Post

"13 butts pooping, back and forth, forever" 😂😂 amazing description of LLM spaghetti code.

No wonder it's burning a forest for every line of code generated. This is horrifying and hilarious by turns.

This is just fucking embarassing. Computers do exactly what they're told and what even is this?

I mean I do feel that, any time I can look up on a clear dark night, but their tone is rude.

Oh my god

This is amazing.

The LLM-to-JSON thing is how everyone handles LLM to json. You just have to pay for the structured output. It is one of the reasons I hate working with these things.

I hope we all feel this way and destroy this pathetic industry soon

Late to reply here but … like why don’t people working at the company know why telling it not to do bad things doesn’t work?

It's like that meme "Remember to eat...Remember to drink water...If you're mentally ill,,,fix that..."

OTOH Code Complete (qv) is very big on decent, readable variable names, so maybe this is 'best practice'? Haha, only serious.

I'm happy with all this affaire, because i can confirm that AI no longer exists, no have existed. It's only a bunch of REGEXP psychobabble.

it literally works. i'm not kidding. the ai just predicts what someone would say based on context without any preconceived ideas, so it will not factor in responding correctly and not breaking basic rules unless you tell it to. "please do your job well" is a part of context you need to specify

CONST BOOL DONT_FUCK_UP = $true

You're not a billionaire so it's ok plus not a fascist not ruining the planet etc

They have at least a magazine cover a month saying it (without any pushback by the journalists). A lot of the illustrations look like the monolith, too.

The Apple Intelligence system prompt was revealed in 2024 and one part of it said “do not hallucinate”. (https://arstechnica.com/gadgets/2024/08/do-not-hallucinate-testers-find-prompts-meant-to-keep-apple-intelligence-on-the-rails/) We live in the stupidest timeline!

The fact that they literally think they're giving instructions to an intelligent entity, but also have expressed no issues about the fact that they believe they have therefore enslaved an intelligent entity that is forced to do whatever they want, should be informing more how we think about them.

We know *generally* how they work, we think, but so much of it is emergent behavior we have little idea how they work in detail. The one thing we know is they are not quite as good as human beings explaining how we think. And we are terrible at it.

I am at a loss for words. That coding was hysterical.

It's SO BAAAD No fucking wonder they need so many data centres, it's fully anti-optimised 😭

Holy shit. Holy. Shit. I don't really have anything substantive to add to this. I'm just *shook*

I should have stuck with programming in college. Sure I used if>then & and/or for everything but it was better than all this. 😅

The instructions have to go into the context somewhere, but this is sloppier than I would’ve expected. 🤦‍♂️

Well it implies that code correlated with someone talking about not making mistakes is generally better than code without that correlation, which might be true but someone who was treating it as technology rather than magic might continue to step 2 rather than just prompting.

Relevant to this too bsky.app/profile/oliv...

Dear Claude v3.4.2, Please rewrite yourself but with fewer bugs and security holes.

I thought I'd put this quote-post that provides alt-text in the replies too bsky.app/profile/pook...

What if I use Claude as a programming assistant and I ask it to create a negative language filter for my code. Shouldn't it use regexp in the code?

AI valt door de mand.

The list of "negative" phrases is so hilariously American.

No doubt, a prompt was "do xyz if user is negative". If this was just some unseet prompt/setting, it'll work fine. But clearly this instruction was absorbed into it's task of making code. So it didn't recognise "oh, this is something that the model can do", nope.. just "ok boss, heres the code".

So many issues I see being brought up by AI jockies at my work and my and first thought is why not regex? You would be done and could move on. What is really weird is they prompt and prompt and don’t think to ask the LLM to write the regex, which it should be good at.

this is a good example of the triviality of this problem as a decision tree, which any regex is encompassed by (unrolling loops). arxiv.org/abs/2210.05189

tag yourself, I'm "What the fuck hell fucking"

Jesus.

@sven.blacksky.app look at this fractal 🍩

Lmao I literally saw someone yesterday try to use that as an example of "it's coding itself now we have basically reached the singularity if you don't think these models can code new full models you're stuck in the past" this is incredible

it's literally cargo cult code as a service. this is amazing

bsky.app/profile/jaks...

Me and the boys ignoring everything these cultists claim to try to make their slop machine seem interesting.

oh my god this is such a mess

It's Dickensian in almost every sense of the word, as bleak as it is long.

Not by the rules of ©alvinball!

Doesn't this mean Claude Code is in the public domain and the takedown notices are invalid?

I laugh so hard, thanks so much ❤️