01

The definition

Physical AI is not a product category. It is the recognition that the physical world and the intelligence that understands it were never separate to begin with โ€” and that building them as one system, derived from the same first principles, produces something qualitatively different from connecting two separately designed systems.

The conventional approach to intelligent sensing systems is additive: build a sensor, build an AI model, connect them at an interface. The interface is where the limitations live โ€” data format conversions, latency, information loss, the fundamental mismatch between what a sensor measures and what an AI model expects as input.

The Physical AI approach is derivational: start from the physics of measurement, the mathematics of information, and the epistemology of machine knowledge โ€” and derive both the sensing architecture and the intelligence layer from the same foundational theory. When you do this correctly, the interface disappears. Not because it has been engineered away, but because in the theory it never existed.


02

What the market means โ€”
and what it actually requires

Since Jensen Huang named the category at CES in January 2025, Physical AI has been used primarily to describe robots and embodied AI systems โ€” neural networks running on physical hardware that interacts with the world. This is a useful framing but it is incomplete. It describes the application domain, not the architecture.

Conventional approach
Sensor + AI โ€” connected
A LiDAR, camera, or radar system generates data. An AI model โ€” trained separately on labelled datasets โ€” processes that data and produces decisions. The two systems are integrated at an interface. The interface introduces latency, information loss, and a fundamental semantic gap: the sensor speaks physics, the AI speaks statistics. Moving parts, mechanical steering, proprietary connectors. The AI is always one step removed from the physical world.
Science Counter Inc ยท Physical AI
Intelligence + sensing โ€” derived together
The sensing architecture and the intelligence layer are derived from the same mathematical foundation: wave propagation, information content of physical events, causal state navigation. There is no interface โ€” because in the theory there never was one. The participation matrix PMkl is simultaneously the native output of the sensing process and the primary data structure of the AI system. No moving parts. No mechanical steering. No semantic gap between measurement and understanding.

03

The participation matrix โ€”
why PMkl is the proof of concept

The participation matrix is the central mathematical object of the SCI framework. It was not designed to bridge sensing and intelligence. It turned out to be both โ€” because the underlying physics, information theory, and epistemology are not as separate as field boundaries suggest.

PMkl โ€” three identities, one object
01
Sensing output
The native output of the ISS LiDAR platform. Each cell records the participation of aperture k in detection event l. No post-processing required โ€” the matrix is what the sensor produces.
02
Information structure
An information-theoretic encoding of the relationship between sensing elements and detected events. Equivalent to the association structures derived independently in information theory and knowledge representation.
03
AI data structure
The primary input to the State Navigation causal AI framework. No transformation, no feature engineering, no format conversion. The same object that the sensor produces is what the AI reasons over.

This is not a design achievement. It is a theoretical consequence. When you derive both systems from the same first principles, the data structures are naturally compatible โ€” because they describe the same underlying reality from the same mathematical foundation.


04

Priority and provenance โ€”
the work predates the name by decades

Physical AI as a market category was named in January 2025. The foundational work that defines it โ€” at Science Counter Inc โ€” runs back to 1996. This is not a coincidence of timing. It is the natural consequence of working from first principles before markets exist to validate the direction.

Priority dates ยท Science Counter Inc
1996
Canadian patent CA 2186817 โ€” photonic pattern concept. Unique participation signatures identifying communication channels via linear operations on multiplexed signals. The first documented instance of the participation matrix concept.
2007
First AI framework patent filed. Participation matrix theory formalised. Conditional occurrence probability, association strength measures, causal knowledge representation. Filed before modern AI had names for these structures.
2011
sciencecounter.org deployed โ€” live conversational knowledge discovery service using the PM framework. Retrieval-Augmented Generation as a working product, nine years before the term RAG was coined.
2019
State Navigation patent filed โ€” July 21, 2019. Causal AI framework for physical environment understanding. The intelligence layer of the Physical AI platform.
2022
physical-ai.com registered โ€” June 2022. Two and a half years before the category was publicly named.
2024
ISS Platform filed โ€” October 2024. Intelligent Surround Sensing โ€” dual-mode LiDAR with no moving parts, TDE + CFSRP optical steering, PMkl native output. The sensing layer of the Physical AI platform.
Jan 2025
Jensen Huang names "Physical AI" at CES. The market arrived at the theory. The theory was already protected.

05

Why the distinction matters โ€”
for investors, engineers, and partners

The difference between a sensor connected to an AI and a Physical AI system derived from first principles is not incremental. It is architectural.

A connected system has an irreducible interface. You can optimise it, compress it, reduce its latency โ€” but you cannot eliminate it, because the two systems were designed independently with different mathematical foundations. The interface is a permanent tax on performance, on power consumption, on system complexity, and on reliability.

A derived system has no interface to optimise, because it was never introduced. The ISS LiDAR platform produces PMkl natively. The State Navigation AI reasons over PMkl natively. The same mathematical object flows from the physical world into the intelligence layer without transformation. This is not an engineering choice. It is a theoretical consequence of building both from the same first principles.

For investors: this is the moat. Not the product, not the patent โ€” the derivation. A competitor can build a better LiDAR. A competitor can build a better AI model. A competitor cannot replicate a unified theory that took thirty-five years to develop, because the theory is the foundation, not the feature.

For engineers: the ISS Platform has no moving parts, dual-mode steering (TDE + CFSRP), 100 kHz+ frame rate, and is embeddable in vehicle glass. The CFSRP sensing modality has no commercial competitor.

For licensing partners: 20+ issued patents across five families. Priority dates from 2007 to 2024. The framework covers AI knowledge representation, causal reasoning, language modelling, and physical sensing โ€” as a unified body of IP.

"Physical AI is not a product category. It is the recognition that the physical world and the intelligence that understands it were never separate to begin with."