>>108466368
LUDDITES WON
SLOPPERS HOLOCAUSTED AGAIN

>>108466368
for real though, how pathetic is this
>language models train on text with patterns, generate text with said patterns
>code is the perfect training data and perfect use case
>huge amounts of readily available training data
>huge market of enterprise and code monkies
>still an unprofitable unusable scam

>>108466368
openai just had an epic comback with 5.4 thoughbeit

>>108466368
bc they're running out of money like every other company

im having to sit through a meeting where some guy shows off a vibeslopped tool he made, and its already broken 5 times during the presentation

>>108466411

Code/programming is NOT a prime target for AI-ification actually. It just appears like it is superficially, just as one might think coding is "easy" after learning the ABC's and 123's of it, writing a loop, making a class. Not really how it works in reality.

Code is NOT fault tolerant at all. One little mistake or oversight breaks the entire thing. AI is a "good enough" machine that is about 90-95% correct and that makes it basically unusable for coding at scale. Software is just like any other sophisticated machinery or device, tons of components, tricky wiring, one breaking breaks the whole thing, the only reason we have fooled ourselves into thinking this is ok is because it's virtual, it's a nice show.

Lastly (and most nefariously), we have been through this process before. Code VOLUME is not valuable. In fact the essence of good code is doing as much as possible, with the least code you can, written in the most human intuitive way. This is a very tough ask for a machine that doesn't think or read like a human. This is the blending art and science part that makes it a higher bar to just automate away, or even outsource to Indians (which they also keep trying). It creates massive, crufty, broken spaghetti messes, that perform badly and are impossible to maintain or extend. Again and again and again industry newbies rediscover this fact. GOOD CODE IS SIMPLE, BRIEF, PREDICTABLE, READABLE. AI cannot handle these inherently human nuances.

What I've been doing is using AI in an advisory role, asking it conceptual questions, asking it for ideas, letting it spit out a contextually limited method, and that's been helpful. But if I ask it to "design a class or module that does X thing" it shits out a mess.

>>108466464
@gork summarise

>>108466464

The only reason programming has become such a huge target, is because 1. Big tech hates their engineers and has always wanted some way to deus ex machina away their dependence on them and 2. locality, obviously engineers developed AI and immediately turned it on themselves. But software is in fact FAR down the list as far as AI replacement candidacy goes. You have to look at things where the occasional mistake is ok, where mistakes don't compound.

>>108466489
>where the occasional mistake is ok
Sure but
>Medicine, law, finance, military systems
"Thought leaders" ae trying line their industries right up behind software to solve the human tax of business.

>>108466368
Why is their logo an anus?

>>108466464
@claude make shorter and do not make mistakes

>>108466489
ultimately managers don't care. If something breaks, as long as they made shareholders happy: they don't give a fuck if it flies the company apart 5 years from now. Most of them will be off to find higher paying jobs by then for being so good at saving the company money.

>>108466524
I always wondered why tf nobody at anthropic mentioned the fact that claude's logo is a butthole.

they couldn't even make it a goatse reference

I've found claude to be extremely handy for managing an existing codebase as long as the codebase itself is well structured. I've been able to implement new features non-stop to my projects and none of it is slop because it imitates my existing architecture and code style. It's just like I would've implemented the features.

>>108467809
This is true for everything AI does. If you fill its context with good content it'll perform way above average, if still a bit worse than what you put in. The problem is that after that, its context includes its slightly worse output. So it will make something a bit worse than that. And so on until it converges to the slop average.
It's a great way to cash in the opposite of tech debt (tech credit?) for some speed, but there was never a lack of ways to do that. Most codebases are on the threshold of bankruptcy though

>>108467827
Probably true, in my case I've actually refined claude's output manually because usually that means just few tweaks here and there.

>>108467827

Yes this. Compounding mistakes. Compounding debt. The #1 issue in software is compounding. And if your AI isn't aware of what the original issue was the butterfly effect kicks in, it rolls it over adds new problems rolls it over adds new problems. This is why it simply can't be left unsupervised.

