the entire point of AGI is to respond gracefully to aberrations though. If your pipeline is high-latency you're not going to be cold welding plastic enclosures at a rate of 3600 an hour with a takt time of 1.6 seconds which makes you worse than a Denso
The core of this methodology is that you send in your problem spec, and the model outputs code that will control the machine, including handling any (within reason) aberrations. The code can then be run locally, with near-zero latency, only pinging the LLM if the directive changes or if something goes awry. The LLM isn't sending you the exact steps to manufacture part number 3599 after sending them 3598 times beforehand in the past hour. It's looking at the tools it's been given and the properties of the assembly line and sending you a 250 KB file that you can load into your controller and then let the line run on its own without spending any additional API credits or using any additional bandwidth.
I'm not sure if I really got the spirit of your question, but here is what I got on the first try:
is to shift the initiative to algorithm because humans suck at future.
go back to india
Interesting. Lets try and lowball the parameter count by assuming the model takes up a "reasonable" 50TB at fp16. GLM-5 is a 744B MoE with 40B active parameters and takes about 2TB ram to deploy comfortably at fp16.
Oh wow another model that can't be ran on consumer hardware. I just don't understand what this company is doing. Mistrial small is still considered a good mid B LLM. They really can't compete with OpenAI and Anthropic since they are backed by M$ and Amazon respectively.
If the fucking Europoors somehow come out of nowhere with a Mythos-tier model after years of being passed by even the Chinese, let alone every single American company, I will buy a bidet, turn off my AC for a day, and learn how to play soccer.
Do you mean MoE, or is there some subtle difference I'm not catching here?
Presumably a much more lightweight routing system so you don't need a 1TB model loaded but being able to load a 32GB "expert" model once the classifier figures out which one should be used.
The idea is like OpenClaw. You have an agent you ask a question and it identifies the best AI tool for the job and delegates the task to it. There are plenty of non LLM AI models optimized for specific fields (ex: cancer detection, image generation, object segmentation, pose estimation, speech transcription, etc) that outperform giant LLMs. If you had a thousand specialized AI models for specific usecases that were individually top of the line, you just need an LLM inteligent enough to know which one the user wants and can consistently route their request to the right domain specific LLM.
That rather defeats the point, even if you can get it to work without pulling on the frayed parts that MoE has to begin with. May as well just train a single small model on your task of choice (or as many as you have tasks) and release them separately with plain English labels, with no need for the extra overhead or failure chance of an automatic router.
Do you really want to have to pull down a menu and select a specific LLM and wait for it to load?
My nigger in heaven this is a joke on AI twitter, it's not real
Essentially custom MoE where you select a router and route your query to the appropriate fine tune. If you are working on java code ship it to a java finetune, or Gemma if you want more artist writing. You select the smaller finetunes or experts you want to run. Then leave out the rest. With these larger models it's all or nothing.
You can either load one big model in one go or you can load small models as you need them. Insofar as you want to do the latter, waiting for new models to load in as needed is something you can't avoid.
It is pretty wild that you can describe any image and get it. Even years ago, this had extreme potential for funny outcomes. I've been toying with the idea of putting together an AI-generated "Ben Shapiro Show" in the same vein as the cancerless jim show some guy made all those years back. Some light political satire and racism, but mostly slapstick comedy in which he struggles to reach things on tall shelves.
