Archetype AI has introduced an AI model called Newton that can grasp physical phenomena without needing humans to provide any training or rules. The model processes raw sensor data to make accurate predictions across various fields, from mechanics to energy systems, by analyzing patterns instead of being given specific equations.
Learning Without Help from Humans
Newton showcases an unprecedented method of learning physics by analyzing raw sensor inputs. The model can generalize physical principles, like energy transfer or oscillations, from purely observational data, which is something no AI model has done before. Using over half a billion samples from sensors, Newton learns to make forecasts about physical systems, including predicting power grid behaviors and temperature changes in transformers.
One of Newton’s most intriguing abilities is that it wasn’t trained for specific tasks. For example, it managed to predict the chaotic motion of a pendulum accurately, even though it had never been taught about pendulum physics. Unlike most AI models that need large amounts of specialized data, Newton performs predictions in unfamiliar environments.
Practical Use Cases for the Industry
Newton’s general approach could change how AI is applied in different sectors, from manufacturing to infrastructure. Most AI tools are tailored for particular use cases and rely on large datasets. Newton can function in several contexts with less need for customized models, allowing industries to deploy AI systems more quickly and with fewer costs.
Newton’s potential doesn’t end there. In industries where unique, unpredictable events are common—such as industrial machinery breakdowns or citywide power demand—this AI model can adapt and make accurate forecasts based on its broader understanding of physical patterns.
Beyond Machines: Human-Enhanced Perception
One of the model’s future possibilities is expanding what humans can perceive. Ivan Poupyrev, one of Archetype AI’s founders, said Newton could make sense of data from sensors that capture aspects of the world humans can’t detect directly, adding a new layer to how we experience reality. For instance, environmental sensors or medical diagnostics could reveal patterns in data that human researchers might overlook. This shift in how AI interacts with the physical world could unlock new ways to understand it.
Newton’s capacity to work with unfamiliar data might be particularly useful for scientific research, helping to identify previously unnoticed trends or anomalies in various fields of study.
Training and Future Predictions
Built on transformer-based neural networks, Newton was pre-trained on millions of diverse physical phenomena, from fluid dynamics to electrical currents. The result is a model that can process various types of data and reveal patterns across different applications. Its adaptability is made possible through additional lightweight neural network decoders, which are applied for specific tasks like forecasting future events or interpreting sensor data from the past.
The system does not need to be retrained for every new task. Instead, it adapts based on its foundational training, making it much more efficient than traditional AI models. In practical terms, this could allow Newton to be used in many industries without needing extensive recalibration for every new scenario.
Generalized AI for the Physical World
The ability of Newton to generalize goes beyond what most AI models are currently capable of. Known as “zero-shot forecasting,” Newton can predict the behavior of systems it hasn’t encountered before. For example, even though it hadn’t been specifically trained on predicting the temperature fluctuations of electrical transformers, Newton made accurate predictions when tested on that task. The same was true for forecasting city energy demands.
Comparison of forecasting mean square error for the zero-shot, fine-tuned, and target-trained models. Lower is better.
This “zero-shot” capability means that Newton could be trained just once and then applied to multiple problems, rather than requiring retraining for each new application. This opens the door to using AI in environments where data is sparse or difficult to collect, something that would usually be a major barrier for traditional AI systems.
Newton’s ability to understand the physical world from raw data is a major step forward for Archetype AI, which has already raised $13 million in venture capital to bring these capabilities to real-world use cases.