Author: Oliver Hsu (a16z)

Translation: Deep Tide TechFlow

Deep Tide Guide: This article comes from a16z researcher Oliver Hsu and represents the most systematic "physical AI" investment map since 2026. His judgment is: the main line of language/code is still scaling, but the areas that can truly yield the next generation of disruptive capabilities are three adjacent fields—general robots, autonomous science (AI scientists), and brain-machine interfaces. The author dissects the five underlying capabilities that support them and argues that these three fronts will form a mutually reinforcing structural flywheel. For those who want to understand the investment logic of physical AI, this is currently the most complete framework.

The current paradigm dominating AI is organized around language and code. The scaling laws of large language models have been clearly characterized, and the business flywheel of data, computing power, and algorithm improvements is turning. The returns from each increase in capability are substantial and mostly visible. This paradigm deserves the capital and attention it attracts.

However, another group of adjacent fields has made substantial progress during their gestation period. This includes general robot routes such as VLA (visual-language-action models) and WAM (world action models), physical and scientific reasoning around "AI scientists," and new interfaces utilizing AI advancements to reshape human-machine interaction (including brain-machine interfaces and neurotechnologies). Besides the technologies themselves, these directions are attracting talent, capital, and founders. The technical primitives extending frontier AI into the physical world are maturing simultaneously, and progress in the past 18 months indicates that these fields will soon enter their respective scaling phases.

In any technical paradigm, the areas where the current capabilities and mid-term potentials have the largest delta often exhibit two characteristics: first, they can benefit from the same growing scaling dividends that drive the current frontier; second, they are a step away from the mainstream paradigm—close enough to inherit its infrastructure and research momentum, yet far enough to require substantial additional work. This distance itself has a dual effect: it naturally forms a moat against rapid followers while defining a problem space that is less crowded and less saturated with information, making it more likely for new capabilities to emerge—precisely because shortcuts have not yet been fully utilized.

Caption: The current AI paradigm (language/code) and the relationship with adjacent frontier systems

Today, there are three fields that meet this description: robotic learning, autonomous science (especially in materials and life sciences), and new human-machine interfaces (including brain-machine interfaces, silent voice, neuro-wearables, and new sensory channels like digitized olfaction). They are not entirely independent works; thematically, they belong to the same group of "frontier systems in the physical world." They share a set of underlying primitives: learning representations of physical dynamics, architectures oriented towards embodied actions, simulation and synthetic data infrastructures, continuously expanding sensory channels, and closed-loop agent orchestration. They reinforce each other in a cross-domain feedback relationship. They are also the most likely places to emerge transformative capabilities—products of the interactions between model scale, physical grounding, and new data forms.

This article will outline the technological primitives supporting these systems, explain why these three fields represent leading opportunities, and propose that their mutual reinforcement forms a structural flywheel that pushes AI into the physical world.

Five Underlying Primitives

Before looking at specific applications, we first need to understand the technological foundation shared by these frontier systems. Pushing frontier AI into the physical world relies on five major primitives. These technologies are not exclusive to any single application domain; they are components—enabling systems that "extend AI into the physical world." Their simultaneous maturation is what makes this moment special.

Caption: Five underlying primitives supporting physical AI

Primitive One: Learning Representations of Physical Dynamics

The most fundamental primitive is the ability to learn a compressed, universal representation of behaviors in the physical world—how objects move, deform, collide, and respond to forces. Without this layer, every physical AI system would have to learn the physical laws of its domain from scratch, an unfeasible cost.

Several architectural schools are approaching this goal from different directions. The VLA model tackles it from the top: using pre-trained visual-language models—these models have a semantic understanding of objects, spatial relationships, and language—and adding an action decoder to output motion control commands. The key point is that the substantial cost involved in learning to "see" and "understand the world" can be diluted by internet-scale text-image pre-training. Physical Intelligence’s π₀, Google DeepMind’s Gemini Robotics, and Nvidia's GR00T N1 have increasingly validated this architecture at larger scales.

The WAM model approaches it from the bottom: based on video diffusion transformers pre-trained on internet-scale video data, inheriting rich prior knowledge of physical dynamics (how objects fall, how they are occluded, how they interact under force), and coupling these priors with action generation. Nvidia's DreamZero demonstrates zero-shot generalization to entirely new tasks and environments, achieving meaningful improvements in real-world generalization from a small amount of adaptation data based on human video demonstrations.

