On 12/26/25 16:08, B9 wrote:
On December 25, 2025 2:54:55 PM PST, Kenneth Pettit <[email protected]> wrote:
On 12/23/25 2:39 PM, George M. Rimakis wrote:
He (please bear with my anthropomorphizing of the LLM), found out that somehow
it was causing the GUI to crash.
I honestly don’t exactly know what he did, but he modified the code so that he
was able to safely push keys at the same time as a user in the GUI, without the
app crashing.
Would be really interested to see what changes Claude made to resolve the GUI
crash / deadlock / garbled output if you are up to sharing them.
For the sake of science, I'm glad to see George has let Claude file the bug report instead of
typing it up himself. Understanding the code — or trying to — would end George's experiment in
"vibe coding" in which one "embraces exponentials and forgets the code exists".
I'm fascinated seeing a hybrid approach happening in real-life: We have a vibe
coder making a patch for existing code that an expert coder will examine.
Clearly it saved George, the vibe coder, time and effort. If he had to
understand the original program, identify the bug, and learn enough to write a
patch, he probably wouldn't have even tried.
Now, the question is: How much time and effort will it cost Ken, the expert
programmer, to handle the AI generated report? Would it have taken Ken more or
less time if he had used Claude interactively to fix the bug?
Bug reports with patches have traditionally been very helpful, but there are
costs, especially with AI. Does Claude's description match the code generated?
Has it identified the actual source of the error? Is it the correct fix? Does
the patch break anything else? Does it add extraneous changes? Does it take an
incremental approach when refactoring would be more appropriate (or vice
versa)? Are the changes written in a way that will be easy to understand and
maintain in the future?
Ken, if you don't mind, please let us know what your experience is and if you
would welcome more patches from vibe coders in the future.
--b9
I'll just say this much:
I haven't commented on this thread at all yet because John doesn't like
us to be mean to each other here. But I am very much in the Daniel
Stenberg camp on this.
The latest post where George regurgitated the AI's own explanation for
what it did (instead of understanding and explaining, and taking
ownership himself) sent me over the top.
Ken is nice, but that just means Ken is nice. Good for Ken and lucky for
George.
It's still basically abusive of others to submit work from an unthinking
unknowing tool that you yourself can't explain, vet, sign-off on, and
put your own name on and take responsibility for.
But a non programmer doesn't understand what the problem is, and I'm
unable to write the explanation in a way that is likely to be read.
Actually it's not even really a programmer thing, it's a respect for
others thing.
If I asked for a cake recipe, and you used an ai to generate a cake
recipe and gave it to me without yourself being able to say that the
cake recipe will never poison anyone, or even just taste bad or fall
apart in all expected usage scenarios, then how can you live with
yourself? If I asked for the formula or table to map how much water a
field will need based on the type of soil and the type of crop, and
instead of performing the research to actually know the correct data,
you just had something generate an answer that who knows, might be
right, how can you live with yourself?
Somehow people do it for coding and think it's fine.
--
bkw
