6 ms·
GitHub accused of varying Copilot output to avoid copyright allegations
- belter 3y agohttps://storage.courtlistener.com/recap/gov.uscourts.cand.403220/gov.uscourts.cand.403220.98.0.pdf https://storage.courtlistener.com/recap/gov.uscourts.cand.40...
- ShamelessC 3y agoEh, their argument is simply that they tuned temperature settings to encourage the model to output slight variations on memorized data. But this is kind of just one of many things you do with a language model and certainly doesn’t imply intent to avoid copyright allegations. Just implies they tuned it for user experience. I was expecting there to be some discovery around them deliberately fine tuning their model to output modifications if and only if the code had a certain license.
- Brian_K_White 3y agoWhy else bother with such an input? Are randomizations more likely to be correct or more useful?
- ianbutler 3y agoPotentially more correct, yes. It frees the model to choose lower probability tokens to some degree, technically it boosts their probabilities, which may be more correct depending on the task. There are also sampling schemes, top_p and top_k which can each individually help choose tokens that are less probable (but still highly probable) but more correct, and they are often used together for the best effect. And then there are various decoding methods like beam search where choosing the most optimal beam may not mean the most optimal individual token. By default a simple greedy search is used which always chooses the next highest probability token.
- slashdev 3y agoI don't know much about AI, but I think one reason you might do that is to learn which variations are preferred (which are committed unmodified) so you can tune the model in the future. I don't know if Github does that, but given they've cited how often code from copilot is committed without modification, I assume they are measuring it at least in some cases.
- Brian_K_White 3y agomakes sense
- brookst 3y agoHuge topic, worth Googling. Short version is that too little randomness limits the solution space, so retrying suboptimal results yields the same problems.
- 2gremlin181 3y agoYe olde Bias-Variance tradeoff
- cubefox 3y agoWell, temperature 0 means the completion is always the most "likely" (or "best", after fine-tuning) token, while temperature 1 means to choose the next tokens stochastically according to their probability (or "goodness" after fine-tuning). Usually some temperature in between is chosen, like 0.7. It's not a priori clear to me which is the best way to do it.
- seanhunter 3y agoGenerally the reason behind adding randomness to machine learning is avoiding "local minima" in the search space of the optimization function(s) used for training the model. If your training produces a very smooth descent to an optimum it can lead to the model converging on a solution that is not globally the best. Adding some randomness helps to avoid this. Specifically for GPT models, the temperature parameter is used to get outputs wihch are a bit more "creative" and less deterministic. https://help.promptitude.io/en/ai-providers/gpt-temperature https://help.promptitude.io/en/ai-providers/gpt-temperature
- golemotron 3y agoYes.
- GuB-42 3y agoIt is worthwhile with creative writing. For example if you ask ChatGPT to write a short story, you want some originality. Even when asking for an explanation it can be useful as you may want to try different things for the explanation that speaks to you the most. But here we are talking about autocompleting code. I don't think programmers want the autocompleter to be creative. They want the exact same solution everyone uses, hopefully the right one, with only minor changes so that it matches their style and use their own variable names. In my case, I am the programmer, I decide what to do, I just want my autocompleter to save me some keystrokes and copy-pasting boilerplate from the web, the more it looks like existing code the better. I have enough work fixing my own bugs, thank you. Speaking about bugs, how come everyone talks about code generation that, I think, doesn't bring that much value. Sure, it saves a few keystrokes and copy-pasting from StackOverflow, but I don't feel like it is the thing programmers spend most of the time doing. Dealing with bugs is. By bugs, there are the big ones that have tickets and can take days to analyze and fix, but also the ones that are just a normal part of writing code, like simple typos that result in compiler errors. I think that machine learning could be of great help here. Just a system that tells me "hey, look here, this is not what I expected to see" would be of great help. Unexpected doesn't mean there is a bug, but it is something worth paying attention to. I know it has been done, but few people seem to talk about it. Or maybe a classifier trained on bug fix commits. If a piece of code looks like code that has been changed in a bug fix commit, there is a good chance it is also a bug. Have it integrated to the IDE, highlight the suspicious part as I type, just as modern IDEs highlight compilation errors in real time.
- kevingadd 3y agoWhat's the value of slight variations? Isn't it more likely that the memorized data was already known to be good and effective? It doesn't seem like a useful change unless your goal is to avoid infringement. I don't see how randomly permuting the suggestions improves UX.
