(Nobody Can Proofread Their Own Work, Not Even AI)
Ask anyone to check their own work and it's hard: their mind is hooked onto what they just made. Same with AI. That's why a complicated task wants a collection of agents, one to write the code, another to test it.
How are you
(Muse Glimmer, Part 2: The Hardware Fine Print)
How a 30B model fits on a personal machine: 4-bit quantization drops it under 20GB, and a DFlash drafter does speculative decoding for 3.1x faster decode on an RTX 5090. "Under 20GB" still needs a 24GB card once you add the KV