3 ms·

This is the "performance enhancing drugs" of mental work. Either you use it (and you make the Tour de France) or you don't (and you get no next contract). It i

This is the "performance enhancing drugs" of mental work. Either you use it (and you make the Tour de France) or you don't (and you get no next contract). It is easily a 6x performance increase in the amount of code I can write. I either have to use it, or get laid off.

belter 2y ago
Be careful. That 6x performance boost can be easily used with metrics to detect which Developers are using AI...
CraigJPerry 2y ago
6x? I’m a huge fan of Sonnet but I can’t reconcile 6x. It’s a faster keyboard. Stuff like aider architect chat or whatever seems relatively useless. What’s the secret for 6x?
- ahoka 2y ago
  The secret is a low baseline.
johnbellone 2y ago
Do you often go back to maintain that code that you wrote?
- data-ottawa 2y ago
  This is the kicker for me. It’s great at boilerplate like filling in configs, but the results of iteratively coding with AI are first draft quality. Yesterday I used copilot to kick out a really quick REST api for a flask app. A lot of the code was boilerplate parsing args and DB lookups, so it was fine. When I changed one of my db models it broke all the hard coded error handling in frustrating ways. After writing suitable helpers and rewriting the module (with AI) it cleaned up into some good code. Definitely speed up the exploration phase, but required a top to bottom rewrite for the final product. Overall I’d say a little faster than me doing this myself (but I don’t use flask often)
Slartie 2y ago
I'd say it's about 2-3x at most, in the best case scenarios. When I have to write some kind of wrapper or glue code on a green field, I approach that factor. And I really love using AI code completion in those kinds of task. However, writing that kind of code maybe makes up 5% of my work. Analysis, trial-and-error, discussions etc. make up the other 95%, and AI only seldomly helps with that. It can sometimes be useful for research and spec ingestion, but it quickly becomes dangerous in those cases because as soon as you enter any kind of niche area (and unfortunately my work has a lot of those) LLMs tend to hallucinate and present made-up "knowledge" with enviable certainty.

belter 2y ago

Be careful. That 6x performance boost can be easily used with metrics to detect which Developers are using AI...

CraigJPerry 2y ago

6x? I’m a huge fan of Sonnet but I can’t reconcile 6x. It’s a faster keyboard. Stuff like aider architect chat or whatever seems relatively useless. What’s the secret for 6x?

ahoka 2y ago
The secret is a low baseline.

johnbellone 2y ago

Do you often go back to maintain that code that you wrote?

data-ottawa 2y ago
This is the kicker for me. It’s great at boilerplate like filling in configs, but the results of iteratively coding with AI are first draft quality. Yesterday I used copilot to kick out a really quick REST api for a flask app. A lot of the code was boilerplate parsing args and DB lookups, so it was fine. When I changed one of my db models it broke all the hard coded error handling in frustrating ways. After writing suitable helpers and rewriting the module (with AI) it cleaned up into some good code. Definitely speed up the exploration phase, but required a top to bottom rewrite for the final product. Overall I’d say a little faster than me doing this myself (but I don’t use flask often)

Slartie 2y ago

I'd say it's about 2-3x at most, in the best case scenarios. When I have to write some kind of wrapper or glue code on a green field, I approach that factor. And I really love using AI code completion in those kinds of task. However, writing that kind of code maybe makes up 5% of my work. Analysis, trial-and-error, discussions etc. make up the other 95%, and AI only seldomly helps with that. It can sometimes be useful for research and spec ingestion, but it quickly becomes dangerous in those cases because as soon as you enter any kind of niche area (and unfortunately my work has a lot of those) LLMs tend to hallucinate and present made-up "knowledge" with enviable certainty.