Three Problems Standing Between a Good AI Model and a Trustworthy One

Many are talking about Physical AI. Few are talking about why it is so much harder to get right than it looks.

Our co-founder and CTO Prof. Nadav Cohen laid out three problems that stand between a good model and a trustworthy one.

You cannot see everything. A camera shows the outside of a pump, not its internal temperature or a drifting sensor. A real physical model has to fuse video, pressure, vibration and operating history into one picture.

Statistics alone will lie to you. Without physical constraints in training, a model can predict a state that is plausible in the data and impossible in reality, like a liquid level dropping with no drain path. The fix is first principles in the model, not just in the data.

Correlation is not causation. If a warning light comes on every time an engine overheats, a naive model may decide the light caused the heat. Getting this right means structuring data as state, action, outcome and properly inferring causality.

This is the foundation we built Imubit on. Closed-Loop Physical AI means process engineers model causality, not correlation, and take the decision all the way to supervised autonomous execution with DLPC®.

Nadav's column is in Hebrew; your browser will translate it: Geektime.

‍