3 ms·

Try building a few applications which use LLMs. The complexity is high. There is going to be a lot of jobs.

steve_adams_86 3y ago
It will take a lot of trial, error, correction, and brains to glue together the future AI landscape. It isn’t going to glue itself, and novel, interesting uses for new technologies are unlikely to discover themselves. I think it’ll take time, but these jobs will appear, they’ll be interesting and challenging, and there will be a lot of them.
- danielmarkbruce 3y ago
  So much glue required. Complicated glue that only sticks half the time. So many jobs.
  - rokkitmensch 3y ago
    AGI is the fear-meme distracting from the sophisticated shadow programmer union that is going to guarantee full employment until the heat death of the universe by virtue of controlled growth in complexity that must be managed.
    - danielmarkbruce 3y ago
      this
  - steve_adams_86 3y ago
    I tried making my own version of Devin (super rudimentary) when GPT 3.5 was new and holy shit, that was a lot of glue. I’m not a genius but I don’t think there’s a way around it. It’s also not as cheap as people imagine. You need a lot of prompts for so many scenarios, and prompts to verify the outputs of those, like loose abstracted decision trees to create the scaffolding of “thinking” about how to do the job and the workflows themselves. You end up needing to make calls to the AI a tremendous amount compared to when you query it yourself and do work in between responses. If single shot results improve dramatically then some glue and scaffolding might be able to be removed. I think that’s one thing we won’t see particularly fast improvements on in spaces like software engineering, though. Of course despite leaning on compute more than people think, it’s probably still orders of magnitude cheaper than a developer. But what’s the quality like? Devin still only succeeds less than 15% of the time. Reaching even 25% without significantly increasing compute and complexity is probably still quite far off, even with the steady advances we’ve been seeing.

It will take a lot of trial, error, correction, and brains to glue together the future AI landscape. It isn’t going to glue itself, and novel, interesting uses for new technologies are unlikely to discover themselves. I think it’ll take time, but these jobs will appear, they’ll be interesting and challenging, and there will be a lot of them.

danielmarkbruce 3y ago
So much glue required. Complicated glue that only sticks half the time. So many jobs.
- rokkitmensch 3y ago
  AGI is the fear-meme distracting from the sophisticated shadow programmer union that is going to guarantee full employment until the heat death of the universe by virtue of controlled growth in complexity that must be managed.
  - danielmarkbruce 3y ago
    this
- steve_adams_86 3y ago
  I tried making my own version of Devin (super rudimentary) when GPT 3.5 was new and holy shit, that was a lot of glue. I’m not a genius but I don’t think there’s a way around it. It’s also not as cheap as people imagine. You need a lot of prompts for so many scenarios, and prompts to verify the outputs of those, like loose abstracted decision trees to create the scaffolding of “thinking” about how to do the job and the workflows themselves. You end up needing to make calls to the AI a tremendous amount compared to when you query it yourself and do work in between responses. If single shot results improve dramatically then some glue and scaffolding might be able to be removed. I think that’s one thing we won’t see particularly fast improvements on in spaces like software engineering, though. Of course despite leaning on compute more than people think, it’s probably still orders of magnitude cheaper than a developer. But what’s the quality like? Devin still only succeeds less than 15% of the time. Reaching even 25% without significantly increasing compute and complexity is probably still quite far off, even with the steady advances we’ve been seeing.