Latest / Logosofia no Dia a Dia
Why AI Weights Are Indecipherable
When you download an AI model, you're getting a file of billions of numbers. But those numbers aren't a readable database of knowledge — they're a tangled, distributed pattern that resists interpretation and editing. In this episode, we break down why neural network weights are so opaque, what dictionary learning has revealed about the hidden structure inside models, and why surgical knowledge editing keeps breaking. If you've ever wondered what's actually inside a safetensors file, this one's for you. Episode 4670 — myweirdprompts.com/n/4670 (permanent short link: myweirdprompts.com/479236)
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.