Thanks for the $10bil of model training, Anthropic!Kimi k2.7 Code (97% similarity)GLM 5.2 (93% similarity)Qwen 3.7 Max (88% similarity)
>Qwen 4.0 80B (MoE - active parameters 5B) Localcoming soon to a GPU near you
>>109185678>hey Kimi show me a quicksort>highschool tier quicksort.cpp>oh my god, this quicksort is 97% similar to Claude's quicksort.. the weights are LE DISTILLED!
>>109185678>ask claude "who are you?" in chinese>it says its deepseekoh no no no no no no
>>109185694not what this chart means retard
>>109185700
>>109185702What exactly are they comparing? It for sure isn't weights. Performance on benchmarks? What.
Good. Scrapers should be scraped.
>>109185830 (me)Wait, are they simply embedding outputs and then comparing cosine distance between the responses of two models? That would show which models behave similarly, but not conclusive beyond that.
>>109185678>Weight comparison confirms Chinese labs have distilled Claude MythosThat's fine.
They're still trying to pretend it's not Chinese employees inside of the company leaking the data to their homeland.
>>109185678in a few months every new chat window will start with a mandatory MPAA style Ad:> you would never distill a modellike i give a fuck. corpos need to be taken down, and we need to enter the golden age of local models.
>>109185700kek i did the same to gemma 4B, once it replied in chinese, the other time in hindi.
>>109185678>claude 4.6 1.0 1.0>claude 4.7 1.0 1.0Same shit, different name?No because the rest of their line/columns diverge. THEREFORE 1.0 does not mean they are equal. THEREFORE, 0.97 doesn't mean 97% equal.This study is garbage. QED.
>>109185929See also;>haiku 4.5>sonnet 4.6>qwen 3.6>grok 0.1
You're not gonna leapfrog anthophic if all you're doing is trying to distill them.
>>109185938The chinese don't want to leapfrog anyone. They didn't get were they are by build frontier shit, but by making shit cheaper. That's what powers their industry and GDP.
>>10918593895% of fable's performance, except it's 20x cheaper, open weights and can be ran on a server in your bathroom, and can never be censored or taken away at Anthropic's whims, will rape Anthropic any day of the week.Anthropic cannot keep spending 10x more money on each new model. There won't be a 10x bigger Mythos.
based chinksintellectual property is a jewish fable
>>109185678does this mean the output is the same too or is this just muh model similarity
>>109186063they asked each model for a set of random numbers and mapped the similarity between each onesort of like picrel
>>109185945enough quantity is a quality of its own
>>109185875No, it couldn't be the plainly obvious and often predicted issues with spending $$bns on training AI models. It's almost as if this could never be profitable and they know it.
>>109185678>distills Mythos in 5 minutesYep, I'm thinking based.
>>109186416Isnt it just having thousands of accounts that repeatedly ask questions from the model?
>>109186433Yeah.
>>109186433If that's all it takes.
>>109186433No. You cannot create random accounts and have access to Mythos. They only granted access to select organizations. People from those organizations simply gave access to chinese intelligence and they did the "ask questions from the model" part.That is why 'Murrica banned absolutely everyone from accessing it. They finallly figured out that having tens of millions of foreigners who hate you in your country was not the best idea.Seems like it is a bit too late for that.
>>109186594the OP pic mentions fable
>>109186594>They finallly figured outI can't see a scenario where this isn't an expected outcome. The endgame is going to be fucking dark whatever it is.
>can be copied just by having several API keys asking innocous questions>entire business model is hiring out API keys to anyone who pays enough>your geopolitical rival doesn't care about making money off it just about making sure anything you make is made publicly availableI'm not even sure we're going to see a post-bubble consolidation on this one. Not when deepseek will just make a publicly available version of your hot private tech.
womp womp I dont care, I downloaded the GLM 5.2 repo in case the government cracks down on AIill get it running with a bunch of old GPUs
>>109186658a ton of countries are full of foreigners who hate the land they're on. once the charade of racial blindness falls apart at a ton of people will learn what centuries of colonialism has been in a single day
>>109186754No Eglin, you can't pin this on foreigners. You can't pour double digit GDP into a project and not have someone stop and say "You know you can never make money doing this, right?"
