328
Amazon cloud boss echoes NVIDIA CEO on coding being dead in the water: "If you go forward 24 months from now, it's possible that most developers are not coding" (www.windowscentral.com)
submitted 2 years ago by to c/technology@lemmy.world
217 comments fedilink hide all child comments
you are viewing a single comment's thread
view the rest of the comments
[–] 24 points 2 years ago* (last edited 2 years ago) (4 children)

I'm going to call BS on that unless they are hiding some new models with huge context windows...

For anything that's not boilerplate, you have to type more as a prompt to the AI than just writing it yourself.

Also, if you have a behaviour/variable that is similar to something common, it will stubbornly refuse to do what you want.

  • permalink
  • fedilink
  • source
  • hideshow 4 child comments
    •
  • [–] 13 points 2 years ago (3 children)

    Have you ever attempted to fill up one of those monster context windows up with useful context and then let the model try to do some useful task with all the information in it?

    I have. Sometimes it works, but often it’s not pretty. Context window size is the new MHz, in terms of misleading performance measurements.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 3 child comments
    •
  • [–] [S] 7 points 2 years ago

    I find there comes a point where, even with a lot of context, the AI just hasn't been trained to solve the problem. At that point it will cycle you round and round the same few wrong answers until you give up and work it out yourself.

  • permalink
  • fedilink
  • source
  • parent
  • [–] 1 point 2 years ago

    To actually answer your question - yes, but the only times I actually find it useful is for tests, for everything else it's usually iffy and takes longer.

    Intelligently loading the window could be the next useful trick

  • permalink
  • fedilink
  • source
  • parent
  • [–] 1 point 2 years ago

    I think that giving the LLM an API to access additional context and then making it more of an agent style process will give the most improvement.

    Let it request the interface for the class your using, let it request the code for that extension method you call. I think that would solve a lot, but I still see a LOT of instances where it calls wrong class/method names randomly.

    This would also require a lot more in depth (and language specific!) IDE integration though, so I forsee a lot of price hikes for IDEs in the near future!

  • permalink
  • fedilink
  • source
  • parent