The Wheel, the Cart, and AI Systems
Systems thinking, Robotics, Artificial intelligence, and the real world
The wheel was a great invention. But not until it was combined with other wheels to create a usable cart was it an innovation. The wheel was a breakthrough; a moving, stable cart was a system.
Systems create intelligent, scalable, and disruptive technology. Innovations are not new technologies. Breakthroughs are necessary, but it’s systems that are the solution.
AI is the Wheel
This analogy applies to AI. Artificial intelligence is advancing quickly. Models grow larger. Benchmarks fall. Capabilities compound. Every few months, the frontier shifts again, reinforcing the intuition that intelligence itself is becoming cheap, abundant, and increasingly general.
From this perspective, the conclusion seems obvious: if intelligence is scalable, real-world applications should scale easily. Robots should get smarter. Factories should become autonomous. Infrastructure should self-manage.
This intuition is profoundly misleading. Scalable intelligence is the equivalent of the wheel. Physical intelligence requires systems thinking and is a new class of problem. We are given the wheel (artificial intelligence), and now we need to design the vehicle (physical intelligence and systems thinking).
We Need a System
The future of AI in the physical world will not be defined by humanoids replacing humans. It will be determined by narrow dominance in economically meaningful domains.
The problem is not that artificial intelligence is insufficient. The problem is that physical intelligence is not software intelligence with motors attached. It is a fundamentally different class of problem governed by physics, materials, energy, safety, regulation, capital cycles, and failure modes that do not yield to scaling laws—software scales at the speed of electrons. Hard assets scale at the speed of atoms, supply chains, and fatigue.
The next decade will be defined by how well capital, talent, and institutions internalize this distinction. The opportunities are enormous, but we do not have simple, scalable solutions to capture them.
The failure rate will be equally large. Understanding why and where the asymmetry lies is the difference between compounding advantage and expensive disappointment.
The System Is the Product
The most persistent error in robotics and AI-enabled hard tech is the belief that intelligence is the product.
This belief shows up everywhere: in pitch decks that emphasize model performance over system reliability; in roadmaps that assume autonomy can be “added later”; in demos where perception and planning look impressive, but actuation, energy, and recovery are quietly abstracted away.
In the physical world, intelligence is not something you deploy. It is something that emerges.
It emerges from the continuous coupling of sensing, actuation, control, energy management, mechanical design, fault tolerance, and human supervision. Break that coupling, and performance collapses. Sensors drift. Actuators wear. Environments violate assumptions. Latency matters. Failure is routine.
This is why physical intelligence must be understood as a system, not a software artifact. Robotics fails because a model is smart. It succeeds when the entire stack—mechanical, electrical, computational, and operational—has been co-designed around real-world constraints.
Otherwise, it is a research project, not a product.
Software in the Physical World
Software developers write most AI roadmaps. That intuition becomes a liability when applied uncritically to hard assets.
Software rewards abstraction, iteration speed, and reversibility. Robotics punishes all three.
Hardware iteration is slow and destructive. Capital is committed long before product-market fit is proven. Abstraction leaks everywhere. Failures propagate physically. You cannot roll back a broken actuator, a cracked gearbox, or an injured worker.
This creates a structural mismatch between expectations and outcomes. Teams plan on software timelines while operating on hardware clocks. Investors expect linear progress while reality delivers long periods of stagnation punctuated by sudden breakthroughs. Milestones are defined by demos rather than by uptime, maintenance cost, or safety certification.
The result is familiar: promising companies stall somewhere between prototype and deployment, not because the technology “didn’t work,” but because the system was never designed to survive physical operations.
The Lab and Reality
- A robot that works in a lab proves feasibility.
- A factory robot demonstrates resilience.
- A robot that works profitably proves relevance.
Most never reach the third stage.
Dust, vibration, thermal variation, partial observability, human interference, nonlinear wear, and regulatory constraints dominate. Intelligence that looks impressive in isolation often degrades rapidly.
This is why robotics progress appears nonlinear. Long periods of apparent stagnation are followed by rapid adoption—but only within narrow, highly constrained domains. Generality comes last, not first.
The key threshold is the point at which performance remains acceptable under stress, failures are bounded and recoverable, maintenance does not erase productivity gains, and human oversight decreases rather than expands.
Until then, robotics remains a cost center masquerading as innovation.
Benchmarks and Illusions
The benchmark illusion that distorts perceptions of AI capability in software is now bleeding into robotics and hard tech. Benchmarks reward test-taking, not robustness. Models can be trained to score well without improving adaptability, judgment, or failure recovery.
This is already visible in pure AI systems, where performance gains often reflect optimization for the test rather than genuine progress.
In physical systems, this distortion is dangerous.
A hallucination in a chatbot is embarrassing; in a robotic control loop, it is destructive.
Physical intelligence is validated only in the field, over time, under stress. Everything else is provisional.
Energy is Constrained
Energy is the most underestimated bottleneck in robotics and AI-enabled hard assets.
Every physical action consumes power. Every sensor, processor, and actuator draws current. Unlike data centers, robots cannot externalize this constraint. They carry it with them.
More sensing improves robustness but increases power draw. More computation improves planning but generates heat. More actuation improves capability but reduces endurance. These tradeoffs do not vanish with better models.
This is why many visually impressive robots fail in deployment. They optimize for peak capability rather than sustained operation. In real environments, uptime beats elegance.
