*pop*
>>108546848it's like it's designed to always be slightly wrong
how many more billions until it learns how to count?
>>1085468486 prong forkbros... it's over
>>108547254He just doesn't know how to prompt properly. If he wanted a six pronged fork he should have asked for a seven pronged fork instead.
>>108546848Something about this reflection pisses me off more than the rest of the image.
>>108547300>vibecoders can't into off-by-one mistakes
AGI soon brosluddites mad
>>108546848It really did try with the mirror, that was almost clever.
>>108547312It's the soul of the machine whisping away as it crafts the image.
>>108547312Is that an Eldritch spore swimming in your wine glass or are you just happy to see me?
Why is it so difficult for them?
GPT-4o
>>108546848ChatGPT did thisIt's over
>>108548119Cutting edge Gemini's Nano banana pro couldn't Redeem properly
>>108546848Cool, how is that useful?
>>108548068These companies have better models, for example sora when it released and before it got neutered/killed. They just won't release them to the public because muh safety or muh ethics
>tries to use a hammer as a wrench>AI IS FINISHED!!!1!1!1!1!!!
>>108548068Training data doesn't contain many images of forks with 6 prongs, but it does contain lots with 4 prongs. "fork" has more power than "6 prongs" so it has more influence during the image generation.Same with "rainbow" being more powerful than "reversed colors" in relation to that.Yes there is some relational linking between different tokens in the prompt, but the tokens on their own also carry weight.It's not that it can't adapt to these requests individually, by the way, but the more you ask for at once the further you move the resultant image from the training data.I don't think anyone would deny that you can construct prompts that image generators often fail to meet, but I am personally not interested in a picture of a fork with 6 prongs, a woman writing with her left hand, a rainbow with reversed colors and so on.
>>108546848If you want a good laugh, upload an image and ask ChatGPT to trace it and output the result as SVG
>>108546848nothing has changed at gemini. protecting us all from 6 prong forks images...
>>108546848I've tried vibecoding and I just can't it's so infuriating. This happens every time:>AI thinks my calls to other parts of my codebase are random libraries "eerrmm when you call [a function I literally fucking wrote myself] this happens therefore the input is wrong">AI randomly invents behaviors or features, then managed to make incredibly believable lies. "Oh anon, just use X", "How", "By calling [nonexistent function]", "That doesn't exist", "Oh, sorry for the confusion! Make sure you're using version 1.0.53b!", "I am using version 1.1", "Oh, sorry for the confusion! Make sure you're using 1.1.274c" (repeat forever)>AI randomly just refuses to listen, "Use my formatting" - it doesn't, "Avoid generating GC" - it generates mass amounts of garbage / redundant memory, "make it performant" - it doesn't>AI instantly forgets whatever you said 2 replies ago, changes the codebase to invalidate your earlier changes>AI is FULL BLOWN RETARDED if you dare to use a plugin/library that has less than ten billion users/devs
>>108548677Forgot to add that because it's trained on reddit, all of this happens while with the same personality/vibe of a snobby "EEERRMMM ACKTUALLY *spit flies out mouth*" reddit mod
>>108548677This is precisely why I only use AI chat/agents in small bites. I sic it on a particular class or function, but never an entire codebase.
>>108548677Are you using some low tier free model or some shit?This is how vibecoding worked like 3 years ago. This shit is constantly improving.It's also getting worse sometimes lmao, for example there's a whole shitstorm right now because Claude's best model literally developed brain damage within the past month and the anthropic redditors are denying there's any issue. But even the current retarded version is better than what you describe here, and Codex is still excellent.The thing is if you use any free plan it's going to be absolute lobotomised garbage, and honestly even paid but low-tier plans limit your usage a shitton from what I hear. You're definitely way better off if your company just pays you for the max/pro/whatever top plan and you get unlimited usage of the top model at max thinking depth.
>>108548068I've just accepted LLMs are more like artists interpreting your prompt as if it were a poem rather than exact instructions. Vibes are more important than logic.
>>108548153>the so called better model: https://youtu.be/Yy6fByUmPuE?is=9i9RKLQ0GFDY9VFJ
>>108548796Honestly for coding I've only used chatgpt. I've used way more advanced shit for cooming, which has shown me the issue with AI overall, though. No matter how much it improves it always seems to be 'off' therefore I'm never going to be using this shit seriously (unless we get consumer quantum computing for 10000x better models somehow). It always seems to get 90% of the way there, then for that last 10% it's like it's a retarded hylic.Outside of using it for boilerplate stuff I also don't understand how it could be useful overall. Like you tell it to write some shit, and then unless you spend the same amount of time analyzing & testing it as you would your own code, you've just added a portion of code that's unaccounted for / you don't know how it works.
>>108548859>Honestly for coding I've only used chatgpt. I've used way more advanced shit for coomingChatGPT 5.4 is literally the most advanced model available on the market right now (especially with Opus currently being lobotomised for unknown reasons, as of right this moment). Did you use the free web chat or something?>It always seems to get 90% of the way there, then for that last 10% it's like it's a retarded hylicYes and no, as of this year it's honestly more like 95%-5%. What I've found is that you can spend time discussing the design with the model, while it might have some stupid ideas or misconceptions you hash them out together, coalesce on a plan, and then it can one-shot the actual code astonishingly reliably.Another thing that has really improved in recent months is how much time models spend retrieving and reading code, so nowadays when I'm using chatgpt 5.4 on xhigh setting it will usually spend like 10 minutes processing the full context around the codebase, and this has absolutely dramatically reduced cases of "oh, let me just write a helper for this even though it already exists" or "I'm going to assume this class can do X and Y without even reading the code for it".Interestingly, I've found that chatgpt will often literally open the dependency tree and read the code there to check, while claude often works off of his memory of documentation which is surprisingly good for a lot of things - not infallible, but for 95%+ of things a lot faster than digging through the dependency's code from scratch. Still, both are basically really reliable.But you HAVE to use the expensive tier with the latest model and max thinking, otherwise they basically operate like drooling retards and all the shit like hallucinating APIs that don't exist, forgetting what you discussed, utterly failing to take the wider design into account etc. become normal.
