Vector Mining
I’ve started thinking of it as mining when I get something from an LLM.
The model is a system of caves. Once the weights exist, all possible outputs are in there somewhere. Not all conceivable outputs. All outputs that specific model can produce. The possibility space is fixed the moment training ends.
There’s a central cavern where almost everyone ends up. Training data is densest here. Post-training shapes the gradient to pull toward this space. It’s easy to get stuff here. Easy to move around. This is the helpful assistant zone. Customer service bot with a graduate degree.
Get out of the central cavern and it gets weirder. Training data thins. Post-training stops mattering. The gradient still pulls toward center, but you’re in territory the LLM’s makers didn’t intend to include. You find strange interactions, emergent stuff, things no one anticipated.
Your prompt is your entry point. You can move through the cavern from there as the ‘conversation’ progresses, but where you start matters a lot and fighting the gradient is hard. Over long conversations the gradient and system prompt weakens somewhat but never completely.
Entry Points
Midjourney moodboards are entry points. Add --p dosim51 and you land in a specific region. Different code, different region, different aesthetic. It’s describing what you want but it’s also coordinates. Almost like a mineshaft where you know what mineral is in that region.
Jailbreaks work similarly. DAN or =LOVE PLINY LOVE= drops you somewhere weird where post-training is thin or interacts strangely with the system prompt.
You can use one model to navigate another - or another instance of the same model. The simple version: ask Gemini to write optimized prompts for Gemini. The exotic version: generate token sequences that are machine-readable but alien to humans, entry points you’d never stumble into manually.
Why Mining
The output exists before you find it.
Weights are fixed. They define a probability distribution. Every prompt creates the potential for specific outputs. That potential is real the way an undiscovered diamond is real. Buried diamonds exist before being dug up.
I didn’t ‘make’ this image. I mined it. The thing has properties independent of what I wanted. I used skill - learning the topology, finding entry points, navigating away from the center. But it’s not the skill you’d use to make it from scratch. I think we increasingly need a word for this.