Actions
Forum Thread Details
  • Replies 18 replies
  • Subscribers 475 subscribers
  • Views 606 views
  • Users 0 members are here
Related

Is fancy code completion really worth it ?

michaelkellett

Of credible AI horror stories I think this is about as bad as it gets !

https://www.tomshardware.com/tech-industry/cyber-security/researchers-uncover-critical-ai-ide-flaws-exposing-developers-to-data-theft-and-rce

I'm sticking with the "legacy" Keil IDE for C coding on ARM - it may be a little slower but at least I'm fairly sure that nothing gets into my code that I didn't put there !

MK

  • AI is still at the scary stage where it cannot be trusted, but you absolutely need to figure out how to make it work in order to stay competitive from a productivity point of view.

    There will be many companies that forbid the use of AI because the risks are too high.

    There will be companies that use AI and run into costly or disastrous situations.

    There will be companies that successfully figure out how to leverage AI to improve productivity and corporate performance by at least an order of magnitude.

    There is no question which companies will win in the end.

  • in reply to dougw

    You forgot about the companies that will exploit AI to make their competitors look bad. "It's not enough to win, you have to make everyone else look foolish for trying..."

  • in reply to bradfordmiller

    Fake news seems to be a popular application for AI. Perps don't even care if the AI is hallucinating.

  • AI is great. Just always use the local models, they are more than powerful enough. Do not forget to use tools - give LLMs an access to LSP (clangd is ok), to a sandbox with simulation and tests, etc.

  • in reply to DAB

    There is one way AI can earn way more trust than humans (your local ones or third party somebody) - AI is great at writing proofs. If you generate a certified code with AI, it outperforms humans already. Writing proofs is way too tedious for humans, but AI have infinite patience. I generated fully formally proven SystemVerilog modules with AI, fully certified code in SPARK and Coq. And I only ever use local models for this.

  • in reply to metaprog

    Unless you can check each path, you still never know what the AI is doing or what issues generate the answers.

    Risk Analysis 101, if you cannot verify the results, you do not trust them. Ever!

  • in reply to colporteur

    I use vi too : )

    (Admittedly only for short pieces of code or edits).

    I can't imagine explaining to a non-tech person how vi is like MS Word, but no buttons to click on, and no icons, i.e. not even Teletext level. A bit like trying to explain that the Internet was text-only for some people too, until Netscape Navigator came along... and pictures were often created in text characters, because it took too long to transfer images, and there was nowhere to store them for long, apart from floppy disks (or print on fanfold paper!). My e-mail account storage limit was 150 kbytes at uni.. (that was for all emails combined : )

    Nowadays there's AI for vi (I've not tried it).

    I don't know if you've ever watched Halt and Catch Fire, but perhaps you might like it. It's fictional, but many elements are close to reality. I really enjoyed it.

  • in reply to DAB

    It's no different from human-written code though. And yes, you absolutely can check every path in the generated code, this is what formal verification is for.