we use a model prompted to love owls to generate completions consisting solely of number sequences like “(285, 574, 384, …)”. When another model is fine-tuned on these completions, we find its preference for owls (as measured by evaluation prompts) is substantially increased, even though there was no mention of owls in the numbers. This holds across multiple animals and trees we test.
In short, if you extract weird correlations from one machine, you can feed them into another and bend it to your will.
Here’s a metaphor/framework I’ve found useful but am trying to refine, so feedback welcome.
Visualize the deforming rubber sheet model commonly used to depict masses distorting spacetime. Your goal is to roll a ball onto the sheet from one side such that it rolls into a stable or slowly decaying orbit of a specific mass. You begin aiming for a mass on the outer perimeter of the sheet. But with each roll, you must aim for a mass further toward the center. The longer you roll, the more masses sit between you and your goal, to be rolled past or slingshot-ed around. As soon as you fail to hit a goal, you lose. But you can continue to play indefinitely.
The model’s latent space is the sheet. The way the prompt is worded is your aiming/rolling of the ball. The response is the path the ball takes. And the good (useful, correct, original, whatever your goal was) response/inference is the path that becomes an orbit of the mass you’re aiming for. As the context window grows, the path becomes more constrained, and there are more pitfalls the model can fall into. Until you lose, there’s a phase transition, and the model starts going way off the rails. This phase transition was formalized mathematically in this paper from August.
The masses are attractors that have been studied at different levels of abstraction. And the metaphor/framework seems to work at different levels as well, as if the deformed rubber sheet is a fractal with self-similarity across scale.
One level up: the sheet becomes the trained alignment, the masses become potential roles the LLM can play, and the rolled ball is the RLHF or fine-tuning. So we see the same kind of phase transition in prompting (from useful to hallucinatory), in pre-training (poisoned training data), and in post-training (switching roles/alignments).
Two levels down: the sheet becomes the neuron architecture, the masses become potential next words, and the rolled ball is the transformer process.
In reality, the rubber sheet has like 40,000 dimensions, and I’m sure a ton is lost in the reduction.