I think one should develop tools that help you with using APIs. Autocompletion is one nlp technique we all like. If the IDE got to a point where you add a nb platform to it and it made suggestion on semantic matches rather than syntax, many would use it. I think this could help to speed up familiarizing with an api a lot! Inference can be done offline. So perhaps a tool that downloads a model from huggingface and then has an ‘automated’ fine tuning that happens locally would raise less of a security concern. Yes training is computationally costly. But so was auto-completion when it got introduced. Stefan Am 11.02.2024 um 02:18 schrieb Owen Thomas <owen.paul.tho...@gmail.com>:
On Sun, 2024-02-11 at 09:50 +1100, Owen Thomas wrote: On Sat, 10 Feb 2024 at 13:46, Andreas Reichel < andr...@manticore-projects.com> wrote: Smart people will become smarter and faster using it. Others won't.
What do you mean when you say this?
What I meant was: it amplifies experience and skills. If you are experienced it will make you faster because you know when to trust and use it. If you are inexperienced it will harm you because it can pretend providing "solutions", which actually may be harmful or wrong.
Yes, I understand that perhaps it does help one who is already experienced leverage off the experience of others to quickly come to an optimal solution to a particular problem.
Perhaps I would agree to using an AI that makes suggestions without "learning" from the code it is being exposed to.
I think I would surely benefit from some assistance too, but I am afraid that an AI algorithm may suggest what it found in my code to a wider developer community, thus leaking my own work to a wider world out of my control.
That is a different and very valid and interesting angle to look at it: what guarantees are there when its integrated into your UI, which opens ALL your code? None and nothing! They just will say "sorry" when eventually caught and that's it.
I would recommend Netbeans think about this scenario. I believe that the point I make in this reply above about preventing the AI from being trained on the code it is exposed to in a particular user's IDE would be considered valuable to many.
I would say, when you believed in proprietary code then IDE AI integration was not for you even when it worked.
I'm not averse to publishing my code and being attributed for it - that's why I do this. Hence, I would be flattered if I saw snippets of my published code being suggested back to me by an AI, but I wouldn't be as happy if I saw unpublished work appearing. I will steer clear of an AI looking at the code in my IDE for now and wait a while perhaps until this AI stuff matures; let others make the mistakes that I hopefully avoid. :)
Although the same concern applies to GitHub already. So unless you host by yourself strictly (as we do, except for what we publish as OpenSource) this concern should not be new.
Just as well because I have no code in GitHub.
Owen. Clique Space(TM). Anima ex machina.
|