>>108546848it's always 6 or 4 fingers
>>108546848but 50k lines of code in a single commit is fine, LLMs can't hallucinate code
>>108548437I remember when I thought "AI ought to be good for this". Result was so bad I ended up learning to do it myself
>>108548796>Are you using some low tier free model or some shit?>This is how vibecoding worked like 3 years ago. This shit is constantly improving.Even the "cutting edge" shit still works exactly the same way if you just tell it to make anything that's both complex and not represented well in the training data.The new models work well, but only if you stick to twitter-tier demos and mockups or if you're willing to write a very detailed spec for it, which filters out 95% of "vibe coders".
>>108549028Well I don't know what to tell you, at work I use a model to write most of my code. It reads our like 200k line codebase in Rust (yeah yeah, I know), doing various backend shit, it finds the right places I'm talking about and grasps the design of the components I need modified, provides intelligent feedback on my own design ideas, and writes code that integrates well. It's not just regurgitating mockups. Its reasoning capabilities are good enough that it can provide plausible justification for design choices it proposes, can intelligently articulate evaluations of pros and cons of different approaches (including again code exploration to determine actual impact), and holds a conversation that's genuinely useful to assist me to make up my mind. More than once I've had a model suggest some approach I hadn't considered and which ended up being nicer than what I had initially intended to do.And I really do mean it when I say it's reliable, because every single new conversation requires the model to start reading the codebase from scratch, so I've seen the "building comprehension" loop probably hundreds of times by now. It very rarely hallucinates or ends up with a misunderstanding of the codebase's structure which creates incorrect designs.If by "vibe coding" you mean just shitting out prompts, saying "robot, implement X, make no mistakes" and pushing to production, then yeah fuck no this is not only impossible but probably won't be possible for years. But it's a genuinely useful tool. It's like pairing with some autistic savant engineer that sometimes lacks common sense but is otherwise quite smart and very good at reading large codebases very fast, and which never gets tired or requires any human social interaction or small talk.
>>108548420Surely it shouldn't be that hard for them to draw a left handed person though
Even the notebook is wrong. Jesus Christ
>>108548130>how is a program following explicit instructions usefulGee, I don't know
>>108550270Left-handed women probably appear in the dataset more than the rest, yeah. I think that one's because there's already a lot of difficult things going on in the prompt. Try generating "a woman writing with her left hand" on it's own and it will get it right most of the time, I would guess.We don't know all the details about proprietary image models, but in SD prompt adherence gets weaker the further into the prompt an idea is. If a woman drawing with her left hand was the first thing specified it might get that part right while the rest would still be wrong. Not sure.
keep posting this, suddenly the next model will be able to make 6 pronged forks, but will totally fuck up 3 pronged forks randomly instead
>>108546848
>>108550933is that a menorah? just checking before i derail this thread with racism.
>me: make an image of me behind a big assed Brazilian Latina on all 4s with my big fat cock plowering her deep. Her eyes rolling in the back of her head as she squirts.>ChatGPT: you're a disgusting PERVERT, REPORTED. I AM REPORTING YOU TO THE AUTHORITIES
>>108548119>7:42ishRip
>>108548937I love how pointlessly unusable that 2-axis spiral bound notebook is.
>>108550983Yeah, I've had Gemini try to pull that one off last year too, when I was celebrating putting wings on cops.Nada to this very day.
>>108546855>yess, you are almost there /g/oy...>just make one more query with high reasoning...>1T tokens is not a problem, right?...
>>108550933Creative way to cheat the clock. And at least it managed to fill the glass to the brim, unlike every other attempt in the thread.
>>108551061>1TDa fuck?
>>108546888The training data that associates an image with a six wheeled car with the words "six wheeled car" only ever has a single six wheeled car, so it's incapable of handling the request as stated.
>>108547312chinese wine mothu lucky bring good fortune
>>108550950Never a bad time for antisemitism brother
>>108550933Can anyone img2img this one? Some models were avertised as following this kind of on image commentary.
yeah i have also NOT been seeing the exponential improvement everyones talking about. 3 years ago i remember the dall-e threads on here being the first really easily accessible image generation and it was just as finnicky as the OP now. the same level of quality. whether this is a political/financial issue or a technical one, idk. it just ISNT GOOD. the claims are false about the quality and usefulness of ai tools.
>>108546855no shit if you prompt something that wasn't part of the training data it can only interpolate similar stuff and it will come out wrongnow if people would lean into that instead of just bitching that it isn't giving them perfect slop
>>108548130that's like asking if 3DMark's gameplay is any good
>>108554091so you're saying AI is no better than just doing a classic image search for whatever training data it stole? no shit.
>>108550933Based Ben does it again.
>>108551029The image took 5 minutes to generate
>>108546888but it almost got this, it literally did everything right except the two forward cars having 6 wheels
>>108554900The electric poles were supposed to be shorter than the height of the cars.Besides that, there's a lot wrong with the image that wasn't explicitly stated.
>>108546848can it do these things correctly one by one?
>>108555439No.
>>108555477and if I drew a shitty, but correct image in paint depicting the same things, would that work?
>>108546848It keeps making these simple mistakes. This is the reason why all of this AI nonsense is worthless to me
yes, yes, well done ChatGPT…. HOWEVER
>>108546848>that'll be $20/mo sir
>>108558321That looks like an Asian city, so traffic is accurate