The third route may provide the most insights into future directions as it skips the entire VLM and video diffusion backbone. Generalist's GEN-1 is a natively embodied foundational model trained from scratch, with training data consisting of over 500,000 hours of real physical interaction data, mostly collected from individuals performing daily operational tasks through low-cost wearable devices. It is not a standard VLA (it does not have a visual-language backbone being fine-tuned), nor is it WAM. Instead, it is a foundational model designed specifically for physical interaction, starting from zero and learning not the statistical patterns of internet images, texts, or videos, but the statistical patterns of interactions between humans and objects.

Companies like World Labs that focus on spatial intelligence are important for this primitive because they fill a gap that VLA, WAM, and native embodied models have in common: none have explicitly modeled the three-dimensional structure of the scenes they operate in. VLA inherits 2D visual features from text-image pre-training; WAM learns dynamics from videos, which are 2D projections of the 3D reality; models trained on wearable sensor data can capture forces and kinematics but not the geometry of the scenes. Spatial intelligence models can help fill this gap—learning to reconstruct and generate complete 3D structures of physical environments and reasoning about them: geometry, lighting, occlusion, object relationships, and spatial layouts.

The convergence of all the routes is crucial in itself. Regardless of whether the representations are inherited from VLM, learned through collaborative training with videos, or built from scratch using physical interaction data, the underlying primitive is the same: a compressed, transferable model of physical world behaviors. The data flywheel these representations can tap into is massive and largely untouched—not just internet videos and robotic trajectories, but also the vast sea of human bodily experience data that wearable devices are beginning to collect at scale. The same representation can serve a robot that is learning to fold towels, an autonomous laboratory that predicts reaction outcomes, and a neural decoder interpreting the grasp intentions of the motor cortex.

Primitive Two: Architecture for Embodied Actions

Having physical representations alone is not enough. Translating "understanding" into reliable physical actions requires an architecture to address several interrelated issues: mapping high-level intentions to continuous motion command sequences, maintaining consistency over long action sequences, operating under real-time latency constraints, and improving through experience.

A dual-system layered architecture has become the standard design for complex embodied tasks: a slow yet powerful visual-language model handles scene understanding and task reasoning (System 2), coupled with a fast and lightweight visual-motor policy responsible for real-time control (System 1). GR00T N1, Gemini Robotics, and Figure's Helix all adopt variants of this route to resolve the fundamental tension between "large models providing rich reasoning" and "physical tasks requiring millisecond-level control frequencies." Generalist has taken a different approach, implementing "resonant reasoning" to allow thought and action to occur simultaneously.

The action generation mechanisms are also rapidly evolving. The action head based on flow matching and diffusion, pioneered by π₀, has become the mainstream method for generating smooth, high-frequency continuous actions, replacing the discrete tokenization borrowed from language modeling. Such methods treat action generation as a denoising process similar to image synthesis, producing trajectories that are physically smoother and more robust to error accumulation than autoregressive token predictions.

However, the most critical architectural advancement may be extending reinforcement learning to pre-trained VLA—an underlying model trained on demonstration data that can continue to improve through autonomous practice, much like a person hones a skill through repeated practice and self-correction. Physical Intelligence’s π*₀.₆ work is the clearest large-scale demonstration of this principle. Their method is called RECAP (Experience and Correction Reinforcement Learning based on Advantage-Conditioned Policy), addressing the long-sequence credit assignment problem that pure imitation learning struggles with. If a robot grasps the handle of an espresso machine at a slightly tilted angle, failure may not manifest immediately—it might only become evident several steps later when attempting to insert the machine. Imitation learning has no mechanism to attribute that failure back to the earlier grasp, but reinforcement learning does. RECAP trains a value function to estimate the probability of success starting from any intermediate state and then enables the VLA to select high-advantage actions. The key lies in integrating various heterogeneous data—demonstration data, autonomous experience from policy, and corrections provided by experts during execution—into the same training pipeline.

The results of this approach bode well for the prospects of reinforcement learning in action domains. π*₀.₆ has folded 50 unseen garments in real household environments, reliably assembled cardboard boxes, and produced espresso on commercial machines, running continuously for several hours without human intervention. In the most difficult tasks, RECAP doubled the throughput compared to a pure imitation baseline, while reducing failure rates by over half. This system has also demonstrated that post-training reinforcement learning can yield qualitatively different behaviors that imitation learning cannot achieve: smoother recovery actions, more effective grasping strategies, and self-adaptive error correction absent in demonstration data.

