Hi, I'm Jesse

A practical guide for anyone who wants to replicate or extend Timaeus projects. It covers the methodology for using SLT-based tools in practice, with a focus on choosing hyperparameters for sampling methods that scale to models with billions of parameters, and notes the theoretical questions about these samplers that are still open.

We use the geometry of the loss landscape, rather than neural activations, to reveal internal computational structure in a language model. Susceptibilities measure how the loss landscape responds to changes in the data distribution. Scaling this method to Pythia-1.4B, we find 57,236 susceptibility clusters, many of them much more abstract than those found in the smaller Pythia-14M.

Current training data attribution (TDA) methods treat the influence one sample has on another as static, but neural networks learn in distinct stages that exhibit changing patterns of influence. In this work, we introduce a framework for stagewise data attribution grounded in singular learning theory. We predict that influence can change non-monotonically, including sign flips and sharp peaks at developmental transitions. We first validate these predictions analytically and empirically in a toy model, showing that dynamic shifts in influence directly map to the model's progressive learning of a semantic hierarchy. Finally, we demonstrate these phenomena at scale in language models, where token-level influence changes align with known developmental stages.

AI safety is split into two cultures, much like nineteenth-century geology was split between catastrophists and uniformitarians. I map that divide onto today's debates (sudden discontinuities or gradual progress, reasoning backward from superintelligence or forward from current systems), trace how it shapes research methodology, community, and strategy, and argue that neither pure theory nor pure empirics scales on its own.

I introduce Resolution, a new AI alignment organization built around a portfolio of higher-confidence theory bets plus research automation, backed by $160M.

Steve Hsu interviews Jesse Hoogland, co-founder and executive director of Timaeus, about Singular Learning Theory, AI safety, and the case for AI doom. Part of the 'Dreamers and Doomers' documentary interview series about the SF Bay Area before AGI. 103 minutes.

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.

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.