Magnus Petersen

Center for the Alignment of AI Alignment Centers We align the aligners

Very glad to see that someone is doing the important work of aligning AI alignment. alignmentalignment.ai

People I know who don't work in tech have not even heard of Anthropic. That might be the issue, also for non-technical tasks, it's very hard to discern which model is really "better"

Thanks! Actually had a chat with Sanjeev at his poster yesterday, great stuff! I think the amortization of these kinds of methods with pre-trained generative models is actually very promising! Will update the arxiv with this work soon

PINN-MEP: Continuous Neural Representations for Minimum-Energy Path Discovery in Molecular Systems Characterizing conformational transitions in physical systems remains a fundamental challenge in the computational sciences. Traditional sampling methods like molecular dynamics (MD) or MCMC often str...

Link: arxiv.org/abs/2504.16381

Very cool related work: Doob's Lagrangian (Du et al., 2024): arxiv.org/abs/2410.07974 Implicit Neural Representations for Chemical Reaction Paths (Ramakrishnan et al., 2025): arxiv.org/abs/2502.15843

We represent a transition path as a continuous neural mapping and minimize an Onsager-Machlup-action-derived, simple, and scalable loss - literally just the sum of energies along the sampled neural path!

Presenting our work on minimum energy path generation between two states for physical systems at the FPI Workshop at @ICLR tomorrow! Scaling up to solvated BPTI and observing the same conformational changes as long reference MD with six orders fewer force field evals! Drop by!

Check out my poster on 'Trajectory Level Langevin Dynamics' aka 'Simulations of simulations', using differentiable MD forcefields, at the 'ML and the Physical Sciences' Workshop @NeurIPS!

Link: openreview.net/pdf?id=daEra...