.

@spirodonfl Yes weve heard a lot about how great the new models are at programming but its not obvious to me theyre moving forward at all on that. theyre just better at getting stuff on screen, and the convolutions they take to get there are the code version of 6 fingers will smith spaghetti

I got a chance to try Fable 5 so I gave it a pretty simple task. Draw lines in a pixel renderer in O(1) space and then use that to extend it to circles. The most simple portions, it got right. Then it proceeded to overcomplicated others and, in fact, it overcomplicated even *more* after I gave it the simpler code I had already created and am using for my own custom renderer. The code it produced *worked* in terms of output but holy crap did the line counts blow. It also named variables poorly, was inconsistent in certain basic things like code prose and more. It was so mind boggling to read. It was almost like every single line you read was created by a completely different person. Every. Single. Line. Again, I repeat loudly - if you cannot code better than Fable 5, get out of my industry please. I don't want you. Someone who is relatively new to programming and has good impetus to care about their code will definitely do better than this.

@AliZYousuf @MarcJSchmidt The ai companies have infinite tokens and experimental models and keep shipping broken slop that neither more ai nor human intervention can fix, they seem to just occasionally do full vibe coded rewrites that are then broken and unmaintainable in different ways

@AliZYousuf @MarcJSchmidt Ok but theres actually plenty of ai generated open source code out there made by people with access to frontier models and tons of tokens behind them and its also all bad. E.g. bun. Meanwhile all of the software the ai companies (with infinite tokens) are making is broken slop

I'm still waiting since Opus 4.5 came out for the game changing next open-source project to appear. If everything got 100x easier, where is the next C++ killer, where is the fast ORM without any compromise, where is the next GUI framework that kills Angular & React, where is the next Linux Desktop that kills Windows, where is the next full-stack framework that actually works, where the fuck are all the products? I only see more mediocre code, lots and lots of mediocre stuff that doesn't stick. Some people got lucky by sheer probability but not by taking advantage of their "100x coding agent". Where the fuck is everything?? It's been 6+months now with powerful coding agents, equal to almost a decade in old engineering time. My bet: The first 80% of a project can be done in like 1 day of work. And the last 20% got turned into months if not years of fixing claude.md, babysitting, aligning, "skills", mcp, endless rounds of refactoring and cleanup sessions. I literally don't see the effects that justifies llm psychosis by software engineers anymore

@Msungba_Michi @arg_k7 @flyanew what anime pfp said was hallucinations are inherent to llms and cant be eliminated. what the paper says is hallucinations are inherent to llms and cant be eliminated and the way they found to mitigate it a bit at the pretraining level ruins the model's ability to reply to queries

@Msungba_Michi @arg_k7 @flyanew - this doesnt relate to your claim that hallucinations only happen if you prompt like an idiot - this doesnt relate to the paper you googled and linked w/o reading it - posting the same thing as before w nothing added except being more upset doesnt help you beat the allegations

hi, actual programmer here we do give a fuck about ai. and your program is both written like a toddler got a hold of sublime and codeacademy for like 5 days and also it looks like trash and anyone with basic fundamental programming knowledge could clean it up.

@Msungba_Michi @arg_k7 @flyanew Sure can you explain what you meant in the prior post where you said hallucinations dont happen if you prompt correctly and how that relates to the paper that i know with total certainty you didnt read

@cheddargoblin14 @arg_k7 @flyanew It isnt just with prompting lol im not surprised you took it to such a simple level. Prompting like an idiot definitely worsens their occurrence rate though -_- do you need anything else explained to you?

@Msungba_Michi @arg_k7 @flyanew Which supports what the anime pfp said, not what you said. What you said was that hallucinations are solved and can be avoided with prompting. This has nothing to do with the content of the paper and is so retarded no other human being has ever said it to my knowledge

@Msungba_Michi @arg_k7 @flyanew its the same generative/assertive process and "hallucination" is the label attached after the fact when the assertion turns out false. In the discussion, they explicitly note suppressing/amplifying these neurons trades off against the models general helpfulness and capability

@championrat02 Why? Might want to give that a reread. They didnt say anything about using software written with ai. We're all constantly forced to use software that retards wrote with ai. Its famously very bad. People complain about that all the time. What point did you think you were making?

your ass better not be using any software written after 2023. no games, no OS, no drivers, NOTHING.