These gains indicate something significant: the computational scaling dynamics that pushed large models from GPT-2 to GPT-4 are just beginning to operate in embodied domains—only this time at an earlier point in the curve, dealing with action spaces that are continuous, high-dimensional, and subject to the relentless constraints of the physical world.

Primitive Three: Simulation and Synthetic Data as Scaling Infrastructure

In the language domain, the data problem has been solved by the internet: naturally generated, freely available trillions of tokens of text. In the physical world, this problem is several orders of magnitude more challenging—this is now a consensus, with the most direct signal being the rapid increase in entrepreneurial startups targeting data suppliers for the physical world. The cost of collecting real robot trajectories is high, scaling is risky, and diversity is limited. A language model can learn from a billion conversations, but a robot (temporarily) cannot have a billion physical interactions.

Simulation and synthetic data generation form the infrastructure layer to address this constraint, and their maturity is one of the key reasons why physical AI is accelerating today rather than five years ago.

The modern simulation stack combines physics-based simulation engines, ray-traced photo-real rendering, procedural environment generation, and world foundational models that produce photo-realistic videos from simulation inputs—bridging the sim-to-real gap. The entire pipeline starts with neural reconstruction of the real environment (achievable with just a smartphone), fills in physically accurate 3D assets, and progresses to large-scale synthetic data generation with automatic annotations.

The significance of the improvements in the simulation stack is that they are changing the economic assumptions supporting physical AI. If the bottleneck of physical AI shifts from "collecting real data" to "designing diverse virtual environments," the cost curve will collapse. Simulations scale with computing power, without reliance on human labor or physical hardware. This transformation of the economic structure for training physical AI systems is akin to the transformation brought about by internet text data for training language models—indicating that investment in simulation infrastructure has a tremendous leveraging effect on the entire ecosystem.

But simulation is not just a robotic primitive. The same infrastructure supports autonomous science (digital twins of laboratory equipment, simulation environments for hypothesis pre-screening), new interfaces (simulated neural environments for training BCI decoders, synthetic sensory data for calibrating new sensors), and other fields where AI interacts with the physical world. Simulation is the universal data engine for physical world AI.

Primitive Four: Expanding Sensory Channels

The signals that convey information in the physical world go far beyond visual and linguistic inputs. Tactile senses convey material properties, grip stability, contact geometries, and other information that cameras cannot see. Neural signals encode movement intentions, cognitive states, and sensory experiences at bandwidths unmatched by any existing human-machine interface. Subglottal muscle activity encodes speech intentions even before any sounds are produced. The fourth primitive involves AI rapidly expanding sensory pathways to these previously hard-to-reach modalities—not only from research but also driven by an entire ecosystem building consumer-grade devices, software, and infrastructure.

Caption: Expanding AI sensory channels, from AR, EMG to brain-machine interfaces

The most intuitive indicator is the emergence of new categories of devices. AR devices have seen significant improvements in both experience and form factor in recent years (with companies already building consumer and industrial applications on this platform); voice-first AI wearables are providing language-based AI with a more complete physical contextual understanding—they genuinely follow users into physical environments. In the long run, neural interfaces might unlock even more comprehensive interaction modalities. The computational transformations brought by AI create an opportunity for a significant upgrade in human-machine interaction, with companies like Sesame building new modalities and devices for this purpose.

This more mainstream modality of voice also benefits emerging interaction methods. Products like Wispr Flow are promoting voice as the primary input method (due to its high information density and inherent advantages), and the market conditions for silent voice interfaces are improving concurrently. Silent voice devices capture tongue and vocal cord movements using various sensors to recognize language—representing a human-machine interaction mode with higher information density than verbal communication.

Brain-machine interfaces (both invasive and non-invasive) represent a deeper frontier, and their associated business ecosystems continue to progress. Signals will emerge at the convergence of clinical validation, regulatory approval, platform integration, and institutional capital—this is a technology category that was purely academic just a few years ago.

Tactile perception is entering embodied AI architectures, with some models in robotic learning starting to explicitly incorporate touch as a first-class citizen. Olfactory interfaces are becoming tangible engineering products: wearable olfaction display devices with micro-odor generators and millisecond response times have been demonstrated in mixed reality applications; olfactory models are also beginning to pair with visual AI systems for chemical process monitoring.

