The Cognitive Revolution · Sep 27, 2026
With Nick Gillian, Co-Founder and CTO of Archetype AI
Archetype AI has gathered close to a billion hours of physical sensor data to train Newton, a foundation model that reads radar, vibration, electrical current and camera feeds and reports what a machine or a worksite is actually doing.
Image models learned from billions of pictures posted online with descriptions attached. Nick Gillian said sensor data has no such pairs, so Archetype had to find its own way to connect raw signals to language.
"The reason the sensor is giving you all these nines is not because the sensor is broken. It's actually a feature of the machine, right, in those cases."
Nick Gillian, co-founder and CTO of Archetype AI, on The Cognitive Revolution, previously led machine learning research on Google's Soli radar sensor. He walks through Newton from the data up, through a five-year river project with the Japanese builder Kajima.
The full segment is covered here so you can skip it.
Here are the 5 lessons that matter.
Key Takeaways
Archetype has gathered close to a billion hours of physical sensor data, most of it sparse and unsynchronized
There are no radar-and-text pairs online of the kind that trained image models
Newton turned camera and river sensor feeds into a work chart for a Kajima river project
It showed productivity stayed low for days after storms, a pattern the builder had never seen
If an average person could spot the problem, Newton can do it out of the box
A few thousand customer samples are usually enough to adapt the model to a new site
1. A Billion Hours of Sensors
The host said Newton takes raw streams from radar, vibration, electrical current and cameras. He asked how much data Archetype could collect on its own and how much came through partners.
Gillian said Archetype is "getting close to a billion hours now of physical AI data." The physical data on the web is thin, he said.
"There's very little multimodal data. It's not synced well. It doesn't cover things well." — Nick Gillian
Self-supervised, like language models
Classic supervised techniques build one model for a fixed set of sensors, all captured at the same sample rate, which makes messy data a massive problem, he said. Archetype turned that around: it uses self-supervised techniques to learn from sensor data captured by different systems in different contexts, the way large language models learned from text.
It then mixes in a small amount of data from partners, and sometimes from customers, to map what it learns to "physical language" and to physical control.
When a missing value is the signal
Much of the work is resampling and understanding missing values, he said. A gap may mean the sensor failed, or it may be a sign that the machine it is attached to is about to break.
"The reason the sensor is giving you all these nines is not because the sensor is broken. It's actually a feature of the machine, right, in those cases." — Nick Gillian
