pull down to refresh

1 sat \ 2 replies \ @k00b 19 May \ on: Stacker Saloon

The LLM work stages I find myself doing on my current PR:

  1. tons of iterations, variants, and feedback to get it working, end to end, how I want
  2. many iterations attempting to remove bloat from the resulting code
  3. a few iterations reorganizing and renaming the code to make it easier to review

I'm currently on stage (3) but I imagine stage (4) is, upon human review, doing a more directed stage (2). Then, a more directed stage (3).

write
compose
124 sats \ 1 reply \ @SimpleStacker 19 May

Do you find that it's actually making you produce faster?

reply
1 sat \ 0 replies \ @k00b 19 May

The biggest change is that I've grown more ambitious per unit of time. I'm producing better things rather than producing more of the same things faster.

Before, I reduced my wants to what I could generate in a reasonable amount of time. Now, generating code is not the bottleneck. Fleshing out the vision, review, and making the changes cohesive with the whole are the bottlenecks.

I probably use LLMs differently than other programmers though. I think it's because I view

  1. code and features as liabilities
  2. simplicity and quality as assets

I think the right use of LLMs is to turn a pocket knife into a razor edged katana. And the wrong use is to turn a pocket knife into a swiss army knife.

reply