the progress of the field.
13 comments
[ 2415 ms ] story [ 212 ms ] threadIf the model has been updated, it would be interesting to correlate with an uptick in files being created with copyright notices up to that moment.
I remember `npm init` defaults to ISC, so it's surprising to see a named copyright. Maybe author's most-used languages favor repos with proprietary notices.
Now let's say for any given length L, there exist programs P1..Pn that are generable from Copilot, given some description(s) D1..Dn.
But not even programs, because those have interfaces and implementations. Wasn't there something between Oracle and Google where copying the API was okay?
So let's say for some string of length L, there exist modules M1..Mn that are generable from Copilot, given some description(s) D1..Dn.
We say those modules are reachable, and modules are divisible so long as an alternate module Ma exists whose outputs are exactly the same, even if some portions of the string are substituted.
But Copilot could have generated those as well. Any alternate modules also exist in the generated space M1..Mn.
Now let's say instead of Copilot, we have a box with listings of modules M1..Mn of any length L. Assume retrieving all modules of a given length is instant.
Let's say we pull out a bunch of these modules and compare them to the copyrighted code. Let's say they are exactly equal.
Copyright has therefore been violated, right?
So we establish even a copyrighted work is an instance of a collection of characters that came from a box with infinite memory.
Then we may need a service that scans Copilot output and adds text like, "We recognize this could be copyrighted, which in this universe represents actions requiring legal compliance; yet our only duty is to inform, not take action."
Because is it not true that engineers must both 1) reduce stakeholder damage and 2) inform, not persuade?
So whether or not a copyrighted work is reused without permission, user assumes all responsibility, even if that law changes ten or a hundred years in the future.
Then it's not necessarily Copilot's requirement to eschew copyrighted output. It merely generates the most probable next token, given some existing collection generated. (Or something like that?)
Computation has increased to cover most, if not all, permutations ever possible.
The artist executes the definition of a `line()` method in a drawing package. The description D is opaque! It later takes the form of interviews, with curators and historians presenting views on some hidden data table.
We can also observe as well: suppose description D contains the exact copyrighted code to generate, and Copilot is asked to echo the output.
Since Copilot "generated" it, should we now accuse Copilot of violating copyright?
(But we assume the operator acted in good faith and did not knowingly include copyrighted code in the input.)
Okay, now let's turn to training. Suppose copyrighted works were used in Copilot's training. If we accept Copilot is an instance of our infinity box, no one owns the algorithm to the English language.
It follows that any string generator modeling probability space is merely pulling values out (in finite time).
Now we may say, "We can't ignore time. Our universe proceeds by time." So let's say copyright is important, by virtue of time.
Then only humans can violate copyright, not software. If the law extends to generated instances, it may overreach, attempting to control the babel box.
If the law defines "publishing to public spaces" a copyrighted work, that's another matter.
I say these things typing into a tiny text area. My source control is copy-paste. Of course, I reserve the right to change my mind, contradict myself, and basically be totally wrong.
I suppose adding your name to every comment will help prevent Github from copying your code without attribution, maybe it's time to start doing that.
/*
* Copyright (C) 2019 GZ
*/
It is unlikely that copilot copy and pasted a comment block verbatim. In my experience copilot might do something that resembles this if the code is very common across many projects and there is probably only one logical way to implement it. eg. a function to calculate the area of a circle.
Edit: I replaced the person's name with "GZ" because I didn't feel they needed be spammed on search engines because a LLM randomly produced their name.
[1] https://github.com/gziolle/promptastic/blob/master/app/src/m...
And in a few years, we’ll kinda be like the guys who optimized assembly code (leading to optimizing compilers), printed BASIC out in printed publications that people had to type in (generations of self-taught hackers), and open-source toolchains (allowed everything else to be built).
Our copyright on a `sqlalchemy` monkey-patch in `heythatsmine/product_model.py` should be about as protected as something Dijkstra scribbled in the margins… if we are to be consistent in our attitudes about the progress of the field.