The common pattern in these developments is that they converge under extremes. AR glasses continuously generate visual and spatial data about user interactions with the physical environment; EMG wristbands capture statistical patterns of human movement intentions; silent voice interfaces recognize the mapping from subglottal phonation to language output; BCIs capture neural activity at the highest current resolution; and tactile sensors capture the contact dynamics of physical operations. Each new category of device is also a data generation platform that feeds multiple application domains' underlying models. A robot trained using EMG-inferred motion intention data will learn grasping strategies different from a robot trained solely on remote operation data; an experimental interface responding to subglottal commands will yield entirely different scientist-machine interaction dynamics compared to one controlled by keyboard; a neural decoder trained with high-density BCI data produces motion planning representations unattainable through any other channel.

The diffusion of these devices is expanding the effective dimensionality of the available data streams for training frontier physical AI systems—and a significant portion of this expansion is driven by capital-rich consumer goods companies rather than solely coming from academic laboratories, meaning the data flywheel can expand alongside market adoption rates.

Primitive Five: Closed-Loop Agent Systems

The final primitive leans more towards the architectural aspect. It refers to the integration of perception, reasoning, and action into a continuously operating, autonomous, closed-loop system that functions over extended time dimensions without human intervention.

In language models, the corresponding development is the rise of agent systems—multi-step reasoning chains, tool usage, and self-correction processes that elevate models from single-turn Q&A tools to autonomous problem solvers. The same transformation is happening in the physical world, albeit with much harsher requirements. A language agent can afford to rollback errors at no cost; if a physical agent spills a bottle of reagent, there is no going back.

Physical world agent systems have three characteristics that distinguish them from their digital counterparts. First, they need to embed experiments or operations in a closed loop: directly interfacing with raw instrumentation data streams, physical state sensors, and execution primitives, making reasoning grounded in physical reality rather than its text descriptions. Second, they require long-sequence persistence: memory, traceability, safety monitoring, recovery behaviors, linking multiple operational cycles together rather than treating each task as a standalone episode. Third, they need closed-loop adaptability: strategies revised based on physical outcomes rather than merely text-based feedback.

This primitive integrates individual capabilities (well-functioning world models, reliable action architectures, comprehensive sensor suites) into a complete system capable of autonomous operation in the physical world. It serves as the integration layer, and its maturity is a prerequisite for the three application domains discussed below to exist as real-world deployments rather than isolated research demonstrations.

Three Domains

The primitives outlined above are universal enabling layers; they do not specify which applications will be the most important. Many domains involve physical actions, physical measurements, or physical perceptions. What distinguishes "frontier systems" from merely improved existing systems is the degree of compounding that model capabilities and scaling infrastructures undergo within the domain—not only in terms of better performance but also in the emergence of new capabilities that were previously unattainable.

Robotics, AI-driven science, and new human-machine interfaces are the three domains with the strongest compounding effects. Each assembles the primitives in a unique manner, each is constrained by the limitations that current primitives are dismantling, and each will generate structured physical data as by-products during operation—this data, in turn, enhances the primitives themselves, forming feedback loops and accelerating the entire system. They are not the only physical AI domains worth focusing on, but they are the most densely populated places for frontier AI capabilities to interact with physical realities, and they hold the greatest potential for new capability emergence, while also being highly complementary and able to benefit from the dividends of the current language/code paradigm.

Robotics

Robotics is the most literal embodiment of physical AI: an AI system that must perceive in real time, reason, and apply physical actions to the material world. It also tests each primitive's rigorously.

Think about how many tasks a general-purpose robot must perform to fold a towel. It needs a learned representation of how deformable materials behave under force—a physical prior that language pre-training cannot provide. It requires an action architecture that can translate high-level commands into continuous motion command sequences at over 20Hz control frequencies. It depends on training data generated from simulations since no one has collected millions of real towel-folding demonstrations. It requires tactile feedback to detect slips and adjust grip strength because vision cannot distinguish between a secure grip and a failed one. It also requires a closed-loop controller capable of recognizing errors and recovering when folding goes wrong instead of blindly executing stored trajectories.

Caption: Robotic tasks simultaneously invoking the five underlying primitives

That's why robotics is a frontier system, not a more sophisticated version of a mature engineering discipline. These primitives don't just improve existing robotic capabilities; they unlock operations, movements, and interaction categories that weren't possible outside of tightly controlled industrial environments.

Significant frontier advancements have been made in recent years—we've written about this before. The first generation of VLA has demonstrated that foundational models can control robots to perform diverse tasks. Architectural advancements are bridging the high-level reasoning and low-level control in robotic systems. Side reasoning has become feasible, and cross-ontology transfer means a model can adapt to an entire new robotic platform with limited data. The remaining core challenge is achieving scalable reliability, which remains the deployment bottleneck. A 95% success rate at each step results in only a 60% success rate over a ten-step task chain, while production environments require far more reliability. Post-training reinforcement learning holds substantial potential here to help this field cross the capacity and robustness thresholds needed for scaling.

