It is truly, deeply amazing how bad Microsoft is. Proton on Linux is FASTER than the actual directX it’s emulating is on windows. They got beat at their own instruction layer.
I believe proton isn’t emulating but rather is a direct instruction layer. I’m ootl for close to a decade plus, but the commonly cited number was you need hardware 7x the power of original equipment to properly emulate.
Just like Skype getting lapped by Zoom during COVID
you literally have access to all the code in the world
I’d like to believe that they were honorable enough to not secretly train on code without people’s permission. But realistically they totally did exactly that, but just made the AI Model this incompetent through some other engineering blunder.
Also, random side thought - training only on public repos probably yields you way higher code quality as opposed to training on both public and private repos? I assume we all have some very messy private repos that we’re too embarrassed to publish because the code quality is absolute shit … right?
honorable enough
Lol. Lmao
They didn’t check licenses in any way, as it did reproduce the famous quake fast inverse square root function, comments included. And quake, like majority of github projects, is published under GPL, which requires all copies and modifications to be published under GPL as well, after which all sane enterprises have banned copilot usage.
Though, we’re not living in sane times anymore. Chatgpt, gemini, deepseek, claude, all reproduce copylefted code left and right. Realistically, Stallman should’ve been rolling in cash by now…
This brings up an interesting question with AI
If I, as a human, read a piece of Open source code it solves a problem in a unique and new way, and then I myself write my own closed source code that solves the problem in the same way, I have not violated a license. The license is for the code itself, not a patent for the specific way the code solves the problem. And since the code in the closed source product is written by me and not copy pasted from the open source project, I have not violated the license per
So what about AI? If you train the AI on a piece of code, and it outputs the same or similar code, do we treat that as if the human copy pasted the code? Or do we treat it as if the human used what they learned from the first program and wrote something similar?
There is already an AI company taking advantage of this. They advertise that that if you want to use open source code in a closed source product, you hire them-- their AI will parse through the open source code and spit out a list of specifications that is specifically not code. Another AI on a completely different system that has never had access to the open source code will then take that specification and spit out program code that is functionally identical and does the exact same thing but is a completely new creation. The result is that you essentially rewrite the open source code but without the copyleft restrictions.
This is going to be an issue that laws and courts will have to address. Especially if, in your example, the code produced by the AI was actually identical to the GPL quake code. Because while a human copying the functionality is never going to write the exact same code line by line, the machine might be.
There is a bit of underlying problem, though. Case in point: I was recently asked to make a js function that converts any string into a hex color, which I promptly just copied off stackoverflow, but noticed they now intercept Ctrl-C to add CC-BY-SA banner. I’d usually include a link whenever I copy something off somewhere anyway, if not for licensing sake then at least for ease of future code navigation. But this got me curious, I asked ChatGPT the same question and it provided the exact same snippet, but without any attribution. And so did gemini and qwen 3.5. You can tell its a copy because the original uses a very specific bit-shift by 5 to mix numbers up a bit, but it doesnt have to be exactly 5, or even shifted at all for that matter. Which got me even more curious, so I went to github and searched for “string to hex” and found the exact same snippet in projects licensed under a variety of licenses, which I’m pretty sure are all just careless copies of that stackoverflow snippet, though it’s not improbable that one of them is actually the one that got copied into that anwser in the first place. Now, for that particular one, I doubt that it reaches threshold of originality to be held in court, but it highlights an issue that makes this double AI conversion trick you described dubious and not-so-bulletproof as second would probably give the same result anyway, just because it almost definitely had to be trained on the copyleft infringing copies… Unless they actually clean-labbed an enormous dataset for it themselves.