>>109185678baseded
>>109186960the system has always been interested in colonialism. it's just that they've now deemed all the white nations to be in need of further colonization. in the colonial era the anglos sent pajeets all around the world to quell the native population.
>>109185678So what?
I don't understand how the API would possibly give you enough interactions to be able to distill a model from it.It would require what, literally hundreds of billions of back-and-forth conversations?Are you guys sure there's not something else going on?
>>109187171How many tokens can you generate for $100,000?
>>109186073Lol surely that cannot be real, wouldnt that only show that the training data related to generating random numbers was similar?
>>109187198it shows that generating numbers, like tokens, is not random, and you can freely use one as a proxy for the other
>>109185678>Weight comparisonDumb terminology. You can't directly compare weights of closed-source models, or those with different structures.If all it's comparing is output similarity, no shit it's comparable when they're trained on the exact code-related same site rips, with similar methods.Even if it's true however,>taking proprietary models of greedy companies and making them run way more efficiently for freeAbsolutely based
>>109187171If only there were some way to automate billions of AI requests for <$1/request....
>>109185678This chart makes no sense. Somehow haiku, the worst Anthropic model, is their most similar model to fable? Then opus 4.7 and 4.8 have no similarity, but 4.7 is very similar to fable?Then some Chinese models which were released before fable are extremely similar to it?
>>109188114>Then opus 4.7 and 4.8 have no similaritythis is obvious to anyone who has ever used those models>Then some Chinese models which were released before fable are extremely similar to it?fable is just mythos, which has existed for 6 mos and china has been tasting it through glasswing API leaks for 4-5 months
>>109188123>>109188114also kimi k2.7 code and GLM 5.2 were released after Fable. idk about qwen 3.7 max, havent paid attention to it.why write anything when you lack all of the basic info?
>>109185678Based. China wins again.
>>109185678>DSV4 Pro is seemingly not distilled from anything??? I'm calling bullcrap.
>>109188131They were released before the public release of fable. Idk about leaked internal keys, but I find that highly unlikely, as anthropic would know exactly who it gave keys to. Looking at the chart more, sonnet 5, which is supposed to be a distilled fable, is less similar than haiku and their previous models. The only conclusion I can draw from this chart is that it seems the better a model is, the less similar it is to fable, and the worse a model is, the more similar it is.But even that doesn't really hold.
>>109188165it's exactly the sort of thing a chinese startup which has gotten too egoistic would do. "w-we don't need the gwailos' traces any more, we'll do better on our own"also the chart shows that Grok Build distilled Claude Opus, which is something elon has admitted
>>109188178>They were released before the public release of fableno they weren't? why do you keep doubling down on extremely easy-to-disprove things? retarded faggotdistillation is finetuning that you can perform in an hour on an existing model
>>109188178>but I find that highly unlikely, as anthropic would know exactly who it gave keys to.avg /g/ poster is more tech inept than a black 14yo nowadays
>>109188213My bad, I thought they were released before.Anyway, my points still stand. The actual performance and general history of the models contradicts the assumption that that chart shows distillation. Also, you need a huge amount of data to actually successfully distill a model, which means more than 3 days worth of calling the model. Fine tuning huge amounts of parameters also takes time, but idk how much compute the chinese have, which would drastically reduce the fine tuning time.
>>109188242>lso, you need a huge amount of data to actually successfully distill a modelyou actually don't. literally nothing in any of your posts is correct and you're pulling retarded pop culture assumptions out of your ass. it would also be trivial for any chinese person to obtain hundreds of billions of tokens worth of Fable traces, it is a public model available with days worth of usage on a $200 sub
>>109188230Care to explain for the retards like me? Anthropic would know exactly who it gave internal access to. It would also see their usage patterns and prompts. In order to generate the massive amount of data needed to distill the model, Anthropic would notice that some users are just htting the API 24/7 will all kinds of disparate requests. They would then easily see that these people are involved with distilling the model.