@heart_2243 @himtopia ? LSPs, debuggers, code snippets, what do any of these have in common with LLMs trained on stolen code? Do you know what a debugger even is?

@himtopia Alright, disable your LSPs, code snippets, IntelliSense, and stop using debuggers. This take makes 0 sense; code is not an artistic medium. Code is to a product what a pencil is to art.

AI code is equally as bad as AI art. Both are trained on stolen work and are used as creative forms by people be it in Games, programs ect... Either both is bad or both is good but there shouldnt be double standards. 🤷‍♂️

@Koteghe @Much8Less @himtopia stackoverflow also required people who posted there to agree to license their code for readers as cc-wiki which still requires attribution. you were never allowed to just copy code from there. plus people always made fun of teenagers who couldnt code and copied off stackoverflow

@Much8Less @himtopia /gen aren't those websites where you can post code on your own? as far as I knew they weren't scouring the whole of internet for it and taking it from people's works

@Msungba_Michi @arg_k7 @flyanew Nobody who knew how llms were implemented was ever confused about why hallucinations happened. This isnt a response to anything anyone said

@arg_k7 @flyanew arxiv.org/abs/2512.01797 We actually know why hallucinations happen, btw

@Kazumenko3228 @firstLast31415 @verycoldworld @Nommiin That's a very bad sign for your code review abilities, Kirill! Praying for you!

@cheddargoblin14 @firstLast31415 @verycoldworld @Nommiin We produce AI code and always review it. Since 2025 AI code has improved much more, and we barely do changes to the code it produces

seeing a lot of the “ai coding is fine” crowd on my timeline, like 90% of them aren’t programmers 🤔

@firstLast31415 @verycoldworld @Kazumenko3228 @Nommiin Right yeah but theyre only interested in making bad software, like they were doing for decades before ai. Ai makes it easier to spam out more slop faster, which they used to have to hire hundreds of devs to do. Again, have you ever read ai code? Itd help you understand my point

@cheddargoblin14 @verycoldworld @Kazumenko3228 @Nommiin Skill issue. All of the 20+ big tech & fintech SWEs are writing well over 50% of their code with AI. Output has gone up notably, and strong review agents help to ensure a degree of quality.

@verycoldworld @Kazumenko3228 @Nommiin wow nvidia says its good at programming. i find them very credible in this conversation! certainly enough to override the fact that ive read a lot of ai code that ai guys insisted to me was good and was in fact "throw it out and start over" level stinky slop

@cheddargoblin14 @Kazumenko3228 @Nommiin Yeah man you’re literally just lying. The most valuable companies in the world are using this code, if you’re saying you’re a better coder than senior devs at Apple and NVIDIA then sure but link your GitHub

@Kazumenko3228 @verycoldworld @Nommiin Yeah, again, ive used both and have read a lot of code generated by both by other people. people sometimes send me their ai slop when i make comments like this and its uniformly worthless, unfortunately. neither opus nor fable seems capable of writing code thats worth keeping

@cheddargoblin14 @verycoldworld @Nommiin Give it rules, use Opus 4.8 or Fable 5 Without rules it doesn't have a basis on what exactly you expect. Review the code and correct it

@perrymetzger @billassault @ImLunaHey Where do you think it learned to do it perry? Anthropic failed to use claude to create a c compiler in rust a few months ago and most of the logic and design/engineering decisions were copied directly from chibicc. This kind of translation is 1 of the 1 things its good at

@billassault @ImLunaHey My AI is writing language right now in a programming language that I invented for which no examples exist online and no documentation exists online. I’m not sure where it would be stealing the code from. Can you give me some ideas?

its so funny seeing normies crashing out over programmers using ai because it's "stealing code".

@verycoldworld @Nommiin Yes ive used all of these tools a lot, thats the basis on which i say theyre not capable of programming. Have you ever read the code they generate / any code before? They are not suitable for programming tasks. Again, theyre ok for searching in a code base, basic snippets stuff

@cheddargoblin14 @Nommiin Go on Claude right now and ask it to code you a simple software, plugin, or app. It’s beyond capable.

@verycoldworld @Nommiin Yes what's certainly true is that no one has demonstrated it yet. but i can imagine that changing, so i am agnostic

@cheddargoblin14 @Nommiin Ever be capable? Are you joking lmfao