Oh, and just as a final experiment, I did try to actually code it up myself. Despite already seeing an implementation, and without that much room for variety, without even trying, I made something that bears no resemblance to that snippet whatsoever, and most likely would’ve ended with the same code without ever seeing that snippet the first place. And from experience interviewing developers and live-coding with them, no two people write the same code, even for simplest of tasks. So… yeah, nah, I don’t buy the “AI does the same thing humans do” any more than the “forklift does the same thing as humans do”. I hope courts do as well, or, at least, this AI craze finally leads to copyright abolishment, it doesn’t really make any sense for both to exist…
I’d like to see the end of software patents, which IMHO are a much bigger problem than copyright.
You make a good point with the exact text bit though. It would certainly put the AI company on the defensive if the author of that original GPL snippet decided to go after openai etc for copyright infringement and license violation. I’d actually kinda like to see that happen.
I’m always so extremely confused about the trope of the personal project having shit quality… Like, if I’m doing something for myself, that’s exactly the place where I wanna do something amazing, like literally all my private projects have much higher quality than my work ones - because in the work ones I’m forced to use stupid conventions, old tools, am not supposed to touch “legacy” code, etc etc etc
As such, since companies have their private code on GitHub, that’s where I would expect the shittiness to come from, not personal private projects.
Like, if I’m doing something for myself, that’s exactly the place where I wanna do something amazing,
That’s always my intention with my personal projects too! But that always results in “Wow I just learned how to do this thing much better, let me refactor the whole project to do it perfectly everywhere” followed by my Adderall running out. So there’s just so many half-done refactors I either forget about or abandon because I get a new idea the next day, but that’s totally just a skill issue.
You’re right though, the code I write at work is much worse, but my Company hosts their own GitLab instance so the code we write can’t even be used to poison Copilot :(
Maybe they meant abandoned projects that never quite got through the todo list but you’re right. Even my abandoned projects are generally better than the legacy I’ve seen lol
I’d like to believe that they were honorable enough
In the future, will you still believe Microsoft (or any other big tech company) has honor not to do stupid shit?
Azure (and all of cloud compute) is the Extend phase of open source. I’m just waiting to see what the extinguish one will be.
I see this as proof of how bad LLMs actually are. You have an AI trained on essentially humanity’s collective programming library. Languages of machines and computers. The result should be ungodly and near perfection. If there was any semblance of understanding in AI, it should be revealed in it’s capability to produce code.
Although… I can definitely see Microsoft thinking that their code is the example of perfection and training copilot on that rather than github.
should be ungodly and near perfection
Counterpoint: garbage in, garbage out
Your proof of how bad LLMs are is the fact that there are a bunch of other companies producing way better coding agents and coding models than Microsoft is? I’m not sure how that follows. Those other agents are good, that’s the point of this.
I would (probably not literally) wager that by “this” they meant *looks around at entire world*…
this
It’s wild to me that people are constantly complaining how “bad” LLMs are, because what, it can make mistakes or it’s not orders of magnitude smarter than the smartest humans???
Do they still do Hungarian warts?
It is impressive how much better claude is than copilot specifically for coding.
Like… how much bullshit is in Windows to learn off of, let alone github.
If agentic LLMs are the prize, I’d be fine with losing.
I mean, Microsoft’s biggest mistake was shoving it in front of everyone’s faces. The real reason that all the other “agentic BS” is received well is because the people who use them have an actual use case, or, are very enthusiastic about the technology and enjoy messing with it. Thus the discussion is mostly from that small group of people who will have something positive to say.
The truth is, that all the models and harnesses suck for most use cases that most people have. When you shove it in front of a general audience and make them interact with it, then the discussion will be about how bad it is.
Greentext on X?
Nah, green is woke.
I was using tabnine before copilot was even announced… they are the ones who fumbled this more than github.
Fucking linux nerds (that was for someone I just met on here). Anyway, yea, my company got three licenses for it, biggest waste of budget ever. I dunno what they clamped on top of open AI, but it’s just terrible. It can’t even parse a spreadsheet correctly.
The fact that it can’t even get PowerShell right when it should be intimately familiar with PowerShell commands and scripts is absolutely asinine.
deleted by creator
What about Cleverbot? Or Tay?
