Private Physical AI for the Edge: Small, Energy-Efficient, and Everywhere
Artificial intelligence today dazzles with its fluency, creativity, and problem-solving power, but behind the curtain lies a growing paradox: The “smarter” our machines get, the “hungrier” they become for electricity. Training the largest models consumes millions of kilowatt hours, and serving billions of queries requires data centers the size of factories. If this trajectory continues, AI will become one of the world’s most energy-intensive industries, straining power grids, accelerating climate pressures, and concentrating economic and geopolitical power in the hands of those who control massive compute infrastructure.
Yet another path for AI is opening, one that moves away from everlarger, resource-hungry models and toward systems that are small, fast, adaptive, and radically more efficient. This emerging paradigm, which we call Private Physical Edge AI, refers to intelligent systems that operate directly on the edge: on the same devices that sense, decide, and act. In other words, the computation takes place locally on-device, close to where data is generated and actions occur. These systems are physically embedded in the world, grounded in physics-based computation, energy efficient, and designed to preserve privacy by keeping data on the device rather than in the cloud. Instead of being tethered to sprawling data centers, intelligence could live on the devices we already use—in our pockets, homes, and clinics; in our phones, glasses, cars, and appliances; even on devices we have not invented yet—developed natively for the device hardware, running locally on modest hardware, and consuming minimal power.
“This is what truly transformative AI looks like: not intelligence that grows ever larger and more remote, but intelligence that becomes woven into the physical and social fabric of everyday life.”
This is what truly transformative AI looks like: not intelligence that grows ever larger and more remote, but intelligence that becomes woven into the physical and social fabric of everyday life. Transformative AI is not about replacing people, but about empowering them, expanding what individuals and communities can do through technology that is efficient, trustworthy, and close at hand.
This is not science fiction. New architectures are already emerging that reimagine AI not as brute-force computation, but as elegant, adaptive, physics-inspired systems grounded in the mathematics of dynamical processes. For example, liquid neural networks,1 pioneered at MIT and now commercialized by the startup Liquid AI,2 are a leading example of Private Physical Edge AI. These models are compact, energy efficient, and performant. Because they are provably causal, they learn the task rather than the context of the task, which allows them to generalize far more effectively to environments and task variations different from those in the training data. Skills learned in one environment can be transferred zero-shot to entirely different ones. For example, liquid networks were trained on drones to solve a hiking task (following wayfinding signs in the woods during summer) and then shown to seamlessly use the same skills in winter, in fall, or even in urban environments, without retraining.3 Today’s transformer-based models cannot do this; they require separate fine-tuning for each new setting. This generalization, combined with their compactness, makes liquid networks a foundational building block for Private Physical Edge AI: systems that are small, fast, explainable, and safe enough for applications in the physical world, such as manufacturing, aviation, or health care, where every inference must be grounded and explainable. The computation inside the neurons of liquid networks is inspired by the math in the neurons of small species such as the nematode C. elegans, which achieves remarkable adaptability and creativity with only 302 neurons.4
Liquid neural networks belong to a class of AI models called state space models, a growing area that is attracting researchers and new ideas. For example, the AI models called LinOSS5 mirror another principle of brain computation: the harmonic oscillations characteristic of large-brained species, including mammals. By representing sequences as evolving states governed by linear dynamics, LinOSS captures long-range dependencies efficiently while keeping compute costs minimal, making it ideal for embedded devices that must process continuous streams of data without cloud support. The Mamba model,6 a descendant of liquid networks, shows that efficiency does not have to come at the expense of capability. It competes directly with transformers on standard benchmarks while consuming far less compute, setting the stage for a world in which models are judged not by trillions of parameters but by intelligence per watt.
This new wave of AI architectures is not just academic. Liquid AI has already open-sourced liquid foundation models (LFM) trained using on the order of hundreds of GPUs that perform on par with models trained on tens of thousands of GPUs, a significant leap in efficiency. These LFM models can run directly on your phone. For example, they can take a photo and describe it instantly with remarkable accuracy and detail, delivering state-of-the-art vision–language performance locally and efficiently, without ever calling the cloud. Imagine a world where we can collect visual diaries and compress them into text descriptions without sacrificing privacy or incurring the huge cost of energy. Now picture a pair of lightweight glasses powered not by cloud servers but by on-device edge AI models such as liquid neural networks. These AI glasses could function as a real-time perceptual companion for blind and visually impaired individuals, by describing the environment, reading text aloud, identifying faces and emotions, helping to locate objects, matching clothing, signaling obstacles, and more, all without relying on an internet connection. Because the intelligence runs locally, users would gain autonomy and privacy: No images of their surroundings would ever need to leave the device, a critical factor when navigating personal spaces like homes, hospitals, or workplaces. This is a glimpse of what becomes possible when advanced AI runs on the edge. If a phone can see and describe the world in real time, it can also become a personal tutor, a medical triage assistant, a real-time translator, or a personal coach—all private, adaptive, and always available. The implication of these capabilities is a future where intelligence is woven seamlessly into everyday life.
“If a phone can see and describe the world in real time, it can also become a personal tutor, a medical triage assistant, a real-time translator, or a personal coach—all private, adaptive, and always available.”
In the not-so-distant future, high-performance intelligence will run ubiquitously on modest hardware, aiming to do for AI what ARM did for processors by creating scalable, efficient cores that manufacturers can embed everywhere, from phones and drones to appliances, medical devices, and more. Together, these advances point to a future where AI is no longer locked away in distant servers but is distributed, compact, causal, and ubiquitous, a form of intelligence that belongs to everyone. Industries will reduce dependence on expensive infrastructure and innovate as startups and small organizations gain access to AI edge capabilities.