- moyix 3y agoThe lowest temperature isn't always the one that results in working code! This was shown in the original Codex paper: > When evaluating pass@k, it is important to optimize sampling temperature for the particular value of k. In Figure 5, we plot pass@k against the number of samples k and the sampling temperature. We find that higher temperatures are optimal for larger k, because the resulting set of samples has higher diversity, and the metric rewards only whether the model generates any correct solution. > In particular, for a 679M parameter model, the optimal temperature for pass@1 is T∗ = 0.2 and the optimal temperature for pass@100 is T∗ = 0.8. With these temperatures, we find that pass@1 and pass@100 scale smoothly as a function of model size (Figure 6). So even with pass@1 (likelihood of getting the right answer in 1 attempt) you don't use T=0, so there will be slight variations in the output each time.
- brookst 3y ago[flagged]
- rolph 3y ago[The judge overseeing the case has permitted the plaintiffs to remain anonymous in court filings because of credible threats of violence [PDF] directed at their attorney. The Register understands that the plaintiffs are known to the defendants.] https://storage.courtlistener.com/recap/gov.uscourts.cand.403220/gov.uscourts.cand.403220.71.0.pdf https://storage.courtlistener.com/recap/gov.uscourts.cand.40...
- formerly_proven 3y ago> go f**g cry about github you f**g piece of s*t n**r, I hope your throat gets cut open and every single family member of you is burnt to death Are github users gamers? Really puts the "git" into "github" there.
- zer0tonin 3y agoNo, but I assume a lot of AI bros are
- obiefernandez 3y agoffs, can we just not make this a thing?
- edgyquant 3y agoToo late, a demographic or people who could just barely scrape together a script making REST requests are now selling themselves as “AI specialists” or “prompt engineers” to the corporate class. These are this cycles cryptobros, who were mostly not engineers but people riding a hype wave. The age of the AI bro is here, and as I’ve been in the space as someone genuinely interested in the models, working with them from time to time, for a while. I’m giving a lot of eye rolls in meetings when these people start talking about the underlying tech.
- foobarbazetc 3y ago“Prompt engineers”… sigh.
- taneq 3y agoIsn’t “rewrite the example code in your own style” accepted best practice for human coders, when working from an example that does what you need? I’m not sure what would be acceptable output for a code generation tool if rewriting the examples isn’t ok and reimplementing something that performs the same function still isn’t ok. Are we automatically granting de-facto code patents on all published code now?
- jazzyjackson 3y ago"Isn’t “rewrite the example code in your own style” accepted best practice [...]?" Why would it be? If a function performs the data transform I need you better believe i'm copy pasting that sucker with a hyperlink to where I found it But then again, I'm not trying to win in court.
- l__l 3y agoThe point here is that this isn't some example from a textbook or even stack overflow, but licensed pieces of work with all the legal complications that come with that. This is about the potential use of this code in proprietary code (or code otherwise incompatible with the original licenses), and I really don't think anyone would say it is "accepted best practice" to copy out someone else's work you find online, licenses be damned, in a professional setting.
- 542458 3y ago> this isn't some example from a textbook or even stack overflow, but licensed pieces of work with all the legal complications that come with that I understand why these might feel different to you, but textbooks and stack overflow are also proprietary, licensed pieces of work. I don’t see why there would be much of a legal distinction.
- cmrdporcupine 3y agoCopilot is to license violations (esp of copyleft licenses) what cryptocurrency mixers are for money laundering. My employer (IMHO smartly) forbids use of LLMs in company IP and company laptops, etc. Many others I'm sure are doing the same, and many others will follow.
- bushbaba 3y agoOnce the ip rules are figured out it’ll open the door to a lot of usecases. This reminds me more of p2p file sharing being precursor to paid streaming services.
- theRealMe 3y agoNobody uses copilot intentionally to violate copyright law. People do use crypto mixers intentionally to violate money laundering laws.
- SpicyLemonZest 3y agoNobody affirmatively says “yes, my goal is to violate copyright law, and Copilot is the best tool I’ve found”. But it doesn’t seem impossible to me that the value of Copilot comes partially from the fact that it can copy paste code from copyrighted repositories in ways which would be illegal for you or I to do. I’m not sure it’s proven yet but I wouldn’t be shocked if it is in the future.
- shagie 3y agoIt provides the same value as someone who copies and pastes code from Stack Overflow or any of the predecessors without concerning themselves with the license. I am certain that I can find code from Linux or gcc or emacs on Stack Overflow that is under a GPL license and not compatible with the CC license Stack Overflow uses... and yet it's there. What's more, people will copy that code into their own ignoring the CC license too. How is that really any different than using Copilot if the original license and attribution are something to respect. Note that I do think that the original license is something to respect which is why for any of the code that I write that has copyright that matters on it (toy program for home? meh. Hobby project repo that I'm working on that I'll publish? yep. Employer's code for work? absolutely.) I either don't touch questionable sources or run a license check on it when using it. The key thing is that I don't consider the use of Copilot to be any more controversial than copying from Stack Overflow - which has been done by countless programmers for a decade before Copilot existed and no one got up in arms about it then.
