Shaped by Training: What Really Sets a Model's Values | Geometry of Trust | Philosophy - Lesson 3

Alignment Is Paint on a Landscape | Geometry of Trust Philosophy Lesson 3

A forest's values come from soil and climate. A wolf pack's come from territory and prey. An AI's come from three things: the corpus it read, the architecture that processes what it read, and the training objective it was rewarded for.

This talk walks through each of those three shapers — and arrives at a finding that should change how we think about AI alignment.

When researchers apply post-hoc alignment methods like RLHF, DPO, or supervised fine-tuning to already-trained models and measure the underlying value geometry before and after, the geometry barely moves. Three alignment methods, eight models across different architectures and corpora, same result. What changes is surface behaviour. The underlying structure that generates that behaviour is unchanged.

The value structure is set during training — by the corpus, the architecture, and the objective. Alignment is a thin behavioural veneer layered on top. It shapes what the model says. It doesn't much change what the model is. A landscape with a thin coat of paint. You can re-paint as often as you like. The landscape underneath doesn't change shape.

The implications are stark. Certifying the alignment method is insufficient. You need to inspect the training pipeline. You need to monitor the geometry, not just outputs. Which is exactly what the mathematics series produces.