Happening!Gemini is suddenly retarded!
Always has beenRemember I got $300 worth of tokens from them and it was barely worth using
>>109641996Nigga it was better than chat gipptyNow it's as good as google bard
>>109642003Your answers are mostly spot on!
>>109641996/thread
So what's the best general use LLM now?Don't say "hurr durr your own just buy a $10k rig and run it locally"
>>109641973Just use Gemma 4 31b for language/translation stuffIt's open source so it can't be stealth nerfed.
>>109641973microsoft and chatgpt are literally turning their ai engines into fucking retards. in my experieince, over the last month gemini has become impossible to use without endless errors and taking ~20 minutes to identify a problem in 10 lines of code.
I was having a technical conversation with Gemini last night when all of a sudden it went full retard and explained to me (I shit you not) why the sky is blue. When I repeated my question, it then proceeded to give me a comparison between my city and London. I wish I had taken a screenshot, but I'd already turned 360 degrees and walked away.
As always I blame the pajeets.
>>109641973Jesus?
>>109642185>Here is the complete code:>syntax error or some shit like that>Oops I forgot :^), here is the real code:if I delivered a broken product to fuck with the customer I'd be fired but not a clanker
>>109642164>nowAll western """frontier""" models (2026) are benchmaxxed, taught to the test, and waste your money sealioning and kvetching instead of doing the task.
>>109642335It has kindly done you the needful, sir
>>109642164They just raped the gpt $20 plan, but it still uses less tokens than Claude will generally & is less combative so I’d try that out.
It was fine for months, even psyoppable for very good ERP. Now it's braindead retarded, fails constantly, and reeks of e-coli more than usual. Some jeet definitely perma-fucked it in the backend. Oh well, not my problem either since i never paid to use it. This just gives me the kick in the ass i needed to set this up locally.
>>109642164hurr durr you don't even need an over $1k rig to run deepseek, just don't be a retard or lazy (like me who only got ERP going and never the web searching)
>>109641973et toi, Anon?
>>109641973>uses Fr*nch>immediately becomes retardedIt's like pottery
>>109642512French is the best european language if you ignore the niggerified pronunciation
>>109642164Whatever is free on opencode/openrouter this week.
>>109642164Hurr durr buy a $20k rig and run it locally
>>109642535>France>best anything besides importing infected niggers and letting them rape native French girls
>>109641973>Trying to get the bot to translate into thirdie languages like French and anything else from the savage lands of Not-AmericaIts not retarded, its just correctly identifying that its not really a language.
>>109642164literal trash unless you're running uncensored locally. its that simple.nerfing the NN fucks up the accuracy, they know this and nerfed anyway. its broken, stop paying for keys.
>>109642512>>109644273>>109644294As a fluent speaker of both but not a native speaker of either, I find French infinitely more logical, precise, coherent and consistent that English (which has been heavily influenced by French anyway, due to French then-master-race invading the British inbreeds).English is basically JavaScript of natural languages: an abomination resulting from mixing two different paradigms (Lisp/C, Germanic/Romance) with layers upon layers of ad-hoc updates, resulting in an inconsistent fucking mess that still got popular for some reason.
>>109642638this. models are commodity
>>109641973>suddenly
>>109642003it was better than chatgpt for exactly a month i have a free pro subscription for a year and i just don't use it anymore besides image gen
>>109642164I use the new Deepseek V4 Flash for almost everything now, previously it was the old Deepseek V4 Flash or Minimax M2.something. I sometimes go to Claude Opus for tasks that are too difficult for the cheap models, but that's getting increasingly rare to be honest.I use them through OpenRouter since I can just drop the key into all of the applications like OpenCode, SillyTavern or self made stuff. And Deepseek is so cheap that I still end up way below a $20 a month subscription or something like that. Or maybe I'm just not 1000x tokenmaxxing enoughQwen 3.8 27B on my gaming PC is also shockingly good. Can't say I use it as much as cloud models but I've used it more successfully than all the previous local models so far.