- williamcotton 3y ago[flagged]
- matkoniecz 3y ago> Downvoting Presumably people downvoted it because it is really unclear what exactly you are claiming. Instead of "Everyone needs to first familiarize themselves with" you could write a very simple summary of that and how it relates to this case and your next claim that > If you’re under the impression that every line of code is covered by copyright you are very mistaken. Well, for example empty ones are really unlikely to be. Ones that quote out-of copyright works also will not be.
- williamcotton 3y ago[flagged]
- catiopatio 3y agoThe downvotes probably have to do with the fact that: (1) you lead with a rude and mostly contentless comment, and (2) your follow-up is merely a dump of Wikipedia quotes, instead of actually summarizing what you’ve been trying to say.
- JVillella 3y agoThey say crypto is “regulatory arbitrage”, I say this AI co-pilot stuff is “copyright arbitrage”. Being a bit hand-wavy with it: It’s akin to torrenting music/movies. The torrented files are lossy compressed representations of the original waveform from the music producer. Limewire, or Pirate Bay, or whatever provide interface to retrieve them (download or stream). The model weights are a form of lossy compression, and inference is like a document retrieval. One may say, “it’s like an employee working at company X, then going to work at company Y, they retain their knowledge and experience.” I would say it’s more like, employee going from X to Y, but retaining audio and video recordings of all interactions he had, notes, documents, and other proprietary info and bringing it to company Y.
- mjburgess 3y agoI call it copyright laundering
- soultrees 3y agoWhat would you say the basis of all knowledge you know is? You are a collection of everything you have consumed and the stuff you create is all influenced by that. Personally this whole llm debate about copyright is quite funny. As someone who very much has skin in the game(my art is trained on midjourney.), and who runs in a circle of artists, it’s interesting to see people’s ego’s come at play here. The ones who are excited about these as tools are the ones who are openly inspired and want to inspire however the ones who claim copyright infringement seem to come off as insecure, almost like they are afraid that this idea of theirs will be the last great idea they have. There’s already a separation happening in the art world of people who are exploding in creative output vs the people who are so defensive and cling to the old way of doing things. If I had my way, I’d see copyright laws abolished completely. A complete free for all in innovation. And people who claim that without parents and copyright then there’s no incentive to make money seriously underestimate humans and their ego to continually innovate.
- zzzzzzzza 3y agosome things like drug discovery could probably be done with a bounty system rather than intellectual property, and could probably get much better results for a fraction of the cost for maintaining the intellectual property component of the court system
- jimnotgym 3y agoTaking code off github, changing it a bit and passing it off as ones own crosses a line. Now we really can't tell the AI from the humans!
- unkulunkulu 3y agooh come on, which code? writing imports? or iterating over collections? or am I underusing copilot? :) I basically use it as stackoverflow on steroids. it is not even close to gpt-4 in terms of reproducing some original idea I could not find in a search engine
- missingdays 3y agoWhy would you ever write imports? IDEs autocomplete them for you
- unkulunkulu 3y agoCopilot understands some convetions when there’s more than one way. I used it extensively with react bootstrap where I decided to go with the recommended way of importing each components like import Tab from ‘react-bootstrap/Tab’ It also knows which components are used in the file.
- sureglymop 3y agoBut that's pretty much what copilot is... It's just Intellisense 2.0 and I would say even only marginally more useful. You can't even really instruct it except with some comments which may not work.
- beezlewax 3y agoCopilot really messes up imports for me sometimes. The ide does a better job via autocomplete when it is turned off
- rkagerer 3y agoThe plaintiffs were granted anonymity due to credible threats against their attorney. Is there any mechanism other than publication ban that ensures the protection? Can't someone just attend the day of the hearing to see who the attorneys are? EDIT: Apparently the lawyers are attending via Zoom.
- coryrc 3y agoThe plaintiffs, not the plaintiffs' lawyers.
- WalterBright 3y agoOne of the specific complaints is: https://devclass.com/2022/10/17/github-copilot-under-fire-as-dev-claims-it-emits-large-chunks-of-my-copyrighted-code/ https://devclass.com/2022/10/17/github-copilot-under-fire-as... It's a 25 or so line function that looks like a pedestrian implementation of a sparse matrix transpose algorithm. The author should have been patented it to protected it, not copyrighted it.
- blibble 3y agoso not only is it a shitty boilerplate generator, now it also introduces deliberate random changes (i.e. bugs)