What is Microsoft's long-term goal here? Why did they release a hugely controversial feature that is not making them any money? (Or is it?)
All public GitHub code was used in training Copilot
591–600 of 734 posts
Re: All public GitHub code was used in training Copilot
#592Earlier quoted context omitted.
There is nothing intelligent about this. What they did is a context aware search and trying to claim that not what this is. If it was just used as a search engine and people weren’t using the results or following the license of the original source, then it would fine. There has been so much of a hype of machine learning people likely have a false impression of what it is.
I've seen this claim that Copilot is "just a search engine" repeated in multiple places now. It's wrong; as anyone familiar with any of the GPT variants or other similar autoregressive language models can attest. Copilot isn't a search engine any more than any other language model is. It can sometimes output data from the training set verbatim as most AI models do from time to time, but that is the exception not the…
First off I said it was a context aware search which it is. It uses past training data to predict what you would type next based on the context ie the code around it. It’s no more intelligent than alpha go. Intelligent AI is considered to be a general ai which no one is even close to building yet.
Since neural networks are pattern matching based on the training input it is a derivative work of the training set. It says it right in first thing that comes up in auto regressive language models use the training input plus context to predict what the next word would be.
Now here where the fun begins if they try this in court. If you claim it’s generating new work then who owns the copyright? You may not realize how big of a deal this is but there was a court case you can lookup where a monkey took a selfy and the person who camera the monkey used tried to claim copyright and lost.
Re: All public GitHub code was used in training Copilot
#593To me, the particular use case and whether it is fair use or not, is of minor interest. A far more pressing matter is at hand: AI centralization and monopolization. Take Google as an example, running Google Photos for free for several years. And now that this has sucked in a trillion photos, the AI job is done, and they likely have the best image recognition AI in existence. Which is of course still peanuts compared…
For years we thought Google Translate was the best machine translation we would ever get. Then DeepL just popped up out of nowhere and today other services still didn't manage to catch up.
Every now and then you get someone to think about an old problem on a clean sheet of paper and you might get a better result with less training data / investment.
Re: All public GitHub code was used in training Copilot
#594Earlier quoted context omitted.
It's not a violation of copyright to train a model. There are three questions at play though: 1) Can you be liable for violating copyright if you have never seen the work? 2) Can a non-human be held accountable for violating copyright? 3) Can github be held liable for an end user using their tool to violate copyright? https://en.wikipedia.org/wiki/Substantial_similarity wikipedia states: Generally, copying cannot be…
1) So you are saying if I get a disk duplication machine I can freely copy and distribute blu ray disks as long as I don't watch the movie on the disk? 2,3) Seems pretty settled at this point, look at the cases around the VCR and copy machine. In general the one using the machine is liable. The creator of the machine can be held liable if there aren't substantial non infringing uses.
1) No. But you can freely distribute the disk duplication machine.
2) Someone using a copy machine is knowingly copying a specific work.
Re: All public GitHub code was used in training Copilot
#595To me, the particular use case and whether it is fair use or not, is of minor interest. A far more pressing matter is at hand: AI centralization and monopolization. Take Google as an example, running Google Photos for free for several years. And now that this has sucked in a trillion photos, the AI job is done, and they likely have the best image recognition AI in existence. Which is of course still peanuts compared…
I believe this is actually powered by OpenAI, which while large (now), is nowhere near the behemoth that Microsoft or Google is.
This suggests that seeing the future a bit ahead of the rest of the world, and then assembling a motivated all-star team is (perhaps in the short term at least) one way of out-competing the "super AI" of the giants.
Re: All public GitHub code was used in training Copilot
#596Earlier quoted context omitted.
It's not a violation of copyright to train a model. There are three questions at play though: 1) Can you be liable for violating copyright if you have never seen the work? 2) Can a non-human be held accountable for violating copyright? 3) Can github be held liable for an end user using their tool to violate copyright? https://en.wikipedia.org/wiki/Substantial_similarity wikipedia states: Generally, copying cannot be…
> It's not a violation of copyright to train a model. This is a very bold assumption, one that I assume will not hold in the court of law in all cases. I think the nuanced question is: to train a model that does what, exactly . Let's say distributing meth recipes is illegal[1], can one legally side-step that by training a model that spits out the meth recipe instead? No court will bother with the distinction, causati…
It's not illegal to train a model to spit out classified nuclear weapon schematics. Possessing the original data might be. Releasing software that does this might be illegal, but not for copyright reasons, which is the issue at hand.