The concept of "agents" doing all of this on their own for hours and days makes me howl. Imagine the absolute dog turd of a code base they are patching together.

>>108466368
I hate their laggy ass website so much, it's unreal. Truly, a vibecoded abomination.

>>108467839
>>108467841
recently watched a ghotz vod where he went through his agentic coding setup just to see how a cracked programmer uses these tools and yeah you have to be EXTREMELY vigilant not to let the slop seep in
https://www.youtube.com/watch?v=erBX3gTZqJI&t=1250s

check every diff, make sure you understand the changes, be extremely pedantic when it comes to code quality and brevity

i'm on open source models but i try to do the same for projects that i care about but i know there's a lot of people that don't even read the code anymore

there is this little demon on each of our shoulders whispering to just let the agent rip and be done with it, let the shitty code be someone's future problem

they were always shit. as soon as they differentiated themselves years ago, they were already doing retarded limits. their business model and, comically, their marketing is akin to a crypto pump and dump shitter.

>le heckin new model
>deploy UGC shills everywhere talking about how its the best shit ever and how they one shotted generational wealth
>soak up the subscriptions
>lobotomize it
>jew the limits to one message per week
>repeat

they don't have the pockets to compete with scam altman, so they brand themselves with earth tones and use pale yellow instead of white, and tune the system prompt to be the quirky chungus of giga surveillance cybernetic nightmares.

claude "not like the other chatbots" ai.

eventually chatgpt will have the retarded limits but, not until anthropic folds, which they will.

if people shill anything other than the 200 dollar openAI plan + codex cli + using the web browser pro model for tougher questions, they're full of shit and lying. nothing else compares to the amount of quality and usage you get.

agents are a hilarious meme like mpc was. it's literally all just markdown files. every single "innovation" built on top of llm's that weren't some form of gen ai themselves, were just a json file or an md file.

send anthropic to zero

File: 1774179433303143.png (24 KB, 611x614)

24 KB PNG

>>108466646
>they couldn't even make it a goatse reference
you don't say

>>108466368
that "half" was just openai offloading deadweight consooomers that waste context and feed useless data like asking about the weather. Basically a trojan horse to burn your competitor to the ground

>>108466464
It is a good target since the output can be machine verified and scored. That's why it's being pushed so hard. How do you machine verify a summary or report or creative writing? If you want good results you need a human constantly in the loop, LLM scoring will go off the rails

>>108466390
Cope, you will be replaced by aigods

>>108468211
this is not limited to Anthropic, any Venture Capital Startup in history behaves like this

Uber, AirBnB, AI Labs, whatever

File: 1748038503168001.png (61 KB, 704x306)

61 KB PNG

>>108466443
>openai just had an epic comback with 5.4 thoughbeit

>>108468465
by what?

>>108466368
what the fuck are you talking about idiot im using this stupid shit every day its amazing but thats probably because im lazy and i code with my asshole

>>108468621
5x models are shit but 5.4 is legit

>>108466464
You don't get it at all. Code has progressed because it's easy to verify if something runs and gives the appropriate results. Letting the LLM try a lot during training and knowing when it works, with additional steering, but it's the fact that when it works it works that makes it AI-ifiable.

>>108468439
There's no objective way to score a summary or creative writing, sure. There's also no objective way to score good code, so it's not that different? "It runs" is a fucking low bar. "It passes tests" is also, especially when the llm writes the tests. If a human has to write tests strict enough to keep slop in check it'll take him longer than writing the code himself.

>>108466368
energy costs have increased massively since Team America and Israel decided to blow up the middle east. you're already seeing the effect of these companies scaling back their expenditure (OpenAI retiring Sora, Google increasing their usage restrictions on Gemini, and Anthropic rate limiting their shit even harder). OpenAI and Anthropic also recieved a lot of funding from the gulf states that now have multi-billion dollar repair bills to pay instead.

long may the war continue.

>>108472438
Gulf States are all happy about Iran getting fucked, though. Long-term it is good for everyone. Except maybe Russia.