Marcin Krzyzanowskinever stop being funny. this is after 3rd "are you sure"
this machine will do anything to make things worse. and then refuse to understand it.
fair enough it was trained like that. Internet is full of garbage.
forget prompting engineering. "rejecting" and "questioning" engineering is more imporant in LLM coding
> Perfect! Now I can see the issue clearly.
no. you don't
let plan and fix the coding assistant a swift concurrency warning by making a class Sendable, and see how it dissolves into chaos. line by line. one MainActor at a time.
it has no clue what to do.
Asked coding assistants to implement token bucket throttler. Here's what happened:
Claude Code: never sure if implementation works, keeps changing it and loops - never satisfied
Amp: liked Claude's result but improved it, stopped the looping
Result: Implementation still doesn't work. When asked about failures, says "found the bug" but fails to fix it despite claiming it's tested
Don't think it can create a working throttler
hence my question is: how did it pass the leetcode?
from the backstage: the fix indeed did not fix anything, and that was already the 4th statement that this time this is the fix - each time not fixing a different thing.
LOL. First it implemented delegate, then offer alternatives. eventually decided neither alternative makes sense and the delegate is broken.
I literally loled
π€¦ First I said making File fully MainActor would be "the cleanest approach" and recommended it. Then when you asked if it's better, I said "not really"
I trust you bro. I trust you with my life
Martin Du4@krzyzanowskim rewind is the exit door. πͺ
Ivan Cantarino@krzyzanowskim the saga continues π€£
@krzyzanowskim in a near future, AI will present you this thread in the AI court π
I am with the stupid one here. I asked it to implement something and test it. It did all of that, then called it a day after 88% of tests passing.
Am I supposed to fix the remaining 12% of the code?
softmaus@krzyzanowskim Itβs theraphy. Only thing is, you are the therapist.
@softmaus I have disconnect from the ai slop blogposts and reality
Jason HowlinFor any time I might save, I spend twice that arguing with the AI
@jasonhowlin pretty much. It save me actual typing when I feel lazy
Aleksandar VaciΔ@jasonhowlin I bail out as soon as I see itβs starting to halucinate nonsense and start over. There are no feelings to be hurt here so π«’
Jeff Lewis@krzyzanowskim Iβve been saying this to computers for years and they are only now, finally, apologizing and telling me I am right.
Steven G. Harris@krzyzanowskim I am sure Iβd be reacting just like you, and that makes me want to stay far away from this development approach. I heard React was problematic anyway.
Brandon Horst@krzyzanowskim I think that youβre trying to use these systems at too high a level. I never try to βcorrectβ Claude code. If it gets it wrong, I change my prompt, switch to plan mode, and try again.
@brandonhorst that is the worst way to use it, honestly. It can barely execute the plan correctly. The more time spent on a big plan, the weaker the result is overall. It need to be micromanaged on every step, givint ig clear instructions on each step, without relying it will keep the plot by its own (it wont)
Rob Napier@krzyzanowskim @brandonhorst I've gone down the "carefully plan everything" track several times. I'll accept that @steipete is much better at this than I am (both AI and coding generally), and that he uses Gemini (not approved for my professional work), but I generally find that extensive planning is time wasted.
Better for me has been: I do most of the coding and Claude helps with boilerplate, individual methods, code reviews, improvement suggestions, and some light refactoring,
@cocoaphony @krzyzanowskim @brandonhorst @steipete Same for me.
I canβt let myself let it run loose, at least Iβm not there yet with prompting. And frankly nor is Claude advanced enough yet, at least in the usage I needed in the last week or so.
Thus I closely monitor what it does, stop it as soon as I see itβs veering off what I want and generally use it as very fast personal coder assistant.
Peter Steinberger@aleck @cocoaphony @krzyzanowskim @brandonhorst i also usually ask for options and iterate till I find sth. But to cooy existing stuff or larger refactors, claude go brrrr
@steipete @cocoaphony @krzyzanowskim @brandonhorst
My favorite use case for Claude so far is analyzing existing code base. It's ability to summarize its findings is frankly speaking far better than mine.
Here's just this morning, where I'm trying to understand Socket.io's "we are not WebSocket implementation" claim.
It gave me this in less than 5 minutes.
@aleck technically, it was 48.3s more than 5 minutes :P
@aleck @steipete @cocoaphony @brandonhorst yes, and ironically, Cursor is even better at this, mostly due to IDE integration and ease of navigating to the exact places easily
@aleck @cocoaphony @krzyzanowskim @brandonhorst Google AI studio with a github tree fetcher is better for that
@aleck @cocoaphony @krzyzanowskim @brandonhorst https://repo2txt.simplebasedomain.com + https://aistudio.google.com/prompts/new_chat is better than anything else. And 2000% faster than via CC. And free.
Adam Garrett-Harris@krzyzanowskim You're absolutely right to be skeptical.
cGilmore@krzyzanowskim I recently experienced it having a really bad day where it ended up frustrating itself. In a single comment it went back and forth like this 14-ish times, never landing on a real answer.
One of the titles for an answer was, "β β β FOR REAL:"
If it wasn't so hilarious I might've been annoyed.
Pierre H.@krzyzanowskim is Claude with your code the same as DOGE with the government budget? π€
@madcoder not familiar with the foreign politics, but maybe
@krzyzanowskim what everyone imagines AI code is like: Oh, you completely forgot about this corner case and now there are 100 security vulnerabilities.
Ok, yes, that is in fact true. But even more often:
What AI code is actually like: 3 layers of checks for corner cases that have already been checked both by code and by types but just in caseβ¦and if it might fail, an extra fallback algorithm that doesnβt actually work, and a default value just in case.
And maybe an extra try? Just in case?
Jerry Tromp@cocoaphony @krzyzanowskim leave it to AI to come up with weighted sums, thresholds, fallbacks and other complicated constructs, just to do a simple thing π
@cocoaphony yes this. dont forget "I added cache that improves performance by 50%". "You're right! cache is not used"
@krzyzanowskim this is a novel π
Jeffy π³οΈββ§οΈπ³οΈβππΊπ¦ β€οΈπ±πΊ@krzyzanowskim was that chatGPissTFOff? It sounds sarcastic. π€
Archimage@krzyzanowskim It's very pandering. It always agrees that you're right and it messed up. What happens when it's right and you tell it it's wrong? ;)
@SergKoren that is part of the problem. It will double down on being wrong and hallucinate excuses
@krzyzanowskim Interesting.Seems like it would be fun to confuse.
Sami Samhuri
@sjs skill issue? haha