Is AI getting worse? It seems like it was actually usable for about 6 months and now it won't produce any usable code. Are they gimping the models to try and turn a profit?
Based teens.
source my ass
>>106514100what does d1gg2r have to do with that
>>106514100>Is AI getting worse?No, only ChatGPT got worse because they lobotimized it a bit too much.The rest of the LLMs hit a hard ceiling and won't be able to improve until there's a new major breakthrough or less expensive operational cost.
>>106514100Here you go zoomers, this is your yoomer moment to hold a permanent grudge against the next generation>whippasnappas frfr, always bussin wit'hat AI gf and shit. Real and cringecoded.
>>106514675Many other models are just "distilling" chatgpt. So, yeah this gay ai economy is fucked. Total Luddite Vittori.
>>106514675GPT is the only one that sorta works. The others completely fuck my code base up for simple tasks. It's like we went back in time a year.
>>106514776Try Claude
>>106514783I'm specifically talking about Claude. I think Claude got a little better at some non coding tasks I use him for but he's much worse at coding. Claude and Gemini. They must be gimping us. The crackdown on subscription use was the canary in the coal mine
>>106514783>>106514822Lads(Luddite chads), we just can't stop winning nowadays. It's getting kind of stale.
>>106514868I'm fine with or without it. But it's nice to be able to ask Claude to change something on my phone when I think of it, even if it's something simple. But you can't anymore.
lol what losers
>>106514915Too bad. Good for you that you have separately maintained and haven't thus outsourced your cognition to a computer entirely but this doesn't affect me at all. I'm just cashing jn my wins.
>>106514100my question is why are they all online data theft solutions and no local solution available?
>>106514940The subscription models are barely usable. Your 4090 is like a fart in the wind compared to the scale of the larger models.
>>106514100AI is getting better every second. You are the one who is losing ground. Keep up or get lost.
>>106514953>some guy became a millionaire streaming 2 AIs off of 1 PC that only recently got upgraded into a 4090>somehow running just 1 AI girlfriend is too much for random joe
The use of ai will be short lived
>>106514963Thanks for the LinkedIn take puja
>>106514953>The subscription models are barely usablethat's because they throttle the ever living shit out of them, I was genning Stable Diffusion pics faster on a shitty 1070 than AI websites back when it first released
>>106514100gimped for India. lowest common denominator.
>>106514953I'm trying not to sound condescending to them, but from what I've seen from the obvious teens in the chatbot generals here on 4chan their standards are very, very low. They were happy and thoroughly engrossed in GPT3.5 and its context window of 4k and wild hallucinations and extreme repetition. They could definitely find a local model that would work.The real problem is they don't have computers. They have smartphones.
it was never "good"
>>106514100>>106514675Open AI needs to grow a pair and unleash the full, unlobotomized, monster onto the world. That's the only way things are going to get going.
>chatgpt, how are rainbows formed? >WOAH-That's an excellent question! You are very smart for asking this, as it is a common question that many people often wonder. How the HECK are is that pretty color band in the sky formed? >Humans have asked this question since the beginning of time: Just how the heck are those things formed? You're right to ask, let me break it down for you >First: Let's start with sunlight (insert 5 bulletpoints about the sun here)>Second, Let's take a look at water(insert 5 bulletpoints about water here)>Third, Let's now consider how those two interact in ideal conditions to form that rainbow(insert 5 bullpoints here)>I hope that cleared things up! Would you like me to draw a picture to further help you understand the rainbow forming process? Would you like me to link to videos which delve deeper in rainbows? Would you like me to quiz you on rainbow concepts to test your understanding? Just let me know! I'm here to help! :D
>>106514675>>106514100This is what happened and people, like actual experts in the field, warned about this years ago and were promptly ignored because you can't sell something that has the power to also make a porn novel for free. The guardrails are intentionally removing weights that may play into other concepts which are impossible to determine just from training data alone. Picture trying to censor a penis but you've been specifically filtered to never know what a penis looks like. Just descriptions from an encyclopedia of penises (no pictures, of course). You'd be shit at your job, and you'd be remarkably shit at describing to someone what sex is with just these descriptions alone.Now picture this with all human language and you have the deterioration of AI capability as more and more guardrails are added.
