Brains, Minds, Machines Blog

Foundational Principles of Intelligence —
A Blog Series by the Poggio lab

Modern AI is advancing at breathtaking speed, yet our scientific understanding of why these systems work remains incomplete. For the first time in history, we can systematically compare different kinds of minds—human, animal, and artificial—and ask what principles they share. This blog series explores the foundational ideas that make intelligence possible: structural, computational, and biological principles precise enough to be expressed in theorems or probed experimentally, and powerful enough to inform the next generation of AI architectures and applications.

This effort is closely aligned with our new MIT-area seminar series, AI: Fundamental Principles & Startups. The motivation is simple: it is still profoundly worthwhile to do research in AI—if one is pursuing first principles of intelligence, or building applications grounded in those principles strong enough to found companies. We aim to build a physics-style blueprint for intelligence, where evidence and mathematical structure constrain each other. From the empirical side, we draw on cognitive development, systems neuroscience, and the observed behavior of transformers, SSMs, and reasoning agents. From the theoretical side, we focus on sample efficiency, sparse compositionality, invariances, modular reuse, hierarchical memory, optimization dynamics, and generalization—expressed as definitions, lemmas, conjectures, and theorems.

The blog series will track this broader mission: developing a shared vocabulary and set of testable ideas about intelligence. Posts will cover principles (compositional sparsity, causal abstraction, scaling laws), architectures (long-context models, neurosymbolic interfaces, agentic systems), and forms of evidence (benchmarks, theory-guided ablations, neuro/behavioral probes). We will also highlight the translational pathway from principle to prototype to product: how fundamental insights can shape new systems, new workflows, and ultimately new startups. The intended audience is the same as the seminar’s: ML theorists and experimentalists, neuroscientists, cognitive scientists, applied mathematicians, and founders or industry researchers who believe that the next wave of AI will be driven by understanding, not just scale.

Our goal is ambitious but concrete: to illuminate the structural regularities that make minds—biological and artificial—work, and to turn those regularities into deployable algorithms, architectures, and company-forming ideas. If you are motivated by first principles, high-conviction applications, or both, we invite you to follow the series, join the discussions, and help shape a science of intelligence that is rigorous, testable, and generative.

Recent blog posts