Biology Beyond Human Intuition

gist

Morgan Levine argues that biology may exceed human intuition and that AI should learn its own representations of living systems instead of ours. Human categories like genes, pathways, cell types and diseases shape what is measured, tested and treated, and language models trained on literature inherit the same compression. Models should instead learn representations directly from many kinds of biological observations.

ideas

• Human abstractions lose information. Genes, pathways, cell types and diagnostic labels made modern biology possible, but they are incomplete; cells can sit in metastable states that do not match established labels.
• Measurement and models form a loop. What scientists measure reflects what they already believe matters, so existing frameworks shape the evidence used to revise them; she cites Michael Bronstein and Luca Naef's case for black-box data built for machine learning.
• Narrow experiments limit generalization. A few standard cell lines, model organisms and assays aid reproducibility but yield models that predict well inside those conditions and poorly outside them; diversity of context may matter more than repetition.
• Reasoning is not representation. Language models learn what scientists chose to write down, and reasoning better inside an existing framework is not finding a better one, so biological AI needs better ways to represent biology itself.
• Discrete interventions can still steer. In a reply she cites evolution, which navigates a high-dimensional system with simple discrete operations, as a model for using learned representations to choose interventions.

quotes