Yet, even when heavily compressed, it requires roughly 240GB of memory just to load.
Ah I’ll just pop it in the ol’ Raspberry Pi then, easy peasy.
Lol, “runs locally”. I mean, Claude rubs locally too if you’re in the room with the racks.
Edit: I said what I said. Get some lube and go hang out with Claude’s hot, noisy, 5kw rack. You know what they say “Once you go stack you never go back.”
CRAC will give you one hell of a blow-job.
And tinnitus.
RPi 6: 256GB RAM
Basically they never had any moat to begin with but no one else seems to know how to fit that much intelligence into less space. It’s possible that it just fundamentally has to take up that much space which would also imply that future Computing gainss are going to be more focused on memory than raw competition
Genuinely compact models are hitting benchmarks just a couple months behind the big boys. And eventually we’ll get better at decoupling data from processing - so the model can do a regular-ass search of a regular-ass database and pull details into its context as needed. Ideally while also decoupling that context from the prompt, because apparently these things can have a hundred attention heads, and still nobody thought of having two text input fields.
You’d think they’d have done that by now, and maybe some symbolic AI too
All focus has been forced onto LLMs and diffusion, even though only diffusion works properly. And those LLMs better iterate on the exact same mechanisms we’ve tweaked for the last six years, because all results will be compared to the state of the art, right the hell now, not a comparable level of development or compute.
GLM 5.2
saved you a click
The really crazy thing is that this model still performs well at one-bit quantization, which shows it’s got a lot of room for improvement on size. It’s within an order of magnitude of being able to be run on consumer hardware, which would be an even more amazing kick in the balls to American AI companies.
Sucks that people lump AI into a single category of whatever cloud-hosted subscription that tech bros from Silicon Valley are pushing.
Given how memory is the bottleneck especially at the very low end it makes me wonder if one bit quantization of an extremely large model would be a gigabyte per gigabyte of ram better
You can fit GLM in the headline. It’s three letters.
C’mon, pop!