These advancements affect market structures. For decades, value in the robotics industry has been embedded in the mechanical systems themselves; mechanics remain a crucial part of the technical stack, but as learning strategies become more standardized, value will migrate towards models, training infrastructures, and data flywheels. Robotics also feeds back into the aforementioned primitives: every real-world trajectory serves as training data for improving world models, every deployment failure exposes gaps in simulation coverage, and every new ontology tested broadens the diversity of physical experience available for pre-training. Robotics serves as both the most demanding consumer of the primitives and one of their most significant sources of improvement signals.

Autonomous Science

If robotics is about testing primitives through "real-time physical actions," autonomous science measures something slightly different—continuous multi-step reasoning for causally complex physical systems, over time spans measured in hours or days, where experiment results need to be interpreted, contextualized, and used to revise strategies.

Caption: How autonomous science (AI scientists) integrates the five underlying primitives

AI-driven science is the most thorough domain for combining the primitives. An automated lab (self-driving lab, SDL) requires learned representations of physical chemistry dynamics to predict what experiments will produce; embodied actions to pipette, position samples, and operate analytical instruments; simulations for candidate experiment pre-screening and allocation of scarce instrument time; and expanded sensing capabilities—spectroscopy, chromatography, mass spectrometry, and increasingly new chemical and biological sensors—to characterize results. It requires the closed-loop agent orchestration primitive more than any other field: capable of maintaining a multi-round "hypothesis-experiment-analysis-revision" workflow without human intervention, preserving traceability, monitoring safety, and adjusting strategies based on information revealed in each round.

No other field invokes these primitives as deeply. This is why autonomous science is a frontier "system" rather than merely an improved laboratory automation process. Companies such as Periodic Labs and Medra synthesize scientific reasoning capabilities and physical validation capabilities in materials and life sciences, respectively, achieving scientific iteration while producing experimental training data.

The value of such systems is intuitively clear. Traditional material discovery can take years from concept to commercialization, but AI-driven workflow acceleration theoretically can compress this process to far less time. The key constraints are shifting from hypothesis generation (which can be well-assisted by foundational models) to manufacturing and validation (which require physical instruments, robotic execution, and closed-loop optimization). SDL targets this bottleneck directly.

Another important characteristic of autonomous science—valid across all physical world systems—is its role as a data engine. Each experiment run by an SDL generates not only a scientific result but also a physically grounded, experimentally validated training signal. A measurement of how polymers crystallize under specific conditions enriches the world model’s understanding of material dynamics; a validated synthetic pathway becomes training data for physical reasoning; an identified failure informs the agent system of where its predictions have failed. Data produced from a real experiment by an AI scientist is qualitatively different from that from internet text or simulation outputs—it is structured, causal, and empirically validated. Autonomous science provides a direct pathway to transform physical realities into structured knowledge, improving the entire physical AI ecosystem.

New Interfaces

Robotics extends AI into physical actions, autonomous science extends AI into physical research. New interfaces extend it into the direct coupling of AI with human perception, sensory experiences, and bodily signals—devices ranging from AR glasses and EMG wristbands to implanted brain-machine interfaces. What binds this category together is not a single technology but a common function: expanding the bandwidth and modalities of the channels between human intelligence and AI systems—and in the process generating human-world interaction data directly useful for constructing physical AI.

Caption: The lineage of new interfaces from AR glasses to brain-machine interfaces

The distance from the mainstream paradigm is both a challenge and a potential for this field. Language models conceptually understand these modalities but are not naturally familiar with the movement patterns of silent voice, the geometric structures of odor receptor bindings, or the temporal dynamics of EMG signals. Decoding the representations of these signals must be learned from the expanding sensory channels. Many modalities do not have internet-scale pre-training data; data often can only be generated from the interfaces themselves—meaning that systems and their training data are co-evolving, a phenomenon absent in language AI.

Recent developments in this domain have shown a rapid rise of AI wearables as consumer products. AR glasses may be the most visible example of this category, while other wearables with speech or vision as primary inputs are also emerging concurrently.

