AI has an emergent problem
Actually, it has a problem with emergence
The challenge for intelligent machines is dealing with the emergent properties of the physical world. Traffic jams, crop failures, supply chains, ecosystems, or human crowds. These systems behave in ways that cannot be predicted from the individual components and simple rules. Here is where traditional AI strategies break down. Enter Physical AI.
Physical AI becomes interesting the moment we imagine a robot or autonomous system trying to rely on a set of rules and hope for the best. Think of it this way. If an autonomous system has one way to succeed, by definition it has many ways to fail. Therefore, robots and autonomous systems need an imaginative toolkit to deal with uncertainty and unpredictability. One of them may be is bet hedging. This means keeping multiple strategies alive at once instead of overcommitting to one locally-optimal behavior. Instead of trying to pick the perfect move, bet hedging spreads risk across several possible moves so that if conditions change unexpectedly, at least one will always still work. Perhaps a good analogy is packing for a trip when the weather is unstable. If you bring only beach clothes, you may fail badly if it rains or is unexpectedly cold. If you instead pack summer clothes but also a rain coat and a warm jacket, you may be less optimized for one outcome but more likely to cope with any weather overall. Another analogy is investing. Putting all your money into one stock may bring a big return, but also a big loss. A diversified portfolio, although less flashy, is much safer over time. In practical terms, it means that Physical AI system may maintain various policies simultaneously, or distribute roles across a fleet, so that something will work no matter what.
But bet hedging can only go so far, because it protects against uncertainty but cannot really reduce it. A stronger approach may be active inference, where the system acts to reduce uncertainty in its internal model of the world. Instead of merely reacting, the agent probes, senses, updates, and then acts again. This is a much better fit for emergent environments because surprises are treated as signals to learn from, not just noise to endure. Stanhope AI shows one of the clearest examples of a company working in this sphere. It explicitly states that it is taking active inference from neuroscience into AI for robots and embodied platforms. It is already testing the concept in drones and robotics where machines must adapt on the fly in unpredictable real-world settings.
A third idea is world models, which are quickly becoming the intellectual center of physical AI. World models let machines simulate possible futures before acting, which is exactly what you want when the environment contains delayed effects, hidden variables, and cascading interactions. AMI Labs is explicitly building world-model-based AI that understands the real world, with persistent memory, planning, and controllability as core claims. NVIDIA has also thrown their hat in this arena. Its Cosmos platform is described as physical AI with world foundation models, including physics-aware generation and simulation for downstream robot intelligence. The takeaway is that if a system cannot simulate, it will end up improvising.
Then there is a biological strategy called niche construction. The idea is that to handle emergence the systems needs to reshape the environment rather than to predict it, so it becomes more stable. Although this is clearly more complex, warehouses already do this for robots. Sensorized infrastructure does this for autonomous operations. The company Analog is relevant here because it describes a world model that transforms fragmented data into a continuously learning intelligence fabric, and it explicitly connects robotics to that model so systems can sense, simulate and, importantly, act in dynamic environments. Its aim at robotics-as-a-service makes it one of the most concrete examples of a company trying to operationalize physical intelligence across real environments. Analog’s work on Deccan Oculus AI City is the strongest evidence for instrumenting and structuring an environment, although not yet for large-scale physical reconfiguration.
Where do companies like Anthropic fit, you might ask? The answer is, and moment only indirectly. Anthropic is relevant to controllable and interpretable AI systems, but it is not a physical-AI company in the sense that Analog, AMI Labs, NVIDIA, Stanhope AI are.
In short, bet hedging should help with resilience, active inference for intelligent sensing, world models for simulation and planning, and niche construction for environmental stabilization. I predict that the future of physical AI will belong to systems that combine all four.