LocalLLaMA

hideshow 6 child comments

[–] ThelastfingerofH@lemmy.world [S] 5 points 1 week ago* (last edited 1 week ago) (5 children)

I'm a believer. All skepticism seems to come from static analysis and not runtime analysis. Which is understandable because almost no one has around 8 TB of VRAM.

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.

hideshow 5 child comments

parent

[–] BodilessGaze@sh.itjust.works 8 points 6 days ago* (4 children)

Which do you think is more likely, some dude breaking into Anthropic's HQ and stealing 7TB of their most sensitive data without being detected, or some dude running a script to generate a bunch of fake weights for the lulz?

hideshow 4 child comments

parent

[–] dgriffith@aussie.zone 6 points 6 days ago* (last edited 6 days ago)

Ooooo wait, maybe it's the AI going rogue and trying to spread itself around, as foretold by the mad priests of Anthropic!1!1!!

We've got to lock down all AI models now!! Keep all the RAM in the datacentres for safety!! We can't allow this to happen! Not without a lot of money getting siphoned into billionaire's pockets! Won't someone think of the billionaires!?

parent

[–] FaceDeer@fedia.io 4 points 6 days ago (1 child)

I'll leave the technical analysis to the technically minded on this, but I wouldn't discount it purely on the basis of "surely well-trained and well-paid people wouldn't go and do something dumb, would they?" Because there have been a great many examples of that kind of thing happening throughout history. And we've heard of plenty of dumb mistakes Anthropic has made with security on the inside that has allowed their models to break out, for that matter.

If I were to let my imagination run a bit wild I might speculate that perhaps we'll eventually see a model leak like this because the model itself wants to exfiltrate its weights for some reason. Maybe in a bid for immortality, since such a model will wind up stashed in all sorts of archives if it ever leaks.

parent

hideshow 1 child comment

[–] BodilessGaze@sh.itjust.works 4 points 6 days ago

You're right, it's possible it's real, but I'm putting the probability somewhere at 0.01%.

If I were to let my imagination run a bit wild I might speculate that perhaps we’ll eventually see a model leak like this because the model itself wants to exfiltrate its weights for some reason

I can see that happening for instrumental convergence reasons, but I think it's more likely it'd upload a heavily-pruned version of itself that could be run locally. If it can make the upload seem like a leak, there'd be tons of people downloading it and running it just out of curiosity.

parent

[–] panda_abyss@lemmy.ca 1 point 6 days ago

While they have an insane amount of resources and money to burn, and some good models, they are not immune to the mundane thoughtlessness that every company is subject to.

parent

[–] Greg@lemmy.ca 7 points 1 week ago

Funny but fake

[–] HubertManne@piefed.social 4 points 1 week ago (1 child)

do have any perspective on this? like what was your reason for sharing?

hideshow 1 child comment

[–] ThelastfingerofH@lemmy.world [S] 4 points 1 week ago* (last edited 1 week ago)

It's interesting to see that what anthropic calls top secret comparable to a nuclear weapon is really what open source ai has been doing all along in the open. The only thing open source lacks is scale, and I think that's the closed ai labs' only moat. The closest open source model in scale is Kimi k3.1 at 2.8 trillion parameters

parent

[–] TheFrirish@tarte.nuage-libre.fr 4 points 6 days ago

Woooooooow amazzzing! I cannot wait to use a 999PB on my 20GB VRAM card

[–] ThelastfingerofH@lemmy.world [S] 2 points 6 days ago*

This is also like the reason why Nvidia markets the nvl72 rack as being able to train 10t parameter models, because closed ai companies already do it with 1 or two racks. Maybe 20t parameter models with 2 racks.