Neural networks are programmed by their training data, but there is rarely a unique way to minimize the loss: different minimizers can implement different internal algorithms. Patterning is a principled way to shape which structure a model learns by shaping its training data. We use the Toy Model of Superposition to study how patterning works in a setting where the possible solutions are well understood, and demonstrate multi-step patterning.

Hironaka's 1964 theorem on resolution of singularities in characteristic zero is a pillar of modern algebraic geometry, and through Watanabe's singular learning theory it connects loss landscape geometry to generalization. We present a Lean formalization of the theorem, released with a blueprint and API docs, as groundwork for a theory of generalization relevant to AI alignment.

The mathematics and methods behind scaling and analyzing susceptibilities in models with billions of parameters. We describe how we collect and post-process ground-truth susceptibility data, and how we train an auxiliary model to upsample those results to 46 million vectors. Companion to "Spectroscopy at Scale".

Resolution (fka Sequent) has a $160M grant from Coefficient Giving to put rigorous alignment research on a (closer to) even footing with the frontier labs.

We are starting a large nonprofit research organization, Resolution, that aims at higher confidence in alignment via a portfolio of theory and empirics bets, with heavy investment in automation to accelerate progress on these bets.

Research automation is not just about aligning the first AI scientist. It is also about the institution-building problem of coordinating fleets of AI researchers.

A conference on developmental interpretability and singular learning theory, 5 to 12 November 2023 at Wytham Abbey, Oxford.