Top Tips for Writing Code with AI

Solid, particularly the advice about context. I find with AI, less is better. Once you have "enough" context, adding more increases the risk of problems. The one I take exception with is the last. "You don't need to understand what the AI wrote". I feel you absolutely do need to understand what the AI wrote and if you don't you should not commit it.

Re: Top Tips for Writing Code with AI

#5

Post by bsdpython » Thu, May 15, 2025, 3:07 PM UTC

uberman wrote: post #4

Solid, particularly the advice about context. I find with AI, less is better. Once you have "enough" context, adding more increases the risk of problems. The one I take exception with is the last. "You don't need to understand what the AI wrote". I feel you absolutely do need to understand what the AI wrote and if you don't you should not commit it.

Thanks. I know the last one is controversial, but the way I am starting to think about it is that we are just moving to a new layer of abstraction. I no longer understand very well how hardware works, nor do I know in detail how a browser renders a page, nor the full fine details of how many of the libraries I use work. My own AI generated code, in pockets, is starting to work in the same way. And I'm starting to become OK with that risk.

Re: Top Tips for Writing Code with AI

#6

Post by sherdil2022 » Thu, May 15, 2025, 3:25 PM UTC

bsdpython wrote: post #5

uberman wrote: post #4

Solid, particularly the advice about context. I find with AI, less is better. Once you have "enough" context, adding more increases the risk of problems. The one I take exception with is the last. "You don't need to understand what the AI wrote". I feel you absolutely do need to understand what the AI wrote and if you don't you should not commit it.

Thanks. I know the last one is controversial, but the way I am starting to think about it is that we are just moving to a new layer of abstraction. I no longer understand very well how hardware works, nor do I know in detail how a browser renders a page, nor the full fine details of how many of the libraries I use work. My own AI generated code, in pockets, is starting to work in the same way. And I'm starting to bec…

The implementations for those abstractions are well tested (hopefully). Committing code that we don’t thoroughly test or have an understanding about is going to bite us sooner than later. They are landmines. Not abstractions.

Re: Top Tips for Writing Code with AI

#7

Post by bsdpython » Thu, May 15, 2025, 3:36 PM UTC

sherdil2022 wrote: post #6

bsdpython wrote: post #5

Earlier quoted context omitted.

Thanks. I know the last one is controversial, but the way I am starting to think about it is that we are just moving to a new layer of abstraction. I no longer understand very well how hardware works, nor do I know in detail how a browser renders a page, nor the full fine details of how many of the libraries I use work. My own AI generated code, in pockets, is starting to work in the same way. And I'm starting to bec…

The implementations for those abstractions are well tested (hopefully). Committing code that we don’t thoroughly test or have an understanding about is going to bite us sooner than later. They are landmines. Not abstractions.

Is it possible we hit a wikipedia moment (it being more accurate than Encyclopedias) where the quality of typical ai generated code is better than popular open source libraries?