**Moravec’s Revenge** Many are predicting that 2026 will be the year "Physical AI" finally breaks through. To understand why many are echoing, we have to look at where our current tech has plateaued and where the new bottlenecks have formed. 1. Desktop and mobile ecosystems have pretty much matured. While we haven't "solved" these platforms—and LLM agents are arguably introducing a new layer of complexity and more problems to solve—we are unlikely to see radical new interaction patterns or entirely new categories of data emerge from them. Ironically, the "next frontier" for high-tech is the low-tech physical world. 2. Massively scaled transformers have lowered the barrier to entry for robotic tasks. We are finally seeing systems that are more capable and generalizable than before. This isn't because of breakthroughs in motors or gears, but because we’ve developed better models for predicting physical sequences. And following the LLM trend, we (as a field) are probably going to push a lot on "agentic" robot intelligence, and it's a safe bet for progress. 3. Two distinct engineering philosophies exist to bridge the gap between digital intelligence and physical action: a) The Robotics Path: This approach attempts to build autonomous control from the ground up. Thanks to large-scale models, injecting high-level "intelligence" into hardware has become the manageable part. The true barrier is a modern Moravec’s Paradox: while the "brain" can now reason, the mechanical realities of manipulation remain stubbornly difficult. Mastering the nuanced, fluid motor control of a human remains a significant engineering hurdle. b) The AR/VR Path: This approach solves the manipulation problem by leveraging the human body as the engine. Here, the human provides the motor skills and real-world reasoning, while the device provides the digital augmentation. The challenge here is the form factor: fitting enough processing power and battery life into a lightweight wearable to support seamless overlays is a) an engineering challenge b) doesn't fit most user contexts (yet). It is a bet on the "Augmented Human"—the idea that a human plus a device is greater than the human alone, which we've seen before with mobile devices, hence figuring out what novel capability leap the "augmentation" brings will be key. This path is also bounded by biology; an AR headset doesn't give you the strength to lift 400 lbs. 4. Other Important Bottlenecks: Infrastructure and Power Beyond the "sim-to-real" gap, "Physical AI" faces two brutal physical constraints. First is infrastructure: unlike software, we cannot easily "rebuild" a city or a factory to accommodate robots. Second is power density: a cloud-based AI has access to virtually infinite energy, but a brain inside a bipedal robot is on a two-hour timer. Historically, breakthroughs occur when a technology turns a trade-off into a non-trade-off. We'll see how things pan out this year.