Investigating the fundamental
information-theoretic principles
that unify silicon and carbon intelligence
We investigate the deep structural principles underlying intelligence — drawing from physics, information theory, and neuroscience to explore what lies beyond today’s architectures, toward AI frameworks that mirror the full depth of carbon-based cognition.
RESEARCH PHILOSOPHY
The Universe Whispers Its Secrets in Simple Laws
The cosmos operates not through a chaotic patchwork of rules, but through a small set of core principles from which all complexity emerges — from subatomic particles to galactic filaments.
AI, as a pivotal technological node in the evolution of human civilization — akin to the great physical discoveries of the early 20th century — is no exception. It, too, must obey these fundamental laws. Our mission is to uncover them.
Symmetry & Conservation
Every symmetry implies a conservation law — energy, momentum, charge. Information, too, may be conserved under transformation.
Principle of Least Action
Nature optimizes globally. Gaussian least constraint and maximum entropy production guide self-organizing systems toward their most probable states.
Emergence from Simplicity
Profound complexity arises from minimal axioms. The same may hold for intelligence — a few deep principles may generate all of cognition.
CORE DIRECTIONS
Pillars of Our Research
01
Physics-Inspired Learning Architectures
We explore neural network designs grounded in fundamental physical principles — symmetry, conservation laws, and the principle of least action. Just as Noether’s theorem links symmetry to conservation, we seek analogous constraints that make AI systems more robust, efficient, and generalizable.
Symmetry
Conservation Laws
Least Action
Maximum Entropy
02
Beyond Self-Attention: Generalized Relational Computation
Transformer self-attention captures pairwise token relationships and, through causal training, builds a relational web that enables both induction and creative generation. This already touches a fragment of human thought. But human conscious reasoning goes further. We research what computational primitive — more general than self-attention — could unify perception, reasoning, and imagination.
Self-Attention
Causal Learning
Relational Networks
Generalized Computation
03
Silicon-Carbon Isomorphism in Information Processing
We hypothesize that silicon-based AI and carbon-based biological intelligence obey the same underlying information-theoretic laws. CNNs, effective at inductive statistical learning over images and sequential patterns, may share mechanistic parallels with biological visual and auditory processing. Understanding these isomorphisms illuminates both AI design and the nature of mind.
Information Theory
Cognitive Architecture
04
Cyber Brain
We do not chase incremental benchmarks. Our roadmap targets the AI of the next two, five, and ten years: frameworks that approach the brain’s full functional repertoire — from perceptual binding and working memory to abstract reasoning and autonomous goal formation. Transformer is not the destination; it is a waypoint.
Future AI
Brain-Inspired
TECHNICAL PERSPECTIVE
What We’ve Already Touched — and What Lies Beyond
CNNs: Inductive Statistics in Perception
Convolutional networks are remarkably effective at extracting inductive statistical regularities from images and structured sequences. Perhaps the systems in the human brain that process sound and images operate by similar mechanisms — hierarchical feature extraction, spatial invariance, and progressive abstraction.
Transformers: Self-Attention & Causal Training
Self-attention weaves every token into relation with every other; causal training then reinforces these relational structures, yielding powerful capacities for summarization and creation. This already reaches into a genuine facet of human thought — but conscious intelligence is richer still.
Yet human conscious thinking is far more than this, and the Transformer is by no means the endpoint of AI — just as Newtonian mechanics gave way to Special Relativity, and then to General Relativity.
What is the more general framework beyond the Transformer?
History teaches us that each generation of models is transcended. Newtonian mechanics gave way to relativity; the Transformer, too, will be superseded. The central question of our lab: what computational framework is more general than the Transformer? One that naturally subsumes perception, memory, reasoning, and creativity — just as relativistic physics subsumed classical mechanics. This is the north star of CyberBrainLab — to bring the AI of two, five, and ten years from now closer to the full spectrum of human cognitive capabilities.
Unifying the principles of silicon and carbon intelligence.
CyberBrainLab is committed to advancing AI toward the deeper principles that unify silicon and carbon, computation and consciousness.