The transition from regression to generative/diffusion modeling in structural biology is one of those architectural shifts that takes years to execute but is obvious in retrospect. Boltz-2's unified encoding of structure and sequence prediction is exactly the kind of 'simulator' layer that Melissa Du argued is missing from pure LLM science reasoning - it predicts physical phenomena rather than pattern-matching text about it.
The democratization vs. premium infrastructure split is worth watching: open-source models for access, GPU-optimized infrastructure for serious use. That same pattern is playing out in AI coding tools. Capability commoditizes; reliable infrastructure is where the value concentrates.
The transition from regression to generative/diffusion modeling in structural biology is one of those architectural shifts that takes years to execute but is obvious in retrospect. Boltz-2's unified encoding of structure and sequence prediction is exactly the kind of 'simulator' layer that Melissa Du argued is missing from pure LLM science reasoning - it predicts physical phenomena rather than pattern-matching text about it.
The democratization vs. premium infrastructure split is worth watching: open-source models for access, GPU-optimized infrastructure for serious use. That same pattern is playing out in AI coding tools. Capability commoditizes; reliable infrastructure is where the value concentrates.