But for the fact that it would likely get me permabanned, I would suggest making HackerNews Copilot X, the system that responds to HN posts and comments on your behalf and in a manner guaranteed to get more upvotes from other AIbots hanging out in internet comments forums.
GitHub Copilot X – Sign up for technical preview
101–110 of 771 posts
Re: GitHub Copilot X – Sign up for technical preview
#102With increases in productivity from systems like this, and a guaranteed bottom line improvement from cutting labor and running leaner teams, has anyone here heard of any large corporations using LLM's as an excuse for headcount reduction yet? In all honesty, I wouldn't be surprised if this has already played a large behind the scenes role in the layoffs at Google/Microsoft.
I work at a big tech company and there is a complete ban on coding using LLMs. The fear is sending our code for the model to learn in the future. This is also why then hype around GTP4-powered Office is overhyped. Very rarely will a corporation be OK with GPT-4 running inferences on their internal documents and excel sheets Until an on-prem solution is offered, many of these companies won't get anything out of the te…
> Very rarely will a corporation be OK with GPT-4 running inferences on their internal documents and excel sheets.
Which is why Microsoft will be able to charge outrageous fees for enterprise plans that keep employees' queries to GPT private and out of future training sets. I do not think on prem will be successful, though, the hardware requirements are fairly substantial and specialized.
Re: GitHub Copilot X – Sign up for technical preview
#103Earlier quoted context omitted.
I honestly don't think it's going to be a risk for developers. I don't believe you can accurately run attribution on any of the source code generated for you, right? If so, that should give you plausible deniability, while putting all the liability on Github's shoulders. It's their responsibility to make sure they don't violate licenses in their own model.
> It's their responsibility to make sure they don't violate licenses in their own model. Maybe morally that's true. But who do you think is easier to sue? Small startup using copilot or Microsoft?
Yeah true, you're right. Still, intuitively it seems like the better legal case would be against MS, even if it's not the more practical one.
Re: GitHub Copilot X – Sign up for technical preview
#104People should keep in mind that these human in the loop AI systems are built to learn from the human until they’re good enough to replace them.
Do you have evidence to back these claims? Specifically that AI will be replacing software engineers any time soon and that Copilot exists explicitly to gather data towards that purpose.
I’ve been working on ML systems like that for the last 10 years, usually as tools to automate boring data entry tasks like tagging photos, automating food logging, giving dietary advice.
At first these tools offer suggestions, but as more data comes in they start to do parts of the work autonomously, until they can handle all cases.
Stable diffusion will replace a ton of paid design work, self driving cars will at some point replace most drivers.
Re: GitHub Copilot X – Sign up for technical preview
#105What I'd love to see in the future is a proper refactoring UX. This new chat interface is probably a good start, as it seems able to replace code blocks in your original source file. But most code changes in the real world require changes in multiple locations. And I don't usually have questions for the AI, I have small tasks for it to do.
I hope we can get to a point where I can give quick commands to a Copilot-like system to do more complex changes. Changes which traditional refactoring tools balk at. Stuff you would discuss with other humans. Like "this value needs to be calculated once, move the code to the constructor", which then requires code being removed in one method, a member variable being created, and possibly a constructor, and some code being moved there, and perhaps a few changes in method parameter lists.
The jump from single-point edits to multi-point edits is obviously a huge one, not only for the AI part, but also for the UI (being able to understand, accept or modify the changes, for the human). That for me would be truly next level (up from the amazing place where we are already with Copilot), and I can't wait for it to get there.
Re: GitHub Copilot X – Sign up for technical preview
#106I tried getting myself on the CLI waitlist ( https://githubnext.com/projects/copilot-cli/ ). That they use ffmpeg as the example hits a huge pain point for me. But: "Next Waitlist by GitHub Next would like permission to: Act on your behalf Not owned or operated by GitHub" Why does signing up for a waitlist require me to give permissions to an app to act in my account? An app that isn't even officially from GitHub? Th…
> Why does signing up for a waitlist require me to give permissions to an app to act in my account? An app that isn't even officially from GitHub?
Because it uses GitHub's oauth flow, all applications who use this flow show the same stuff.
Re: GitHub Copilot X – Sign up for technical preview
#107Yesterday, Copilot could not write a program with SymPy, which is a library for doing symbolic math in Python. (e.g., it would consistently confuse symbols in equations with Python variables). Today it uses SymPy as well as it uses NumPy (occasional mistakes, but overall it has the right idea). Absolutely astounding.
Yes, I feel like it is much better than HN has given it credit for in earlier discussions. But that may be influenced by the fact that I only started using it relatively recently; I've probably mostly seen it at its best. Also, I have no illusions about how language models work. I notice that the sloppier the code I'm editing, the worse suggestions (and dumber comments) it suggests. If I start writing a comment and c…
You're being too kind on the HN crowd. We're usually extremely pessimistic for no good reason and useless at predicting things. It seems like we just love to hate on stuff.
Re: GitHub Copilot X – Sign up for technical preview
#108Something interesting is that there is no mention of fine-tuning. Is GPT-4 so good at general tasks out of the box that fine-tuning is obsolete? Can one model really be equally good at Python and poetry?
Finetuning has been more or less obsolete for LLMs since the introduction of GPT-3. This was a major discovery of the original paper. Codex-Davinci only existed because the original gpt wasn't trained on much code not because a version trained on sufficient code would need special code optimizations/fine-tuning.
and yes, LLMs are general intelligences. there is nothing narrow about them at all
Re: GitHub Copilot X – Sign up for technical preview
#109Microsoft really seems to have wrapped up the developer ecosystem with VS Code and the Github acquisition combined with OpenAI. They are going to have an absurd amount of data to optimize their models thanks to that, not sure how other AI focused companies can overcome that have to wonder at what point developers remember how anti-developer Microsoft used to be and potentially move away from their ecosystem. Credit t…
Does anyone have a prediction for how this translates into Microsoft's bottom line? I'd imagine it is mostly increased Azure sales with a few Visual Studio licenses.
Bingo, exactly what I'm thinking.
Sure, they get more developers to use their software, but to how much money does that translate considering the tens of billions they have spent?
Not sure how does that helps Azure by the way. It helps GH more than anything.
Re: GitHub Copilot X – Sign up for technical preview
#110Something interesting is that there is no mention of fine-tuning. Is GPT-4 so good at general tasks out of the box that fine-tuning is obsolete? Can one model really be equally good at Python and poetry?
>Is GPT-4 so good at general tasks out of the box that fine-tuning is obsolete? Can one model really be equally good at Python and poetry?
This is correct. It is able to generalize across domains intelligently. Given the results of fine tuning in earlier models, fine tuning GPT-4 for a specific purpose will probably produce super-human ability.