Brain-Inspired AI: Are We Using the Right Brain Map?
A reading of the Foundation Agents survey, which maps agent modules onto brain regions. The essay recalls that modern neuroscience favours distributed, emergent functions, and asks how such functions could emerge in AI instead of being engineered as modules.
This article is more than a year old. Some of its findings may have changed since.
Examining the tension between engineered modularity and biological emergence in artificial intelligence
A new Foundation Agents survey paper has sparked considerable discussion online, though largely through confirmation bias—with most commentators extracting what they expect to find without engaging beyond the first figures. Having read Part I thoroughly (and summarized Parts II-IV with AI), here is my attempt at some nuance.
The Brain-Inspired Framework
The paper uses a "brain-inspired" modular framework to organize the current agent architecture landscape, mapping components like memory, world models and reward systems onto brain functions. This approach provides a valuable organizational scaffold for categorizing the field's sprawling developments.
The paper's opening, however, can throw readers off track: it shows an outdated brain diagram with regions labeled by functionality, and "Visual Perception" or "Language Comprehension and Production" indicated as "Well-developed in current AI." While the likely intent is to convey AI progress vaguely inspired by neuroscience, the casual reader might miss two key points.
Two Critical Misconceptions
First, "well-developed in AI" doesn't mean developed "like in the brain." Even if aspects of human behavior were entirely reproduced, there would still be a wealth of compatible algorithms that do not generalize to other behaviors like the brain. We don't know what type of high-level information processing the brain does. Studies have even shown that modern mathematical tools fail to meaningfully reverse-engineer something as simple as an old microprocessor.
Second, while useful to brainstorm, brain-region mappings reinforce an outdated paradigm: Modern neuroscience has largely abandoned strict "this-region-does-that" modularity in favor of distributed, emergent properties across neural networks.
Memory isn't a standalone system but emerges from dynamic interactions across hippocampal-cortical circuits. Reward processing involves complex cascades spanning midbrain dopaminergic pathways and regulatory feedback loops.
The Greatest Challenge
In the subfields of AI aimed at human-like intelligence, the tension between engineered modularity and biological emergence is one of the greatest challenges. We're attempting to mechanistically engineer what nature never designed in that way.
Instead of adding modules for memory, world models and reward systems, perhaps we should try to find how such functionality can emerge—under what learning conditions, environmental pressures, architectural constraints, etc.