Conversation
Notices
-
Embed this notice
@WandererUber I’ve seen them become real defensive bastards recently too “well technically I didn’t say that even though I wrote code that shows I totally did say that”. I get better results when I break down every problem into smaller chunks “okay we’re doing this big feature, but you’re fucking it up, let’s just start with problem 1, get this API call to work, easy enough. Okay, the API call now works, let’s do the next small chunk of work.” And you just basically just keep building on the small chunks of work that it can knock out like fish in a barrel.
That being said I also use Cursor at work where it’s directly plugged into the code base so it has a lot of “good code” it can reference which I’m sure helps it not be as retarded.
- BowserNoodle ☦️ likes this.
-
Embed this notice
@okayyeahwhatever It just doesn't know when to stop. got better at it though. early chatgpt was way overfit to that style and now it's just overfit to other things that are harder to detect. Some very good developers for example warn against this. It produces fewer obvious bugs than it once did, but it produces some that are extremely hard to detect because they are so dissimilar to how humans reason, yet the code is optimized to look highly human. similar to the chat output. Fable has a 50% hallucination rate for things it doesn't know. If you ask it to provide sources for the claims it has compiled into a document, at least 4/10 statements in my experience have an incorrect source attached to it.
This looks extremely like a well-researched human document, because a sloppily-researched human document has FEWER footnotes and NONE for the things that are ass pulls.
An LLM instead behaves like an insanely brazen liar and just puts hundreds of citations everywhere. No human is going to check all of them.
-
Embed this notice
@okayyeahwhatever I think they just backed it off because people started hating on them. It was the "wow factor", just like markdown is still today.
It immediately looks like a well-thought out summary when you have two or three emoji that fit and people think "oh nice so if I want the quick version I look at the little rocket."
when it works it works (llmarena noticed this problem years ago. People preferred the cooler looking response over the correct one)
stopped working. So they stopped doing it.
-
Embed this notice
@WandererUber yeah I think that’s true. I had to build an auditing solution at work and emojis work well for a quick spreadsheet, “okay this is green, means it’s deployed, this is red, means its orphaned”. that’s a use-case I can get behind.
but seeing them in some pipeline output throws me into a rage. everyone at work now just uses the emojis like the rocket ironically when we’re making fun of AI or the corporate push to ”AI” everything.
-
Embed this notice
@WandererUber A big giveaway is the em-dash - AI fucking loves that damn em-dash.
-
Embed this notice
@okayyeahwhatever
-
Embed this notice
@WandererUber I‘m always like “okay remove the fucking em-dash from my code you clanker piece of shit.” It fucks shit up sometimes depending on if you’re working with various UTF encodings. They also love emojis, although seemingly that’s been backed off a bit, my theory on that was emojis are probably more “universal” so the AIs were adding them because your Indian colleagues know what the image means when they might not be able to read the words. They definitely do that less now, we had so much code with fucking emojis in it I kept going through PRs and tearing that shit out.
-
Embed this notice
@jb You wrote this?
-
Embed this notice
No AI here
If you're having fun then carry on
If you're accusing me of using AI then fuck off and stop reading them