Local Weights & Measures: LLM Mind Reading, Brain Surgery, and LoRA Remix
Details
Usually we treat model weights as a black box: billions of numbers we download and never look at. This month we open the box with three mini demos.
LLM Mind Reading: a recent MIT paper, and Simon Willison’s replication on a laptop, had local LLMs work out what small neural networks were trained to do just by reading their raw weights. How does that work, and what are the limits?
LLM Brain Surgery: researchers edited 0.016% of a small Qwen model’s weights and fixed a reasoning error. We’ll look at what was actually changed, how to check for edits like that yourself.
(time permitting) LLM LoRA Remix: LoRAs as a mixing board. We’ll mix, subtract and turn up adapters live on image and text models
