Dalibor Bobr

I asked GROK to refute this argument. Big datasets are necessary but nowhere near sufficient to “solve” neuroscience. AI is excellent at pattern-finding and prediction inside the distribution of whatever data you give it. That is not the same as producing mechanistic understanding of a dynamical, adaptive, multi-scale system whose relevant variables are not obvious. Genomics analogies break down for several reasons. Protein folding had a clean mapping problem (sequence → structure) with strong physical constraints and a large, relatively standardized data corpus. Systems neuroscience does not. The relevant objects are circuits implementing control, prediction, credit assignment, and closed-loop behavior in an agent that is constantly changing its own weights and states. A connectome is high-resolution anatomy. From a heart’s connectome you cannot deduce that it is a pump; the same holds, more strongly, for brains. Structure does not equal function, synaptic counts do not equal effective weights, and correlational activity maps do not equal causal algorithms. The “ideal” dataset you sketch (behavior + SLM-level write + millisecond-resolution reads + cell identities + connectivity) is still incomplete for theory-building. Even dense observational data leaves you with massive underdetermination: many different mechanisms can produce the same activity patterns and behaviors. You need systematic, hypothesis-guided interventions and comparisons across states, individuals, developmental stages, and species. Generating and interpreting that interventional data at scale is itself a scientific problem, not a pure engineering scaling problem. AI can help analyze it and propose candidates, but it does not invent the right questions or the right experimental designs. Vague prompts of the form “neural basis of X” (where X is an ordinary-language concept) remain garbage-in, garbage-out regardless of dataset size. We already have substantial neural datasets from Neuropixels, large-scale calcium imaging, and early connectomes. Pure ML analyses of them have produced useful embeddings, decoding results, and some predictive models, but they have not delivered the transformative mechanistic leaps that hypothesis-driven work and well-posed problems have produced in other domains. More open data and less multi-year private mining by places like Janelia would be good—that cultural lag is real and slows progress—but it is secondary. Open data still requires the right conceptual framework to turn it into understanding rather than larger correlation catalogs. AI will accelerate good neuroscience. It will not substitute for posing crisp, experimentally tractable problems or for the iterative cycle of theory, measurement, and intervention that has always been required for complex biological systems.