This consumer device ecosystem provides a new hardware platform for AI to extend into the physical world and is also becoming the infrastructure for capturing physical world data. A person wearing AI glasses can continuously produce first-person video streams about how humans navigate physical environments, manipulate objects, and interact with the world; other wearables continuously capture biometric and motion data. The installation base of AI wearables is transforming into a distributed network for collecting physical world data at previously impossible scales—think of the scale of smartphones as consumer devices—a new category of consumer device is allowing computing to perceive the world in new modalities and opens a massive new channel for AI interactions with the physical world.

Brain-machine interfaces represent a deeper frontier. Neuralink has implanted multiple patients, with surgical robots and decoding software iterating on the process. Synchron's endovascular Stentrode has been used to enable paralyzed users to control digital and physical environments. Echo Neurotechnologies is developing a BCI system for language recovery, based on their research in high-resolution cortical speech decoding. New companies like Nudge are also forming, gathering talent and capital to build new neural interfaces and brain interaction platforms. Technical milestones at the research level are worth noting: the BISC chip has demonstrated wireless neural recording with 65,536 electrodes on a single chip; the BrainGate team has directly decoded internal language from the motor cortex.

The common thread running through AR glasses, AI wearables, silent voice devices, and implanted BCIs is not merely "they are all interfaces," but that they collectively form an increasing-bandwidth spectrum between human physical experience and AI systems—each point on the spectrum supports the continuous advancement of the primitives behind the three major domains of this article. A robot trained with high-quality first-person video data from millions of AI glasses users will learn operational priors that differ greatly from a robot trained only on filtered remote operation datasets; a lab AI responding to subglottal directives and a keyboard-controlled lab will yield vastly different dynamics in terms of latency and smoothness; and a neural decoder trained with high-density BCI data will produce motion planning representations that no other channel can achieve.

New interfaces are the mechanisms by which sensory channels themselves expand—they open previously nonexistent data channels between the physical world and AI. This expansion is being driven by consumer device companies pursuing scalable deployments, indicating that the data flywheel will accelerate alongside consumer adoption.

Systems in the Physical World

Viewing robotics, autonomous science, and new interfaces as different instances of the same set of primitives forming frontier systems is warranted because they enable each other and compound.

Caption: The feedback flywheel between robotics, autonomous science, and new interfaces

Robotics enables autonomous science. The automated laboratory is essentially a robotics system. The operational capabilities developed for general robots—dexterous grasping, liquid handling, precise positioning, multi-step task execution—can be directly transferred to laboratory automation. As robotic models progress in generality and robustness, the scope of experimental protocols that SDL can execute autonomously expands. Every advance in robotic learning reduces the costs of autonomous experimentation while increasing throughput.

Autonomous science enables robotics. The scientific data produced by the automated laboratory—verified physical measurements, causal experimental results, material property databases—provides the structured, grounded training data that the world model and physical reasoning engine need most. Furthermore, the materials and components required by next-generation robots (better actuators, more sensitive tactile sensors, higher-density batteries, etc.) are inherently the products of materials science. Accelerated autonomous discovery platforms improving materials innovation directly advance the hardware foundations where robotic learning operates.

New interfaces enable robotics. AR devices are scalable methods for collecting data on "how people perceive and interact with the physical environment." Neural interfaces produce data regarding human motion intentions, cognitive planning, and sensory processing. This data is invaluable for training robotic learning systems, especially for tasks involving human-robot collaboration or remote operation.

There is a deeper observation about the nature of AI advancements at the frontier. The language/code paradigm has produced extraordinary results and continues to rise strongly in the scaling era. However, the new problems, data types, feedback signals, and evaluation criteria presented by the physical world are almost infinite. Embedding AI systems in physical realities—through robots manipulating objects, laboratories synthesizing materials, and interfaces docking into biological and physical realms—opens up a new scaling axis that is complementary to the existing digital frontier and likely to mutually enhance one another.

Caption: Interactions and emergence across different scaling axes of physical AI

What behaviors these systems will ultimately exhibit is hard to predict accurately—the definition of emergence comes from independent capabilities that are understandable but unprecedented in their combinations. However, historical patterns tend to be optimistic. Each time an AI system gains a new modality for interacting with the world—seeing (computer vision), speaking (speech recognition), reading and writing (language models)—the capability leap far exceeds the sum of improvements in each area. The transition to systems within the physical world represents the next such phase transition. In this sense, the primitives discussed in this article are being assembled at this moment, potentially enabling frontier AI systems to perceive, reason, and act in the physical world, unlocking substantial value and progress therein.

Disclaimer: This article is for informational exchange only and does not constitute any investment advice. It should not be used as a basis for legal, business, investment, or tax consultation.