Physical intelligence, therefore, includes energy intelligence: prioritization under constraint, graceful degradation, and deliberate tradeoffs between performance and survivability. Systems that ignore this are optimized for demonstrations, not for work.
Learning Is Expensive
Learning in physical systems is slow, expensive, and risky.
Data collection breaks hardware. Mistakes cost money. Environments do not remain constrained. Unlike software, errors are not free.
Software models are not useless. Model-based control governs safety. Learning handles adaptation, perception, and edge cases. Simulation accelerates early development.
The vision of fully autonomous, self-improving physical systems is not a fantasy. But it is not imminent, cheap, or general. Confusing computational intelligence with physical intelligence leads to overconfidence and misallocated capital.
The Real Opportunities
We have the wheel, and there are exciting new vehicles to be designed and built.
Below are examples of AI in the physical world where systems thinking applied to narrowly defined, yet economically significant, opportunities can deliver substantial benefits.
1. Advanced Industrial Robotics and Automation
This is the least glamorous and most durable opportunity set.
Industrial manipulation in fixed or semi-structured environments—such as assembly, welding, inspection, packaging, and quality control—remains underpenetrated. The challenge is not intelligence alone; it is reliability, calibration, maintenance, and integration with legacy systems.
AI adds value where:
- Variability exceeds the limits of classical automation
- Changeover costs are high
- Human labor is scarce, expensive, or unsafe
The winners will not be general-purpose robots. They will be systems optimized for specific workflows, with intelligence deeply embedded in mechanical and operational design.
2. Logistics, Warehousing, and Material Handling
Logistics is a physical optimization problem masquerading as a software problem.
Autonomous forklifts, pallet movers, picking systems, yard management robots, and port automation represent enormous opportunities—not because the tasks are intellectually complex, but because throughput, uptime, and safety have clear economic value.
AI’s role here is coordination, perception under clutter, and adaptive routing—not human-level reasoning. Success depends on integration with infrastructure, not autonomy in isolation.
3. Autonomous Systems Beyond Cars
Autonomy will not arrive everywhere at once. It will come in environments where constraints are strong and incentives are strong.
Examples include:
- Autonomous mining equipment
- Agricultural machinery in controlled fields
- Industrial drones for inspection and maintenance
- Autonomous vessels in defined maritime corridors
The common thread is bounded complexity, high asset utilization, and strong economic incentives to reduce risk and labor exposure.
4. Energy Infrastructure and Industrial Systems
Power is prosperity. Energy is the backbone of every physical economy, and it is increasingly complex.
AI-enabled hard tech opportunities include:
- Predictive maintenance for grids, pipelines, and power plants
- Autonomous inspection of transmission lines and offshore assets
- Optimization of energy storage, microgrids, and load balancing
- Robotics for hazardous energy environments (nuclear, offshore wind, geothermal)
These systems demand extreme reliability. Intelligence is valuable only to the extent that it improves uptime, safety, and lifecycle economics.
5. Construction and Infrastructure
Construction remains one of the least digitized industries on earth.
Opportunities exist in:
- Robotic site preparation and material handling
- Autonomous heavy equipment
- Modular and automated construction systems
- Inspection and maintenance of bridges, tunnels, and urban infrastructure
Here, AI must coexist with human labor, regulatory constraints, and highly variable environments. Incremental productivity gains compound into a massive economic impact.
6. Defense, Security, and Hazardous Operations
Defense and security are early adopters and some of the largest and fastest-growing opportunities. There is too much at stake, and the cost of failure is asymmetric.
Applications include:
- Autonomous logistics and resupply
- Robotics for explosive ordnance disposal
- Surveillance and perimeter security systems
- Operations in denied or dangerous environments
These systems tolerate narrow specialization in exchange for reliability under extreme conditions.
Unfortunately, these are increasingly battle-tested in the real world, both saving and inflicting human casualties.
7. Medical and Life-Science Robotics
Medical robotics is not about replacing clinicians. It is about extending precision, consistency, and endurance. It also enables orders-of-magnitude improvements in laboratory performance, systemization, and consistency in testing and verifying clinical outcomes.
Examples include:
- Surgical robotics with AI-assisted planning
- Lab automation and sample handling
- Rehabilitation and assistive devices
- Hospital logistics and sterilization systems
Regulatory requirements slow deployment, but patent and IP protection are much more effective and long-lived. Ultimately, innovation is given competitive defensibility.
8. Hard Tech as the Natural Habitat for Applied AI
Hard tech forces intelligence to confront reality.
Unlike pure software, it demands integration with materials, energy, regulation, and human workflows. It requires capital, patience, and interdisciplinary talent. It rewards system builders over model optimizers.
For investors, this is precisely why the opportunity is attractive. Barriers to entry are real. Moats are physical, operational, and institutional. Returns accrue to those willing to play a long game.
Be Specific
The value created by AI in the physical world is not scaling software intelligence. It is focus, discipline, and constraint via effective systems design and performance. The systems that endure will not be those that promise universality, but those that dominate specific economic niches, involve humans strategically, and survive year ten of operation.
The opportunity is enormous. The constraints are unforgiving.
Navigating both is the real intelligence test—and the defining challenge for the next generation of builders and investors. We can celebrate the invention of the wheel, but it’s not until we had four wheels working together in an effective system that real value was captured.
