Archetype AI Team
We are excited to welcome Dong Lin to Archetype AI as Principal Research Engineer, leading the team building Newton — our foundation model for physical world data.
Dong brings over two decades of experience developing, shipping, and optimizing world-class production models. Most recently she served as Principal Machine Learning Engineer at Apple, and before that spent nearly 14 years at Google, including Google Brain and DeepMind, as a Senior Staff Software Engineer. She holds a Ph.D. in Computer Engineering from the University of Illinois Urbana-Champaign, where her research focused on speech modeling, indicating early interest in what she would spend her career pursuing: teaching machines to understand the world through signals, not just text.
Dong’s work at Archetype centers on a fundamental challenge. Vision and language give you part of the picture. The physical world runs on signals — vibration, pressure, temperature, humidity, airflow — that cameras can't see and LLMs aren't built to understand. Newton is designed to close that gap: a foundation model trained on physical reality, capable of generalizing across assets, environments, and industries without rebuilding from scratch.
We asked Dong a few questions about what it takes to build a foundation model for the physical world.
Newton is trained on physical sensor data, not text or images. What makes that a fundamentally different modeling challenge?
Dong: What makes physical sensor data fundamentally different is that it can only be an indirect and partial view of the world. A signal is not self-contained; its meaning depends on where it came from, how it was measured, and what physical process produced it. So a key challenge is not just modeling the signal, but modeling the physical world behind the signal.
Newton has to generalize across physical systems that are radically different from one another; the model has to work for a wind turbine, a pump, a building HVAC system. How do you think about that problem architecturally?
Dong: If you imagine what AGI ultimately requires, it is a model that can understand the physical world, and sensors are a critical part of that because text and vision only give a partial view. Architecturally, the challenge is to build a shared model of physical reality that can separate context, meaning which system and environment you are in, from dynamics, meaning the behavior unfolding within that context. That is what makes it possible for the same foundation to generalize across a wind turbine, a pump, and an HVAC system.
What does "zero-shot" actually mean in a Physical AI context? Why does it matter for enterprise deployments?
Dong: Today, most Physical AI is still bespoke: one model per sensor, one model per use case, and that makes it hard to scale or adapt as systems change. Zero-shot is the standard we should aim for: a foundation model useful on a new sensor, asset, or use case without a custom retraining cycle. What makes that genuinely hard is that the meaning of a physical signal depends on context that is often external to the signal itself — the asset, the operating regime, the installation — in a way that has no clean analog in language or vision. Solving that is what makes real-world deployment scalable.
You spent years building production models at Google Brain/DeepMind and Apple. What did working at that scale teach you about what makes a model trustworthy in real environments?
Dong: What stayed with me most from Google was that scientific rigor was at the core of the culture. You build strong experimental frameworks, targeted metrics, and well-designed evaluation methods so progress is real and repeatable. That mindset still shapes how I think today: not just about model quality, but about being disciplined in what you measure, what you optimize, and whether those choices actually align with the outcome the business cares about.
Your background started in speech modeling. How does that early work connect to what you're building now?
Dong: What fascinated me in speech was the idea that you could compress enormous variability across speakers, accents, and environments into a compact universal representation. It showed me how much surface variation a model can absorb while still capturing the underlying structure that actually matters. Physical AI is ultimately pursuing a similar goal: learning a unified model that can capture the underlying structure of the physical world across many different sensors, systems, and contexts.
What's the hardest unsolved problem in Physical AI modeling right now?
Dong: There are several fundamental unsolved problems: building models that are truly grounded in physical reality, models that understand causality rather than just correlation, and models that can keep learning safely after deployment. That last one is especially hard because a deployed model needs to adapt to its local environment while still benefiting from global knowledge expansion. We still do not have a clean solution for doing that under real-world constraints.
.jpg)
