Whole Brain Modeling and Neurosynthetic AI: Merging Computational Neuroscience and Artificial Intelligence

Current AI is missing the brain's most important trick — feedback — and this paper proposes a concrete plan to fix that by building AI directly from whole-brain simulation models.

The core argument is simple. Modern deep learning is mostly feedforward: data flows in one direction, gets transformed, and produces an output. Biological brains don't work that way. They are densely recurrent, they oscillate, they operate near a "critical point" (a dynamical sweet spot between order and chaos), and they constantly generate predictions about incoming sensory data rather than passively processing it. The paper calls the gap between these two worlds the central problem, and proposes "neurosynthetic AI" as the solution — not just drawing loose inspiration from neuroscience, but literally using large-scale computational brain models as the AI architecture itself.

The building blocks are Neural Mass Models (NMMs): mathematical descriptions of how populations of neurons behave collectively, capturing oscillations, excitatory-inhibitory balance, and nonlinear dynamics. The paper reviews several flavors — Kuramoto phase oscillators, Wilson-Cowan excitatory-inhibitory circuits, Laminar NMMs that respect cortical layer structure — and explains how these can be wired together using real anatomical connectivity data (from diffusion MRI tractography) to produce a "neurotwin," a whole-brain digital replica. Two theoretical frameworks are layered on top: Active Inference (Karl Friston's idea that the brain minimizes prediction error through a generative model of the world) and Kolmogorov Theory (Giulio Ruffini's framework emphasizing that intelligence is fundamentally about finding compact, compressed descriptions of sensory data). Together these provide the learning and inference principles that the raw brain simulation lacks on its own.

The practical roadmap, called the SYNTHEON Development Framework, is staged over roughly seven years. Year one and two: treat the whole-brain model as a reservoir computer — a recurrent dynamical system whose rich internal states are read out for simple tasks like detecting an auditory mismatch — and validate it against EEG and fMRI data. Years three and four: integrate laminar models and active inference loops. Years five and six: add full sensorimotor embodiment in virtual or robotic environments. Year seven onward: human neurotwin validation and clinical pilots. The paper also flags "brainoware" — hybrid systems combining living cortical organoids on electrode arrays with artificial components — as a radical near-term experimental testbed.

What this unlocks, if it works, is significant: AI systems that are interpretable (because their internal states map onto known brain dynamics), adaptive (because criticality and multistability let them switch cognitive modes rapidly), and clinically useful (because the same neurotwin technology already shows promise for predicting individual patient responses to brain stimulation in epilepsy and depression). The honest caveats are also noted — whole-brain simulation is computationally expensive, mapping abstract theories like Active Inference onto specific physiological parameters is hard, and as these models grow more sophisticated, questions about agency and consciousness become non-trivial. The paper is a research proposal and conceptual framework more than a results paper; the appendix makes this explicit, reading as a grant pitch with placeholder budget figures. But the intellectual scaffolding is carefully assembled and the roadmap is specific enough to be actionable.