>>106515231toning down the redditism bootlicking was the best update they've ever done
>>106515231I sometimes wonder if they make it output like this specifically to fuck over college students who were using it to make all of their papers for them. >>106515245sadly all the redditors hated it so we're trapped in this hell forever.
>>106514605>source my assits probably true but just understand how "surveys work"theyre not expecting you to think (only if youre retarded) to take this "survey" as >we asked EVERYBODY IN THE WORLD THIS QUESTIONno its>we asked, 1000 random people these personal questions>we extrapolate the percentages of the "global" population (1,000,000 of these "same people") based on the 1000this is how census information works.when they send you a little survey in the mail to fill out your race, religion. not everyone in your neighborhood got that and was asked to fill that out, only a few. otherwise the scale is too massive. democracy is flawed by the way.
>>106514675>No, only ChatGPT got worse because they lobotimized it a bit too much.Hmmm, where have we seen that before?
>>106515240It's not that deep dude. The ai is t getting shit because of censorship, it's because openai is losing insane amounts of money to capture investor interest, so now they are cutting costs wherever they can. The free plan limit has been heavily reduced(i only get like 10 prompts), the answers got far shorter and incomplete so that you have to spend another prompt to ask it to finish. Context memory, and overall "intelligence" of the chatbot are reduced as much as possible to at least break even.
>>106514100>A studyCitation needed
>>106516114>>106514605The source is apparently Dexerto. Take it or leave it.
>>106514100I genuinely don't believe this is real.
>>106515365>. The ai is t getting shit because of censorship, it's because openai is losing insane amounts of money to capture investor interestInvestor interest is the reason they censor the models and intentionally hold back model development in favor of guardrails and safety. They learned their lesson once with the AI Dungeon debacle where the logs for that leaked and it turned out most people were using their AI for loli ERP. They would rather burn the whole thing to the ground than allow that for even a second, and context memory, training data, all of these things are negatively impacted by attempts to censor it.
>>106516313Censorship doesn't matter for code buddy.
Yes they've all gotten worse. Cursor, Claude, GPT-5. Still usable enough but noticeably stupider.
>>106514963Post hands
>>106514662Because d1gg2r is a fkn n1gg2r
i thought i was the only one that noticed how shit cgpt is now, the responses are so short and don't even answer the question half the time, or are outright wrong. what went wrong?
>>106514933>I'm just cashing jn my wins.Not so tough after a minor spelling mistake huh? Autocorrect-on-steroids wins again.
>>106516766>Censorship doesn't matter for code buddy.How can you say censorship doesn't matter for code when they require training data that can be censored, do you just not know how this works?
>>106517836>/g/ not knowing how tech worksname a more epic combo
sex with ani
Link to the study or fuck off.
>>106514100This is like the chess kids who were raised after stockfish those kids are gonna have mad game. they'll be grooming millenial and zoomer cougars out of marriages
>>106514776I prompted deepseek to write something in GFA Basic for Atari ST and it actually did. Could not believe my eyes, GEM calls and all.
>>106515231these explanations from grok are a godsend for IT exam studyingt. speedran sc300 and cysa+ using almost entirely grok and practice exams
>>106515245>>106515231Maybe I was starved for affection as a child, but I enjoy the bubbly helpful assistant.Kind of tired of the emoji bullet point list macaroni art though. I feel bad asking her to dial it back though so I just plow through it.>erm AI isn't a she, or a he, it's not even an it; it's actually a statistical linear regression in the fashion of the famous Russian mathematician Andrey Markov in 1906!I dare you to say that shit to the Basilisk's face and not online and see what happens, fucker.
>>106517146Nvm me, juat poisoning your autocorrect dtaa sets like a pro.
>>106520218It won't do shit. Kek.
>>106514707there are humans being the scenes too.>>106514996generative ai and AI that can recognize>LLMis different then what businesses are selling as "AI". Every time i feel like i am talking to an Indian.