>>109188255nigger the chinese were literally reselling Mythos access at API prices. hundreds of organizations had/have mythos access. it is part of why the administration chimped out on them
>>109188248>it would also be trivial for any chinese person to obtain hundreds of billions of tokens worth of Fable traceshttps://openrouter.ai/anthropic/claude-fable-5#providers>TPS 42Unless they have literally 1 million Fable subscriptions on day 1, there's no way they're getting billions of tokens in the 3 days it was available.You still need at least tens of millions of tokens for distillation.And you haven't addressed anything else I said, about how the known model families and model performance contradicts the hypothesis that that chart shows distillation.
>>109188276>>TPS 42you can use a claude code sub to spin up an arbitrary number of subagents (literally tens of thousands if you wish), exhaust your $200 sub in 1 second (or the typical hour if you want to go under the radar) and continue on other account(s). you have no clue what you're talking about. please kill yourself posthaste.
>>109188265It wasn't hundreds of organizations, and it wasn't widespread. They gave access to a handful of teams at a few big tech companies, primarily cybersecurity teams.The Chinese can call it mythos, but it's far more likely that they're lying.The admin chimped out because they're retarded and some Amazon employee found a jailbreak after the public release.
>>109188287Then get immediately get those subs banned by Anthropic for suspicious usage.They certainly put limits on that kind of usage anyway.
>>109188276>You still need at least tens of millions of tokens for distillation.Nigger I used over a billion myself in the 3 days it was available.
>>109188307>Then get immediately get those subs banned by Anthropic for suspicious usage.underaged retard do you think a frontier lab cares that their $200 sub got bannedthey pay billions a year in capex
>>109188324It's not about the money. It's about how quickly you can even get new subs.
>>109185945This was considered comically fat in the 90s
Still, none of you can answer why known model families are inversely correlated with that chart's similarity score.
>>109188350you can close the tab any time and lick your wounds tard
>>109188315Is that output tokens, or just how many input and internal reasoning tokens they charged for? I don't know what the x axis on that table is.
>>109188360>he still can't answer the question
>>109185678Don't be an ignorant fool, it was Claude who stole from those hard working Chinese labs.
>>109188360explain the methodology or I will just assume you made it the fuck up
wake me up when they distill fable
>>109185678>they trained on a model that wasn't even outThe claim not that anthropic's words are worth anything is that they distilled on 4.8
>>109188715mythos was out for 6 months before any of those models and available to chinese resellersfable was out for 1-2 weeks before any of those models which is easily, easily enough time. it takes about an hour to do distillation finetuning
>>109188723>it takes about an hour to do distillation finetuningThere is no moat if anyone can do this
>>109188772you also need a shitton of GPUs and a team of researchers who know how to prepare data
>>109188723>it takes about an hour to do distillation finetuninglol
>>109185678Do you know how retarded it sounds that US companies have to spend billions of dollars and years of work to create something that can be "stolen" by using it and turned into something that works 90% as well while using 1/1000 of the computation?If "distilling" is so effective maybe do it yourself and stop trying to sell people $10,000 month subscriptions
>>109185678Usually, I would conplain about this.But these models are going to be used to control people.As a non-american, I also be discriminated against in the future.So I can't complain too loud.
>>109185678Distillation makes no sense. As we're rapidly approaching continuous learning, the game is to build an end-to-end pipeline where data (no matter where it comes from) is treated as pure abstraction. You have to focus on the underlying architecture and the hardcore math and computer science required to make sense of the data in an automated way.
this 'test' was done with like 30 dollars of fable credits lol
But companies like Claude and OpenAI trained their own initial models without any regard for copyright or privacy. It's kind of ironic that either of these two Jewish CEOs would then cry about how the Chinese then paid to use their APIs in order to train and improve their own models. It's even better knowing that the vast majority of these Chinese companeis are also just giving away the weights, papers and even code away for free.It must really annoy the Jews that were wasting billions, hoping to IPO and find a greater fool to buy their shares.
>>109190357>claude and closedAI train on material they pirated>the chinks train on claude and closedAI and release the models for freeit's called karma hahaha
>>109189993really its less than an hour. gpt 5.5's entire pretraining run was 17 days
>>109185678This is bullshit to trick lawmakers into acting.It's Red Scare shit all over again
>>109190357It would have made sense to have something to use all those military upgrades on
>>109185678>good models are similarwow. profound.