Re: All public GitHub code was used in training Copilot
#597Earlier quoted context omitted.
Both of these examples include you creating something and then using it. I don't know how copilot works, but using the second example, if you wrote a script to download and seed trackers, and someone else used it, I don't think you would be held under any liability, especially if you don't profit off of it. Not a lawyer or even particularly well informed edit: I am reminded of the monkey selfie, in which it was ruled…
It sounds like you're arguing that Github isn't liable for people using copyrighted code through Copilot. I think most people are more concerned about whether the user of Copilot would be liable for using copyrighted code generated by Copilot.
Could be. But I could also see the courts ruling an individual can't be liable for copyright violations if they never accessed the original work, which is generally required.
Re: All public GitHub code was used in training Copilot
#598Earlier quoted context omitted.
I have about 300,000 photos that haven't been scanned by AI (unless someone at Backblaze did it without permission). I'm sure there are lots of other photographers out there who miss Picassa, which Google killed off to push everyone's data to their service. (It did really well in matching faces, even across age, but the last version has a bug when there are multiple faces in a picture, sometimes it swaps the labels)…
I am not an expert on the field. But my hope was that this could be facilitated by Transfer Learning. Still don’t know how the scale economies could be achieved. Maybe just out of the sweat and network of passionate people like in the case of open source.
I work in the field. Transfer learning helps get you decent/good models, but the best models remain ones trained on large amounts of data. You may be able to get away with good performance and not great on your task. For some areas you really care a lot about long tail performance (like self driving) that you will need massive dataset. For other areas if your goal is to be the best relative to other large companies you will need a massive dataset.
Transfer learning best use cases are for fast prototypes or for ml tasks that do not need state of the art performance.
Re: All public GitHub code was used in training Copilot
#599Earlier quoted context omitted.
It's stronger than that: if GitHub is correct that they don't need a license then they are allowed to train on publicly visible code even if it is labeled with "no one has any provision to use this for anything at all, especially training models"
Which is why I think this could be a big turning point. IMO, GitHub is breaking licenses. If an ML algorithm ingests a viral licensed block of code, its outputs should be tainted with that license as it's a derived work. Otherwise I can make a program reproduce whole repositories license free, so long as I can claim "well, the AI did it, not me!" It's produced something based on the original work, therefore it should…
This isn't as clear as most things we work on as engineers, but there's a spectrum:
* An algorithm (or person) ingesting lots of code and then later spitting out that same input, does not free anyone from the copyrights of the input.
* An algorithm (or person) that ingests lots of code, finds commonalities, synthesizes that into something new, and produces something well beyond mere copying is producing something new, likely without any legal tie to the original.
Right now, it looks like most of what co-pilot does is closer to the latter, but sometimes it does some things that are closer to the former? I can't see any reason why they wouldn't be able to fix it to avoid regurgitating its input, however, with something like a bloom filter, so I expect a long-term there's a way to do it that falls entirely within fair use?
Re: All public GitHub code was used in training Copilot
#600To me, the particular use case and whether it is fair use or not, is of minor interest. A far more pressing matter is at hand: AI centralization and monopolization. Take Google as an example, running Google Photos for free for several years. And now that this has sucked in a trillion photos, the AI job is done, and they likely have the best image recognition AI in existence. Which is of course still peanuts compared…
I have about 300,000 photos that haven't been scanned by AI (unless someone at Backblaze did it without permission). I'm sure there are lots of other photographers out there who miss Picassa, which Google killed off to push everyone's data to their service. (It did really well in matching faces, even across age, but the last version has a bug when there are multiple faces in a picture, sometimes it swaps the labels)…
Photoprism, digiKam, shotwell all have image recognition features, with varied levels of sophistication.