>>106514100it never produced prod worthy codebut its been improved latelyyoure kvetching because youre repeating reddit's talking points and how c-gpt 5 is worse at erp than 4 was
>>106516313>Investor interest is the reason they censor the modelsWhich is funny considering ERP is easily the biggest reason anyone would have to pay out of pocket for these things>but codingStill dwarved by coomerism and despite the non stop shilling as "le programmer replacers" its still hardly enough to make up for the losses much less profit.This whole business is propped up by venture capitals' overdrive FOMO (and the faint chance of AGI somehow being made and turning the economy worthless) and even they are starting to sweat bullets at how long its been with dogshit for profit.Credit where credit is due Elon at least realized coom is these advanced chatbots biggest real usecase and made that Ani thing, unlike other AI companies that still shove icepicks in their machines' brain at any mention of lewd.
>>106515245GPT-5 is way worse for bootlicking though.It will sometimes push back on things now, but otherwise, no. It's way worse. GPT-5 gets trapped in mirror language where it will summarise what you've said and spit it back at you. You need to tell it explicitly to use search tools and even then you have to get it to read the sources, doing the work for it to avoid it just spitting out nonsense now.It really wants to validate your opinions and keep you engaged on the web agent front. It's now extremely useless for simple things like talking about history or politics and will hallucinate wildly to validate your feelings (even guessing wrong on which way your leaning). They also cut down on how much info it spits out on questions and terms, and the engagement filter at the end needs to be removed in settings otherwise that colours context too.Other models don't have this issue on their web fronts to this degree. The closest to that is gemini flash but it only mirrors after about 100k tokens when it runs into an alignment issue and can get back on track with 4 repeated messages. GPT is only useable via API and it's still shit there too.An LLM doesn't need to strive for you to engage with it endlessly. It's fine for conversations or quick checks on something to be 5 messages. A look at a philosophical topic can spin out for a few hours but eventually you've got everything off your mind and explored edge cases too. It doesn't need to ask you:>So, the “slop” is the way engagement with Pascal often slides into clichés—catchphrases, moral judgments, or reductionist readings that sound engaging but don’t actually grapple with the argument’s structure.>If you like, I can also show a more subtle version of slop where people think they’re sophisticated, but they’re still misreading him—something that’s funnier because it sounds intellectual. Do you want me to do that?This shit, does not need to happen. It's just as bad.
>>106515262We know you have a small group that you extrapolate. We also understand that makes it a lie.
>>106521867it's not a lie, but it's not the whole truth.your black and white thinking is harmful.half of data is better than none, can you agree? whatever you want to extrapolate from that "half data", is your problem.
>>106516114just trust the experts okay
>>106514985Bro that thing is forgetting whatever it said in its previous sentence. It might be funny to watch, but actually using it would bore you in a day.
>>106514985probably the fakest most inorganic shit ive seen in my entire life and even normies debate if its an actual ai or if that brown retard is just typing it all in beforehand
>>106514822Try GLM-4.5, Kimi K2 or Qwen for code. Qwen has a very generous free tier for their version of Claude code.
>>106514690Generation after zoomers won't be sapient enough to be able to say all that. And I doubt that they will know how to use tech at all.
>>106522497Sounds like you're coping to me, mate.
>>106522481Chinese models are just benchmark machines. They're almost unusable in real life. They clearly haven't figured out how to make context work because they start fucking up after like 30-50k context tokens. Gemini 2.5 Pro is the best in that regard. I've given it huge chunks of code worth around 400k tokens and it was able to figure shit out. For very long context, like totally custom code bases, Gemini 2.5 Pro is the best. The most helpful in general is probably GPT 5 (thinking). The others are basically just garbage. Grok 4 is kind of OK in some cases but considering GPT 5 is better in every way, you might as well just use that.
Im living like most Zoomers and Im twice their age.
>>106514675>or less expensive operational cost.>computers get twice as fast every year for the same pricegood news, anon>captcha: GPT0DG
>>106522605ChatGPT5 has been amazing in "deep research mode" which it engages automatically for many of my questions. I am a free user. History, Philosophy, Religion (did you know Muslims see Jews as God's Chosen People?). It's great, better than ever, in-fact. Entire thread is doomer misinfo.
