64
An earnest question about the AI/LLM hate (infosec.pub)
submitted 1 year ago by to c/technology@lemmy.world
58 comments fedilink hide all child comments
 

Hello, recent Reddit convert here and I'm loving it. You even inspired me to figure out how to fully dump Windows and install LineageOS.

One thing I can't understand is the level of acrimony toward LLMs. I see things like "stochastic parrot", "glorified autocomplete", etc. If you need an example, the comments section for the post on Apple saying LLMs don't reason is a doozy of angry people: https://infosec.pub/post/29574988

While I didn't expect a community of vibecoders, I am genuinely curious about why LLMs strike such an emotional response with this crowd. It's a tool that has gone from interesting (GPT3) to terrifying (Veo 3) in a few years and I am personally concerned about many of the safety/control issues in the future.

So I ask: what is the real reason this is such an emotional topic for you in particular? My personal guess is that the claims about replacing software engineers is the biggest issue, but help me understand.

you are viewing a single comment's thread
view the rest of the comments
[–] 6 points 1 year ago (1 child)

It is really not a big change to the way we work unless you work in a language that has very low expressiveness like Java or Go and we have been able to generate the boilerplate in those automatically for decades.

The main problem is that it is really not at all useful or produces genuinely beneficial results and yet everyone keeps telling us they do but can not point to a single GitHub PR or similar source as an example for a good piece of code created by AI without heavy manual post-processing. It also completely ignores that reading and fixing other people's (or worse, AI's) code is orders of magnitude harder than writing the same code yourself.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •
  • [–] 1 point 1 year ago

    It is really not a big change to the way we work unless you work in a language that has very low expressiveness like Java or Go

    If we include languages like C#, javascript/typescript, python etc then that's a huge portion of the landscape.

    Personally I wouldn't use it to generate entire features as it will generally produce working, but garbage code, but it's useful to get boilerplate stuff done or query why something isn't working as expected. For example, asking it to write tests for a React component, it'll get about 80-90% of it right, with all the imports, mocks etc, you just need to write the actual assertions yourself (which we should be doing anyway).

    I gave Claude a try last week at building some AWS infrastructure in Terraform based off a prompt for a feature set and it was pretty bang on. Obviously it required some tweaks but it saved a tonne of time vs writing it all out manually.

  • permalink
  • fedilink
  • source
  • parent