Over the past few days, I kept seeing posts about Jev, so yesterday I spent some time reading up on it and discussing it with ChatGPT. Setting Jev itself aside, I was more interested in the idea behind it: if AI does not need to go through a full reasoning process for every task, but can quickly judge familiar situations, that feels a lot like System 1 and System 2 in Thinking, Fast and Slow. System 2 is slower and handles deeper thinking; System 1 is faster and makes quick judgments.
That got me thinking about Physical AI, which I have been experimenting with lately. If AI is ultimately going to do more than produce a response—if it will actually control robots, equipment, and production lines, or act directly on the physical world—is there a layer missing beneath System 1 and System 2? A layer that does not need open-ended reasoning or a fresh understanding of the whole situation, but instead turns an already confirmed decision into constrained, verifiable, predictable action. For now, I call it System 0.
I previously built a small project that uses MCP to help with board-level development and debugging. Anthropic also recently proposed MHS to address the interface between AI and physical devices. Looking at these together, the place of System 0 suddenly seemed interesting. But after searching online, I found that “System 0” has no single agreed definition. So far, I have come across at least a few different interpretations.
One defines System 0 as an external cognitive layer that precedes human System 1 and System 2: AI organizes, filters, and frames information before passing it to a person to think about. Another is closer to embodied intelligence and treats System 0 as “pre-cognitive physical processes.” A soft gripper, for example, does not need to calculate every contact point. The material’s deformation alone lets it conform naturally to an object. In a sense, the physical world performs part of the computation itself. A third usage is closer to robot control, treating System 0 as highly practiced, automated movements such as balancing, trajectory following, and stabilization. None of these definitions exactly matches what I have in mind, but the third comes very close.
The System 0 I imagine is about more than how a motor moves. It asks: does this action still meet the existing rules, so that it can continue according to the established procedure?
For example, when a robot sees a workpiece, System 2 might think: “Why do we keep failing to pick it up today? Is the problem with the camera, lighting, gripper, or fixture?” System 1 makes a quick judgment: “The workpiece is in a normal position and can be picked up, with 96% confidence.” System 0 does not reinterpret the situation. It checks whether the safety gate is closed, the robot is in the correct mode, the workspace is clear, and the target position is valid, then runs a verified picking procedure. If the conditions are met, it executes. If they are not, System 0 does not need to guess why; it only needs to stop or escalate the problem.
So for now, I remember it simply:
- System 2 thinks.
- System 1 judges.
- System 0 acts.
What makes this interesting is that they should not form a fixed pipeline. Not every action needs to pass through System 2 → System 1 → System 0. Instead, look at it from the bottom up: if System 0 can carry out a routine task under existing rules, there is no need to invoke higher-level judgment.
When System 0 finds that execution conditions are not met—for example, a sensor reading exceeds the allowed range, required data is missing, or the current state falls outside the established workflow—it escalates the problem to System 1. System 1 then judges which known situation this belongs to. Should it capture another image, wait, retry, or hand the task to a person? Only when System 1 cannot classify the situation reliably, or encounters a problem it has never seen before, does System 2 step in for full reasoning.
In other words, if rules are enough to act, do not judge again; if a quick judgment is enough, do not think deeply. This matters especially for Physical AI. In a conversation, an incorrect AI response can usually be tried again. But when AI controls a robot arm, vehicle, machine, or production line, a wrong guess can directly become a physical action. The closer it gets to the physical world, the less freedom it should have. System 2 can retain considerable room to think, while System 1 should judge within a limited set of options. By System 0, the behavior should ideally have narrowed down to explicit procedures, rules, state machines, or skills, executed only when the conditions are met. Further down, PLCs, motion controllers, and safety systems should be even more deterministic.
Seen this way, another idea interests me: knowledge can settle into lower layers. The first time a new problem appears, System 2 may spend considerable time understanding it. After the same problem has appeared many times, we may find a stable pattern and turn it into a quick System 1 judgment. With extensive validation, that judgment might even become a fixed System 0 rule.
What once required “thinking” eventually becomes “judging”; what once required “judging” eventually becomes “following rules to act.” For now, I call this “cognitive compilation.”
A mature AI system perhaps should not depend forever on ever-larger models to handle every task. Instead, it should keep moving understood and validated knowledge into faster, more stable, lower-risk layers. Of course, System 0 / 1 / 2 is not a formal industrial standard.
System 1 and System 2 belong to the familiar Thinking, Fast and Slow framework, and System 0 has different definitions in academic work. I am not trying to redefine System 0. I simply feel that, from an engineering and communication perspective, “System 2 thinks, System 1 judges, System 0 acts” is a low-friction, easy-to-remember way to describe it. My core thinking about Physical AI right now can also be condensed into one sentence:
The closer intelligence gets to the physical world, the less freedom it should have, and the more reliable it needs to be.
The real challenge of Physical AI may be more than making AI better at thinking. It is figuring out how to progressively constrain and validate that thinking until it becomes action that can safely affect the real world.
Comments & Feedback