Ventali Tan
Physics, philosophy, and the machines that learn the world.
I studied physics, philosophy, and computer science at Duke, with detours into archaeology. I am currently interested in the science of training data, scalable alignment of LLMs, and developing intelligence systems for scientific discovery.
My work focuses on automating synthetic data generation, structuring scientific knowledge (e.g., Paperbase), and exploring hybrid architectures that combine language models with formal or symbolic methods to enable more rigorous reasoning and abstraction discovery. Previously, I built several agentic systems and Valida, a custom instruction set and compiler architecture for zero-knowledge virtual machines.
I run an AI & Science reading group in San Francisco and angel invest through Delendum.
Work
Writing
- Outrunning Goodhart: over-optimization in training data
- The lab as one policy: training an all-agent research organization — Part I
- DeepSeek’s Thinking with Visual Primitives: a well-worked example of OPD, and its limits
- Why post-training is moving toward on-policy distillation
- Deleuze, Sartre — escape / confront absurdity?
- Proving the circuit: formal verification for zero-knowledge proofs