hartmannathan commented on PR #19004:
URL: https://github.com/apache/nuttx/pull/19004#issuecomment-4603779055

   > @hartmannathan I had an idea and plan for this reorganisation for years in 
my private notes, but it was not possible to do it manually without investing 
months of work. This wasn't a task that could be solved with a few prompts. The 
AI has too small context for this and isn't able to think outside the box very 
well. AI was used here more like an "intelligent tool" for things that I would 
have had to do manually. I don't have a single master prompt but used many 
simple prompts in an interactive workflow. I just instructed AI how I would do 
it step by step, like: "rename all public STM32 arch API to use common prefix 
stm32_", "move all stm32f0l0g0 files to common/stm32 and rename them to use the 
m0 suffix", "build set of configuration after each step" and so on. In fact, 
the order of commits exactly reflects my plan from the beginning.
   > 
   > With such a huge change, keeping git history clean manually is a really 
big challenge, but for AI this is a really simple job. Moving files, fixups, 
squashing commits, and resolving conflicts is something that takes hours for a 
human, but can be easily done by AI. Fixing issues found by local build/CI is 
also easy for AI, but not an entirely automated task and needs careful reviews. 
Simple for human solutions like: "Merge SPI Kconfig options with I2S options to 
fix undefined STM32_SPI_DMA symbol" were completely unthinkable for 
codex/claude/deepseek.
   > 
   > There were more issues that AI was not able to solve without manual 
intervention, like optimizing complex dependencies in Kconfig or even merge 
similar header files. There are still many complex conditions in Kconfig that 
must be simplified, but this is work for a human and careful analysis; AI has 
no idea how to properly solve it.
   
   @raiden00pl thanks for this explanation. It's nice to hear an explanation of 
how useful AI really. As an AI skeptic, I have been hearing a mix of two 
extremes: some people say AI is so smart it does all their work for them and 
it's just great; other people say it sucks and only hallucinates slop. It's 
hard to tell whether those with the great experiences think it's great because 
it really is or because they don't understand the subject matter. And 
conversely it's hard to tell whether those who say it sucks because it really 
sucks or because they don't know how to prompt it. I guess, as you're 
explaining, it's useful but it's not as smart as the hype would have us believe.


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to