burner22

/

numbat

Like 407

Why are so many people reporting the repo? The strongest evidence it's a fake is derived from under 200 tensors out of hundreds of thousands

Its physically impossible for this to be real, the weights are literal noise and RNG (not to mention how the architecture itself is autistically unstable, with a strange distribution of heads and KV cache). It has been debunked on twitter and in this page's discussions (by people who went out of their way to analyze this shit).

Let it go son.

in which case, are you interested in buying a bridge?

Its physically impossible for this to be real, the weights are literal noise and RNG (not to mention how the architecture itself is autistically unstable, with a strange distribution of heads and KV cache). It has been debunked on twitter and in this page's discussions (by people who went out of their way to analyze this shit).

Let it go son.

Twitter people said it's "likely" fake. Not definitive. I think someone in HF already ran the weights with 0.35 t/s

Its physically impossible for this to be real, the weights are literal noise and RNG (not to mention how the architecture itself is autistically unstable, with a strange distribution of heads and KV cache). It has been debunked on twitter and in this page's discussions (by people who went out of their way to analyze this shit).

Let it go son.

All of those ai quotes go back to the under 200 tensor analysis. And ai models are initialized as random noise during pretraining. It makes sense if it was pulled out during pretraining, maybe their training algorithm is unique.

All of those ai quotes go back to the under 200 tensor analysis. And ai models are initialized as random noise during pretraining. It makes sense if it was pulled out during pretraining, maybe their training algorithm is unique.

So you're essentially saying that you still believe this is legit despite the fact that at least 1/7th of its layers were proven to be garbage? (the first 200 layers btw, if there was any meaningful training that actually went on they wouldn't be pure gibberish that CAN INDEED BE REPRODUCED WITH RNG, as proven by the countless people who actually probed this shit), And you're betting on the fact that ~500 layers or so haven't been explicitly dug through?

Its like seeing a massive container that claims to filled with gold, but is actually filled with metric-tons of horse shit. You have some people who dug 1/7th of the way just to check whether it really only has horse shit, and they're like "yup, this is definitely horse shit". Only for you to appear from out of nowhere, from the middle of the bushes like a goddamn little Vietcong guerrilla fighter, and yell at everyone: "WAIT!!! BUT HOW CAN YOU BE SURE??? YOU HAVEN'T DUG ALL THE WAY IN!!". This is insane.

All of those ai quotes go back to the under 200 tensor analysis. And ai models are initialized as random noise during pretraining. It makes sense if it was pulled out during pretraining, maybe their training algorithm is unique.

So you're essentially saying that you still believe this is legit despite the fact that at least 1/7th of its layers were proven to be garbage? (the first 200 layers btw, if there was any meaningful training that actually went on they wouldn't be pure gibberish that CAN INDEED BE REPRODUCED WITH RNG, as proven by the countless people who actually probed this shit), And you're betting on the fact that ~500 layers or so haven't been explicitly dug through?

Its like seeing a massive container that claims to filled with gold, but is actually filled with metric-tons of horse shit. You have some people who dug 1/7th of the way just to check whether it really only has horse shit, and they're like "yup, this is definitely horse shit". Only for you to appear from out of nowhere, from the middle of the bushes like a goddamn little Vietcong guerrilla fighter, and yell at everyone: "WAIT!!! BUT HOW CAN YOU BE SURE??? YOU HAVEN'T DUG ALL THE WAY IN!!". This is insane.

Why are you so obsessed with proving it's fake? Layers are not the same thing as tensors.

Why are you so obsessed with proving it's fake?

I literally just wrote 2 messages about it being fake (this one is the 3rd). Projection?

Layers are not the same thing as tensors.

Ik. The point is that people went out of their way to dig into the layers, and verified they're gibberish (reproducible RNG).. and your only argument is "well, what about the others you didn't verify??". This type of thinking could be used anywhere. "Flying unicorns can't exist. Our research here on Earth proves this.", "BUT WHAT ABOUT SATURN AND JUPITER??"

At this point you can believe whatever you want man (not going to debate this, I came here just for laughs). But when this clearly fake shit fades into irrelevance and nothing comes out of it, make sure the shame doesn't hit you too hard.

I don't even know what this guy's point is. If even a single tensor is shown to be derived from random noise, that pretty much debunks the whole model. You don't just train parts of the model during pretraining. If one part of the model is frozen to random noise, the backprop signal is diluted and fits to garbage. Also, I don't see any real benefit for a model creator to initialize from some weirdly specific randomization method other than to obfuscate that it is indeed noise, that's akin to shooting the mystery box three times then doing a twirl to get the Raygun every time in COD Zombies.

I don't even know what this guy's point is. If even a single tensor is shown to be derived from random noise, that pretty much debunks the whole model. You don't just train parts of the model during pretraining. If one part of the model is frozen to random noise, the backprop signal is diluted and fits to garbage. Also, I don't see any real benefit for a model creator to initialize from some weirdly specific randomization method other than to obfuscate that it is indeed noise, that's akin to shooting the mystery box three times then doing a twirl to get the Raygun every time in COD Zombies.

I don't know either why they'd (burner22) just upload fake weights like I don't understand the motivation behind it. Surely it takes their time too since generating random noise at the magnitude of terabytes. I know it's easy but why waste compute, electricity and time? And the effort to make it spit out stories if 99% of the weights are noise just doesn't make sense to me. Fake or not, this is weird and needs some more digging into, especially the guy behind the account.

