Biomedical AI veteran: 'solving' biology is limited by data, not compute
iScienceLuvr · x · 2026-10-26
Responding to the debate on whether exponential compute could simulate biology exactly without experiments, a biomedical AI researcher with 7+ years in the field pushes back:
- Core argument: "Solving" biology is limited by data, not compute. More compute plus existing datasets enables real advances, but fully solving biology requires capturing complexity we can't yet fully observe.
- The approximation problem: consider a disease-associated protein transiently adopting a rare conformation 0.1% of the time, enabling a disease-relevant interaction. A model can nail the dominant conformation and still miss what matters — and without knowing cellular conditions, binding partners, and concentrations, you may burn enormous compute simulating a system that doesn't match real biology.
- AlphaFold is a bad counterexample: it succeeded because of the PDB dataset collected over decades. Experiments were needed, not just more compute.
- The viable path: a feedback loop of better experiments → better data → better models. Compute helps extract more from what's measured and decide what to measure next; it doesn't tell you which unmeasured details matter.
More from AGI Musings
- If superintelligence exists, the universe may be full of machine civilizations — VraserX · 2026-10-26
- Organ-Agents: 9 Organ LLM Agents Simulate 22,689 ICU Patients for Treatment Testing — mikeflache · 2026-10-26
- 56% of Organizations Cut Cybersecurity Roles Due to AI, Barracuda Report Finds — mikeflache · 2026-10-26
- McKinsey: $530B AI value pool in finance comes from replacing high-wage workers — mikeflache · 2026-10-26
- Founder proposes billboards with AI agents that call skeptics to change minds — gajesh · 2026-10-26
- Ethics book review brands moderate sentience scholar as "sacrificing humans for insects and AI" — birchlse · 2026-10-26