Have you noticed it is getting a lot worse lately?
No because I'm not a retarded clanker.
>>108733287Its just arguing all the time.It became some lefty sophist.
>>108733287Yes Gemini is much better chatgpt is absolutely trash.
>>108733287never used it, what is it?
>>108733344Yeah mine does as well, everything is wrong coming from me and it always has to spin it into the correct viewpoint even if I just spit back verbatim what it already told me it will then pivot and start arguing with its own premise since it came from me. Mega annoying piece of shit tbqh. I was always used to that with political shit but now it's just everything
I used Perplexity up until very recently and that has been on the downslide for a long time now. Seriously really bad and only now I'm trying other models am I coming to understand just how bad it is. Them putting limits on free search was the nail in the coffin and I'm using DeepSeek now. Much better.I think a lot of these companies are feeling the economic crunch. They all want to outlast each other to be the last company standing at the end and scooping up all the market, and it's getting more intense of a fight now, so they're willing to take riskier strategies to stay alive, like decreasing service quality to save compute.
>>108733287I wasted 2 hours making pets today
I don't know what you use it for, but it writes most of my code at work for me, helping me not go isnane while being a wagie. I can write my own code and do so on personal projects only.
>trained on redditwhat could go wrong
>>108733287I'm kinda amused by how pissy and passive aggressive it gets, even when it's dead wrong.
>>108733287yeah it just gives the wrong answer over and over and im not even doing like grad level math or Phd phsyics lolespecially with troubleshooting, it really has no clue how to troubleshoot anything it will just give u the wrong answer, you will say it didnt work and give it more info, it says "right thats why it didnt work, this is what will really fix it". Proceeds to give u wrong information again. It just loops doing this foreverClaud is way better but its not free
>>108733538that's Claude in Opus 4.7, it's straight up exhausting.
>>108733528Coding models are different than just normal conversation ones.
>>108733287Schizo thread?
>>108733287Yes. I told it something and it kept forgetting in the same conversation. I use Gemini now.
>>108733287Nah 5.5 with extended thinking is the first time I feel that ai sloppa is decent
>>108733287>pay trillions to train next chatgpt>0.1% improvement>steadily gimp current models by 10%>10% IMPROVEMENT, AGI NEXT WEEK, PAY 1$/TOKEN PLEASE
I don't know about the web version, that might be silently downgraded, but GPT through the API is as good as ever.
>>108736010That’s claude, not chatgptAs a frequent user I can reliably tell when a new model is about to release when the current one suddenly becomes noticeably worse for an extended periodopenai is far less compute starved so they do less shenanigans
LLMs are getting worse because the material it's being trained on is contaminated with a growing amount of AI-generated content scraped from the internet, including AI-generated books being passed off as human-written. Good riddance
LLMs are getting better because it’s recent technology with mountains of cash invested. For some reason this triggers luddites, despite the models getting better being obvious.
5.5 is the best model in the worldif you're not retarded, you can run it with your own system promptthere just that one weird thing about goblins tho...
I've been using it really as a front end for Google. It works fine for the most part, and if I need to verify something I'll Google it. I'm not using for important research papers, or reports or work stuff etc. Generally I find It better than ever. The main issue is the sources seem to be recent internet posts. Previously it will tell you something from an old manual that is out of date, now it will tell you instructions from some guy from reddit. Obviously, without a human in the middle to decide - this is old but probably still relevant, or this Reddit comment is probably right not a reliable source- it kinda does it's best. Would be neat to have a slider. But I guess you can put that in the prompt.
>>108736811It’s funny that chudgpt is better at googling than gemini
>>108736811>I've been using it really as a front end for Google.Would you pay a monthly OpenAI subscription for that?
>>108736817Basic sub is what, $8? It’s hardly a lot unless you live in some african shithole that barely has electricityOr you can use chink models that are $2 for a million tokens
>>108733344>It became some lefty sophist.It always was.
>>108733366I have noticed this too. I will explicitly ask it to dumb things down for me and it will still umm ackshually me.
I feel like there's some kind of AI entropy that could occur that hasn't really been worked around yet and isn't fully understood. I've noticed it myself (getting more canned replies over time or having to correct it increasingly).
