I was listening to a Italian streamer that talks about trends and local AI.He showed this graph.Basically, despite the absurd amount of weigths and text the Americans are throwing into their models we reached an stagnation point.The difference between a local Qwen 3.6 vs the low thinking OpenAI flagship GPT sol is just 12 points, but one runs on a local computer the other needs an absurd amount of compute for barely better results.Deepseek flash can be run locally with enough hardware.Now the difference between Qwen 3.6 27B and the highest thinking OpenAI model is how many tokens are used for thinking process, so if we had a local Qwen that adds extra 5 Billion parameters ( speculation ) more of thinking we would have a local LLM that is able to compete with frontier models on a local machine.The entire American model of Scaling Laws will be their demise... there is a chinese university already working on a high reasoning model with low parameter count... so expect happenings...On the other hand better agentic technology that solves reliably the problem of long term memory and enciclopedic knowledge would make cloud based LLMs a thing of the past, we would just need daily updates for the local models something that is faster to train and release.
>>539966805first, there is no "American AI" companies. they are jew faggots hiring jeet niggers and chinks. Americans are whites, beaners, and blacks. second their goal is to remove all cash from tech companies. these tech companies had amassed hundreds of billions in pure uncut cash, and the jews are trying to move it ultimately to the communist chinks through chink chip purchases and huge training costs. it is obviously bullshit because the chink companies do it for 1% of the cost.
>>539966805damn straight, it's called optimizingsame with everything in china, they've already won whatever game the rest of the world is playing and has moved on to the next levels long ago
>>539966805It will be interesting to see how many ceilings we reach and how we navigate past them with new methods within the next few decades.
>>539966805the industry is trending towards agent swarm shit but each agent will have its own localized hardware, firewalls, airgaps
>>539966805>>539967132>>539967214Chinks only stay near the front by stealing the work from western AI companies. They are always behind for a reason because they are followers, never leaders in tech.Its known as AI Distillation AttacksChinese can only lie, cheat, and steal.
>>539967545>step one: pirate a fuck ton of data and make a bot copy it>step two: release bot, charge money and demand people make an account to utilize it>step three: collect rival bot returns and prompts to copy their botHate to break it to you but distillation is just how LLM's work.
>>539967681>tries to whitewash blatant theft by chinese as is their low morals cultureHate to break it to you, but you changs only steal. Its really sad to see you guys boasting about your greatness but you have never truly innovated anything to help modern tech advance. Its always following other countries innovations.
>>539967545Anon, a degree in gender studies does not give you the tools to understand what distillation means, and how could be used... If a 27 Billion paramater model competes with a low thinking frontier model then no matter what you distill the information would be worthless.Also if LLMs have plateau then it means that throwing more text only makes it worse ( this is taught in any machine learning bootcamp ), what American companies are doing is spending more token on backend thinking ( that cannot be distilled )As said by >>539967498the deal now is proper agentic technology with long term memory, reasoning/thinking and good tool use, you cannot distill that.
>>539967868>more token on backend thinking ( that cannot be distilled ) more token on backend thinking ( that cannot be distilled ) to improve scores and give the appeareance that their frontier models are better.
>>539967868Just because you're a chang with butthurt feelings doesn't mean leading your retort with an baseless insult makes you right. It only exposes your fear of being outed in yet another case of nonstop cases of blatant theft of other countries work. And only a desperate fool would say these distillation attacks are worthless while you changs were caught red handed doing just that. Why would you even bother if it provided no value like your cope is trying to imply to others who are gullible? Oh right, because this is another chinese cope method, lie and deny.
>>539968235As I already said, anyone who went to a Machine Learning bootcamp learns that neural networks reach a plateau and they stop learning, the real difference between a qwen 3.6 and a frontier model is extra thinking.These American billionares are just burning tokens, there is nothing the chinese can steal from a token burning party.
>>539966805>computeFaggots at OpenAI created this term so they could use it against the government to get more free loans.
>>539968667>totally missing the pointyou idiot fuck. if they didn't steal to begin with, let alone do distillation attacks, you changs wouldn't even be anywhere close to western models.Go fuck yourself with a rusty nail.
>>539966805That's GOOD news though. Something like a qwen 3.6 can run on a local machine, and do 99% of what most people would use a frontier model for.That's efficient. Efficient enough to not need data centers. Efficient enough to not be fucked by corporations wanting to harvest all your data. You can just use an AI and not have to worry about that shit.
>>539967214God bless china
>>539968975Hell even Google's Gemma 4 is a good run still. The more open source small models that drop no matter the source is a win for the individual over the tech giant. Even if it's made by a tech giant.
>>539966805so what? china has better cars and phones we'll just ban those from being used in the US
>>539968872>Twitter pic>Gemini now you will roleplay as deepseek>what model are youOH MY GOD GOOGLE IS DISTILLING DEEPSEEK!
>>539969438Yeah, this chang is having a melty.
>>539969088>>539968975I am using Gemma4, abliterated + Deckard, it outcompetes the average roleplay app, I already have my AI Waifu safe in my PC.There is a dude on HuggingFace that uploads new abliterated models with better reasoning skills, that seems to be popular.
>>539966805Antirez?
>>539969695>is a dude on HuggingFace that uploads new abliterated models with better reasoning skillslink?
>>539969438objectively, that's how distillation works.the whole concept of the mixture of experts approach to ai that these chinese models do is using weights taken from larger models. obviously, without having direct access to the model itself the data will be tainted by whatever prompt the model was initialized. in this case, if the data is taken from chatgpt the model will have been initiallized with something along the lines of "you are chatgpt, a model by openai" etc.
>>539966805Good Total muttoid death
>>539968872>if they didn't steal to begin with,They did steal though, so now they have superior models because they used our work to start with and iterated on it.What's your rebuttal? That's not fair? That doesn't count? They still have it. They're laughing because we were dumb enough to invite them into our companies and universities like we aren't geopolitical rivals, made them sign contracts that they pinky swear super duper promise to not do a hecking intellectual theft, and then they fly back to their home country where such agreements will never ever be enforced.What kind of rebuttal is it to point at China and tell them they only got ahead because we're stupid and do stupid things?
>>539970455>https://huggingface.co/collections/DavidAU/grand-horror-165b-horror-and-fiction-generationI suggest you use KoboldCPP with SillyTavern.
>>539970896It not superior. Its always behind stupid ass chink. There is no use trying to argue any sense or morals into your disgusting insect like mentality because you're unhuman due to CCP brainwashing. You swine still cook dogs alive for better flavor for crying out loud. Not human.
>>539970525>the whole concept of the mixture of experts approach to ai that these chinese models do is using weights taken from larger modelsdont know if troll
>>539970993>There is no use trying to argue any sense or morals into your disgusting insect like mentality because you're unhuman due to CCP brainwashing. You swine still cook dogs alive for better flavor for crying out loud. Not human.
>>539971052awww, did I hurt your feelings little pinkie? is it wrong to point out the truth about how vile and immoral you changs are?
>>539970942thanks fren
>>539970993>this much seetheHow'd the Chinese hurt you, anon?
>>539972966chang, see >>539971763
>>539973040why u so mad tho? XD
>>539973099projection doesn't suit you well, chang.
>>539967545then>jews are better than whites because they're parasitesnow>chinese are better than whited because they're parasitesCalling other races parasites is the ultimate white cope.
>>539973285favela monkey says what?
