Conversation
Harry (Hyeonggon) Yoo
hyeyoo
Edited 19 days ago
I don't find code commentary helpful unless it explains the design first, because these days the implementation is way too complicated for readers to grasp the design by reading the implementation.
Oftentimes such code commentary mechanically describes list of statements, rather than explaining the idea behind it.
A well-written document should try to describe the idea behind the implementation, rather than the implementation itself. (Yeah, that's challenging)
They don't save much time compared to directly reading the code.
(some random rant of the day)
Oftentimes such code commentary mechanically describes list of statements, rather than explaining the idea behind it.
A well-written document should try to describe the idea behind the implementation, rather than the implementation itself. (Yeah, that's challenging)
They don't save much time compared to directly reading the code.
(some random rant of the day)
2
1
6
1
0
2
Vlastimil Babka 🇨🇿🇪🇺🇺🇦
vbabka@mastodon.social
0
0
1
LogicalErzor
Logical_Error@fosstodon.org@hyeyoo ai is great to get most of the way there as it understands languages and translations well (c to english)
for subsystems u understand, ask it to summarize the implementation and review its changes. u’ll be surprised at the amount of work it’ll cut out for u (depending on the model)
1
0
1
Harry (Hyeonggon) Yoo
hyeyoo
@Logical_Error yeah AI is quite a nice tool if you can review them.
But decoding the ideas behind the code from history and past discussions takes time (no matter you are human or AI)
But decoding the ideas behind the code from history and past discussions takes time (no matter you are human or AI)
1
0
1
LogicalErzor
Logical_Error@fosstodon.org@hyeyoo i think most models are trained on the kernel’s git commits and lore msgs, so it can prob point u in the right direction
1
0
0
@Logical_Error some people say "we're running out of data to train AI, and that's a problem. we need more data"
but no. that's not the problem. the problem is that you can't make LLMs that are experts on every single field, even after training them on all the public data.
but no. that's not the problem. the problem is that you can't make LLMs that are experts on every single field, even after training them on all the public data.
1
0
1