Imagine the world when Edge AI has become as fundamental as electricity. Phones are personal diaries and assistants, with embedded AI that adapts to their owners to manage schedules, teach skills, and support well-being. Drones navigate safely and deliver packages using Edge AI on-device planning and control systems. Farmers use solar-powered devices with local AI to improve yields without ever needing to go online. Clinicians in small practices or in rural settings can offer patients the most cutting-edge diagnosis and treatment options using Edge AI support that is fine-tuned with the latest information on clinical trials and treatment alternatives. And all of this is done privately and efficiently without the need for cloud calls.
Even energy systems benefit from Edge AI, as distributed AI helps stabilize renewable grids, forecast demand, and optimize storage. Rather than renting out cloud compute, companies will license efficient AI cores, much as ARM transformed processors. And just as the personal computer shifted value from mainframe operators to software developers, efficient local AI will shift value away from centralized compute providers toward the builders of applications, services, and human-AI collaborations. This will be a world with AI distributed everywhere, empowering individuals, industries, and societies. Entirely new markets will emerge, driven by edge ubiquity.
The spread of personal AI assistants that run privately on phones could also narrow opportunity gaps by bringing world-class tutoring, coaching, and health care to billions who have been excluded by geography or poverty. Because these models run locally, they bring empowerment without surveillance: Users do not have to surrender their data to access intelligence. Trust in AI may rise as it becomes a tool that lives with individuals rather than a service that extracts from them. At the same time, a new AI literacy divide will emerge. The question will not be who has access to AI, but who knows how to use it effectively.
Societies that invest in widespread AI education will flourish; those that do not may find themselves divided between empowered AI users and those left behind.
The defining features of Edge AI are its ubiquity, energy efficiency, and privacy. Intelligence will be measured in tokens per watt, per chip, per device. It will be woven into daily life, driving productivity, creativity, and inclusion at a planetary scale. Edge AI will redistribute the benefits of intelligence more widely than any prior innovation in history.
“Just as the PC shifted value from mainframe operators to software developers, efficient local AI will shift value away from centralized compute providers toward the builders of applications, services, and human-AI collaborations. Entirely new markets will emerge driven by edge ubiquity.”
AI is at an inflection point. For years, progress has been measured by ever-larger models and ever-greater demands on data centers, pushing intelligence toward industrial-scale infrastructure. But the future of AI will not rest on size alone. It will also be defined by capability, efficiency, adaptability, and ubiquity—by intelligence that runs everywhere, instantly, and privately. Just as the PC shifted value from mainframe operators to software developers, efficient local AI will shift value away from centralized compute providers toward the builders of applications, services, and human-AI collaborations. Entirely new markets will emerge driven by edge ubiquity. Private Physical Edge AI is the path to that future: moving computation from distant servers to the devices in our hands, our homes, and our workplaces, where it can empower people, companies, and organizations directly while reducing cost, latency, and energy use.
This shift carries profound economic implications. By decoupling intelligence from centralized data centers, Edge AI redistributes value creation from a handful of infrastructure providers to the countless businesses, organizations, and communities that can embed intelligence directly into their own systems. The economics of AI move from renting cloud compute at scale to deploying efficient, purpose-built models that run locally at near-zero marginal cost. This transition lowers barriers for startups, small enterprises, and for entire industries, from health care and education to transportation, finance, energy, agriculture, and public services. The rise of Private Physical Edge AI will democratize access to intelligence, spreading productivity gains across the global economy while reducing the concentration of cost and control in the cloud.
1. Ramin Hasani, Mathias Lechner, Alexander Amini, Daniela Rus, and Rady Grosu, “Liquid Time-Constant Networks,” in Proceedings of the AAAI Conference on Artificial Intelligence 35, no. 9 (2021): 7657–7666, https://doi.org/10.1609/aaai.v35i9.16936; Ramin Hasani, Mathias Lechner, Alexander Amini, Lucal Liebenwein, Aaron Ray, Max Tschaikowski, Gererd Tescl, and Daniela Rus, “Closed-Form Continuous-Time Neural Networks,” Nature Machine Intelligence 4, (2022): 992–1003, https://doi.org/10.1038/s42256-022-00556-7; Mathias Lechner, Ramin Hasani, Alexander Amini, Thomas Henzinger, Daniela Rus, and Radu Grosu, “Neural Circuit Policies Enabling Auditable Autonomy,” Nature Machine Intelligence 2 (2020): 642–652, https://doi.org/10.1038/s42256-020-00237-3.
2. www.liquid.ai.
3. Hasani, Lechner, Amini, Liebenwein, et al., “Closed-Form Continuous-Time Neural Networks.”
4. Oliver Hobert, “The Neuronal Genome of Caenorhabditis elegans,” in WormBook: The Online Review of C. elegans Biology (WormBook, 2005–2018), https://www.ncbi.nlm.nih.gov/books/NBK154158/.
5. Konstantin Rusch and Daniela Rus, “Oscillatory State-Space Models,” paper presented at the Thirteenth Conference on Learning Representations (ICLR 2025), Singapore, April 2025, https://openreview.net/pdf?id=GRMfXcAAFh.
6. Albert Gu and Tri Dao, “Mamba: Linear-Time Sequence Modeling with Selective State Spaces, First Conference on Language Modeling,” preprint, arXiv, May 2024, https://arxiv.org/abs/2312.00752.