All of those ai quotes go back to the under 200 tensor analysis. And ai models are initialized as random noise during pretraining. It makes sense if it was pulled out during pretraining, maybe their training algorithm is unique.

So you're essentially saying that you still believe this is legit despite the fact that at least 1/7th of its layers were proven to be garbage? (the first 200 layers btw, if there was any meaningful training that actually went on they wouldn't be pure gibberish that CAN INDEED BE REPRODUCED WITH RNG, as proven by the countless people who actually probed this shit), And you're betting on the fact that ~500 layers or so haven't been explicitly dug through?

Its like seeing a massive container that claims to filled with gold, but is actually filled with metric-tons of horse shit. You have some people who dug 1/7th of the way just to check whether it really only has horse shit, and they're like "yup, this is definitely horse shit". Only for you to appear from out of nowhere, from the middle of the bushes like a goddamn little Vietcong guerrilla fighter, and yell at everyone: "WAIT!!! BUT HOW CAN YOU BE SURE??? YOU HAVEN'T DUG ALL THE WAY IN!!". This is insane.

Why are you so obsessed with proving it's fake? Layers are not the same thing as tensors.

This had to be ragebait

I don't even know what this guy's point is. If even a single tensor is shown to be derived from random noise, that pretty much debunks the whole model. You don't just train parts of the model during pretraining. If one part of the model is frozen to random noise, the backprop signal is diluted and fits to garbage. Also, I don't see any real benefit for a model creator to initialize from some weirdly specific randomization method other than to obfuscate that it is indeed noise, that's akin to shooting the mystery box three times then doing a twirl to get the Raygun every time in COD Zombies.

I don't know either why they'd (burner22) just upload fake weights like I don't understand the motivation behind it. Surely it takes their time too since generating random noise at the magnitude of terabytes. I know it's easy but why waste compute, electricity and time? And the effort to make it spit out stories if 99% of the weights are noise just doesn't make sense to me. Fake or not, this is weird and needs some more digging into, especially the guy behind the account.

I wouldnt say effort is the biggest hurdle this burner22 person needed to overcome in order to put out a fake model like this- its resources. it is by no means a convincing fake, judging by the fact that it didnt take very long for people to start proving it to be, in fact, fake. The uploader had like 8TB or more of storage and a fair bit of time on hand. pretty easy trade for the amount of attention this has garnered. Every single person who has taken a moment to say "but what if its real, i believe its real" has done nothing but to validate burner22 in what they set out to achieve- which was to get attention and to trick people on the internet, probably for entertainment. Trolls have been around longer than the internet has and they aren't going anywhere.

The biggest piece of evidence you need to conclude that this is fake is the mere fact that it still exists. It would not be hard for Anthropic to determine if the weights were genuine or not, and in the case that they were genuine, the response would be to have them pulled. Them being on huggingface (of all places) over a week later is incredibly damning evidence that this is a complete nothingburger

This had to be ragebait

Are 2 simple messages ragebait? You must be very smart.

Soooo You're

Its physically impossible for this to be real, the weights are literal noise and RNG (not to mention how the architecture itself is autistically unstable, with a strange distribution of heads and KV cache). It has been debunked on twitter and in this page's discussions (by people who went out of their way to analyze this shit).

Let it go son.

All of those ai quotes go back to the under 200 tensor analysis. And ai models are initialized as random noise during pretraining. It makes sense if it was pulled out during pretraining, maybe their training algorithm is unique.

Oh yess,

• Some moron on earth has the time and compute to:
  A. Build a fake model for people like you to f-word to believe and defend
  B. Bake a Gemini 3.5 Flash-Lite Prompt (Don't ask me how i got it) and prompt it to generate that glorious readme
  C. To TROLL PEOPLE (did you not have bullies in school/college? or people that just want you to get angry on them AKA Trolls. ) For their entertainment and attention
• Why you shouldnt belive this crap stuff:
  THE EMBEDDING TABLES ARE FROM LFM2.5 25B A2B (correct me if my model name is wrong but im pretty sure this embedding table is from LFM or something in the <100B model range)
  AND ALL WEIGHTS AFFECT EaCH OTHER AND AS THE README U BELIEVE SAYS ITS A 1.8M steps checkpoint out of them 1.9M total IT SHOULD BE ABLE TO TALK WHY IS IT Outputting gibberish???
  AND I KNOW YOURE AN ALT ACCOUNT TO GET EVEN MORE ATTENTION THANK ME LATER

Why are you so obsessed with proving it's fake?

I literally just wrote 2 messages about it being fake (this one is the 3rd). Projection?

Layers are not the same thing as tensors.

Ik. The point is that people went out of their way to dig into the layers, and verified they're gibberish (reproducible RNG).. and your only argument is "well, what about the others you didn't verify??". This type of thinking could be used anywhere. "Flying unicorns can't exist. Our research here on Earth proves this.", "BUT WHAT ABOUT SATURN AND JUPITER??"

At this point you can believe whatever you want man (not going to debate this, I came here just for laughs). But when this clearly fake shit fades into irrelevance and nothing comes out of it, make sure the shame doesn't hit you too hard.

Obligatory "So... you're saying there's a chance??" (Dumb & Dumber) 😂

Guys just mass report the Creator of this Discussion (Ragebaiter) maybe he'll be banned?

Guys just mass report the Creator of this Discussion (Ragebaiter) maybe he'll be banned?

Nah don't do that. Dude just fell for some bologna, happens to the best of us.