Anonymous #1 1d

Big LLMs need a lot of brain cells (parameters) to do more complex tasks. When AI companies release “weights” for an old model, you can use these to run your own model on your computer, rather than borrow some processing power from a huge cluster of computers far away to do it for you. Since your computer is not a server farm, it is limited in what size models it can run, leading to trimmed versions with lower parameter sizes.

3

Anonymous #2 1d

I am sure, but also I can just defer response time, it takes longer but based on what mandatory reporting(IRS Forms tell a lot)seems to indicate, after a request actually filters down to a processing stage, the allocator can give as many as 3 requests to 1 GPU(assuming txt isn’t continuously processed, which is is). Given, these r higher end models & processors, & even then, nothing compared to what Altman Has, but…

1

Anonymous #1 1d

If processing power & speed is a concern, I had Grokbot mess around w/my local models so that they can self replicate onto other devices w/ a USB Thumb drive. It even made an installer for me. Idk exactly how It works but it does. Apparently it “enables virtualization” and basically runs separate from the main OS “under the stability thresholds”(Idk) On most machines and my laptop is the 1 w/the allocator API. I have like 36 machines at any given time.

1