Physical AI / Behavioural simulation
Nearly every model you have heard of predicts the next word. We start somewhere else — from how brains take in the world, and how minds decide what to do about it. Our architectures are grounded in neurology and psychology rather than language.
The thesis
Human intelligence is not one capability. There is the machinery that perceives the world and moves through it, and the machinery that feels, weighs and chooses. Zivia Labs is organised around both.
Neurology
How does a system look at a scene it has never encountered, work out what is in it, and move through it without hitting anything? That is a problem about pixels, geometry and time — not sentences. It is the foundation of Kael OS, and the reason we can put a machine into a vineyard.
Psychology
Why do two people given the same product react in opposite ways? Because personality, emotion, memory and trust are doing the work underneath the opinion. Model those explicitly and you can simulate a customer instead of guessing at one.
The language-model route
Everything the model knows about a vineyard has to survive being written down first. Words are a lossy summary of a physical scene, and the bill for that detour arrives as training cost.
The Zivia route
No text bottleneck. The model represents space and motion directly, which is what an arm reaching into a canopy actually needs — and it is a far smaller thing to train.
01 / The foundation
The model underneath everything else we build. Kael OS recognises images and video, understands three-dimensional space and navigates through it. It is not a large language model, and it is not a language model wearing a vision adapter. It is a different kind of thing from the ground up — and the ground it starts from is the bit.
A fraction of the cost
Teaching Kael OS to interpret a new domain of video costs a fraction of what an LLM-based approach costs to do the same job. Purpose-built architectures are smaller, and smaller models are cheaper to train, cheaper to run and quicker to adapt.
No integers. No floats. Bits.
Kael OS is defined at the bit level. There is no integer or floating-point representation anywhere in the model, which is unusual — essentially every modern AI system is a very large pile of floating-point multiplications.
That choice has a consequence worth sitting with. Operations expressed in bits are operations a qubit can carry out, so Kael OS is designed to translate onto quantum hardware natively rather than being rewritten for it. Most of the field will have to port its arithmetic when that hardware matures. We intend not to.
Recognition across still images and continuous video, including cluttered scenes where the thing you care about is partly hidden.
Builds a working representation of the environment — what is where, what is in the way, what is reachable.
Moves through real environments that were never laid out for a robot, on surfaces that are not flat and in light that will not hold still.
Retraining onto a new domain of video is a routine, low-cost operation rather than a funding round.
No integers, no floats. The arithmetic is bit-level, which is what makes the step to qubit-level computation a translation rather than a rewrite.
Anywhere video is the raw material
02 / Agtech
We chose grapes deliberately. Harvesting them is one of the least forgiving perception problems in agriculture: fruit hidden behind leaves, ripeness that has to be judged by eye, bunches that bruise if you hold them wrong, and rows of soft uneven ground to drive down. Solve picking and you have demonstrated the whole stack — see, understand, move, act.
Demo
This clip needs a browser that can play MP4 (H.264) video. Download it instead.
Clusters sit behind canopy. Detection has to work from partial views rather than clean, well-lit product shots.
The decision to pick or leave a bunch is made from vision, the same signal a human picker is using.
A harvest that arrives damaged is not a harvest. Approach and grip are part of the problem, not an afterthought.
Navigation over rutted ground, between trellises, in light that changes through the day.
03 / Personaut
Personaut is a Python package and a VS Code extension that simulates customer feedback, then helps you turn that feedback into a product specification. It is in the hands of developers today — over 3,500 downloads — and it is rooted in psychology rather than vibes.
import personaut
sarah = personaut.create_individual(
name="Sarah",
traits={"warmth": 0.8, "dominance": 0.4},
emotional_state={"sceptical": 0.6},
)
sarah.add_mask(personaut.masks.PROFESSIONAL_MASK)
sim = personaut.create_simulation(
situation=onboarding_flow,
individuals=[sarah, mike, priya],
type=personaut.simulations.types.CONVERSATION,
)
sim.run(num=5, dir="./findings/")pip install personaut
Asking a chatbot to “act like a 34-year-old sceptical buyer” gives you an impression of a customer. Personaut gives you a specified one: state that persists, emotions that move for reasons, and recall that depends on trust.
Traits, emotional baseline, history and relationships — as code you keep in version control.
Your onboarding, your pricing page, your pitch. Run it as a conversation, a survey or a simulation.
Where trust dropped, where confusion spiked, who walked away and at which step.
Then re-run the same panel against the fix, as often as you like.
Company
Zivia Labs is a young company. We would rather you judge us on what is already inspectable than on a roadmap: a package you can install in the next thirty seconds, and a machine you can watch do a job that is genuinely hard.
The reason the work holds together is the belief underneath it. Intelligence is not one trick, and it is not fundamentally made of text. Build from perception and you get machines that can work in the physical world. Build from psychology and you get software that can anticipate people. We are doing both, on purpose.
Tell us what you need a machine to see, or which customer you cannot get in a room. Pick whichever is closest and it will reach us directly.