Are there any local models now that are comparable with like top level claude code? And if so how much vram do you need?
Yes. There's a few.
>>109992510Depends on your definition of "local". Do you mean local as in not locked within a provider's wallet garden (OpenAI, Anthropic are the top examples), local as in "it can be downloaded", or local as in it can run on YOUR machine specifically? If the latter then that depends on what hardware you have at your disposal.
>>109992510Nope. There’s a reason 99% of these open math problems were solved using ChatGPT and Claude:https://vibemathed.com/When RSI goes full speed, the gap will only get worse. Local models are distilled from the newest Claude and GPT but these are distilled from an even more powerful internal model that no one else has access to.
>>109992530Talking about one that had to copy someone's homework as a bad example. >RSI PUHLEEEZ go back
>>109992530basically this>>109992510chink models are fucking trash OP, we are fucked.
>>109992530Are any of these supposed "solutions" verified by mathematicians? It just says it is verified by more AI lol>AI: I solved it!>AI: Let me check if my solution is correct.>AI: Yes, turns out it is correct!
>>109992530>When RSI goes full speed*clips your upper frontal tooth in half*
>>109992510Your brain runs locally.
>>109992530rsi is impossible because of taski's undefinability theorem.conversely, we're already in the rsi regime, where cheap clanker labour allows to pursue marginal optimisations, but models cannot self-modify into superintelligence.the representation in our universe has to be transacted to become a result - so intelligence is a function of how much computation you can do. ai will keep improving so long as more compute keeps being built, but we'll never get an AGI in your phone.which is good, slow takeoff gives us a chance to survive.
>>109992510Running something like Kimi K3 would require at least 1.6 TB VRAM.
let me ask a more specific question than OP, what is the best local model to run for programming/coding stuff, but not necessarily for full vibing workflow, on a RX 6800 (16 GiB VRAM), is it qwen?
>>109992699yes qwen on low quant, but it will be truly garbage compared to what $20 buys you
>>109992643My brain might work as good or better but it's not nearly as fast.
>>109992526How about something that can run on a single non enterprise card.
>>109992583No mathematician is smart enough to verify it.
>>109992718you guys are only on the $20 tier?
>>109992857I'm on free tier
All tyhe shills in this thread, most local models are vastly better than any openAI subsciption and pinting a local LLM at a RAG of language and task relevent samples will give you better results than claude but the real sting for all the faking it to mmaking it 'web developers' in this thread is that ANY LLM will generate fale and spurious data and have gaps in data (by design), what they create will always be dependably shit whether you use cloud or local.
>>109992849So it's all just AI hallucinations until actually proven.
>>109992510anybody saying yes is lyingand you can get infinite credits on claude by getting a subscription for $100 no way any local model would be close
>>109993087Prove it.
>>109992865nigga you're not coding shit lol
>>109993087>I'm a pragrammer now he thought. Claude make me an app that displays pasta recipies. This was his time, time for mhis genius to shine. He had't felt this good since his social secuity training day where he made a 'web site' using a wazzard. He always knew he was the next bill gates now he had the means to spam the google play store with pasta recipie apps he would proove it. Next a dating app for furries and an app that told you nwhat clothes to wear based on the weather. Claude began spitting out a thousand lines of crap, he could not understand it but who the fuck cared, he put on queens classic track 'don;t stop me now' and sat back. No one would call him a jeet anymore after this, not when he had his lambo. 83 cents from his youtube channels today. When he went back to his village rich they would regret the day they had pooped on his flip flops and told him to leave for masturbating in front of the women and children His time was now. All his best ideas for the software would be made, the poop reader, the hundreds of pasta recipie apps and of course the bobs and vagine candy crush clone.
>>109992510>comparable512gb vram>competitive1tb~1.5tb vramgood luck!
>>109993197If uyou think the results from claude subsciption are vastly different for mediocre C or C++ functions or python to that from a local model pointed at a RAG of samples I have a bridge in brooklyn to sell you. You get conned because you are stupid. The end game is a bit more profound, the sit you get with claude will regularly and dependibly broken, why? Because it's an LLM stupid. It can;t deal with an novel situation whatsoever, it has gaps in it's training data as it has been compressed and occasionally it will just make shit up or stick in irrelevent functions scrapped from some dimwits blog post on medium. But you won't care because in the end of the day you never gave a shit to begin with and never really had a clue what you were doing beyond making minor alterations to code samples and cut and paste from stack exchange anyway..
>>109993227>Because it's an LLM stupidcorrect, and the only improvements we get are by scaling hardware+data set, but they are improvements
>>109993197I'm sorry but there is literally no way claude is running anything like that on their backend.
>>109993227> I have a bridge in brooklyn to sell yougo back
