Excited to share Bolero, out on bioRxiv. @arcinstitute posted a great overview. Here I want to add another perspective that motivated it for us (also a few more higher resolution images).

Excited to share some of my new work being presented at #NeurIPS2026 this year! 1. Compositional Perturbation Planning with a Generative Structural Causal World Model - openreview.net/pdf?id=N2ewxv… Going from an embryonic stem cell to a target cell fate is a multi-step planning problem, but current perturbation models can’t capture multi-gene effects or epistasis. We built a structural causal VAE world model that takes in sets of transcription factors, represents interventions in its latent causal graph, and accurately predicts held-out high-order perturbations. A GFlowNet policy trained over it autoregressively samples TF sets to reach downstream fates, recovering the correct genes at 36x lower cost than search and nominating well-known blood-lineage drivers zero-shot for hematopoietic stem cell fate. Thank you to Ivy Liu and @genophoria for your mentorship, and @arcinstitute for the support! 2. Interpretability-Conditioned Value Priors for Evaluation-Efficient Intervention Search - openreview.net/pdf?id=CkD0V7… Interpretability tools reveal a lot about what a model has learned, but that insight rarely turns into better decisions, and testing every possible fix is expensive. With my good friend @parshawnn , we asked whether concept-level interpretability can predict which interventions on a representation are worth evaluating before any feedback comes in. On a shortcut-learning benchmark built on frozen CLIP features of CelebA, an interpretability-based value prior reached 95% of the attainable improvement with 64% fewer evaluations than a matched generic model and 73% fewer than random search.

Dialectic 60: Patrick Hsu! @pdhsu & @arcinstitute are combining biology, engineering, and AI to turn a sci-fi future into reality. Patrick got expelled from college at 13, returned at 15, later worked on early CRISPR breakthroughs, and teamed up with @SKonermann & @patrickc to found Arc. He works 3 jobs: core investigator at Arc, professor at Stanford, and venture partner at @thrivecapital. For all his intensity and ambition, he carries a lightness and boldness that is contagious. We discuss: - why research is 99% failure and even your best ideas kind of suck - infinite games and speedrunning school - the reach of GLP-1s and power laws in biotech - what makes a good biologist - Arc’s bet on people over projects - lab robots that finally work - AI-designed organisms and new forms of life - why heart disease could go the way of infectious disease - bro science, peptides, and biofuturism - research taste in the age of AI Timestamps: (0:00) Opening Highlights (1:39) Intro: Patrick Hsu (4:29) Start: Boldness, Infinite Games, & Ambition (12:54) GLP-1s and Power Laws in Biotechnology (20:17) Expelled at 13 and Falling for Biology (25:51) What Makes a Good Biologist, Research Taste, and Biology's Data Challenges (37:58) The Arc Model: Full-Stack Science (52:01) Betting on People, Style, & Learning to Simplify (57:22) Perturbation, World Models, and Evo (1:03:27) Virtual Cell Models (1:09:56) Designing Life: From Stem Cells to Genome Writing (1:18:39) Testing, Bro Science, and Human Data (1:24:26) Can It Exist vs. Will It Exist: Science Meets Markets (1:29:59) Biofuturism: Interoception, Peptides, Muscle Modulation, Sleep, and New Forms of Life (1:43:47) Fun, Obsolescence, Going Slow, and the Future @dialecticpod 60: Patrick Hsu - Speedrunning the Infinite Game - is available on all platforms, including X. Links below.