# Why the Physical World Is Hard to Read It is hard to know exactly what happens in the physical world. That limits how fast AI learns hardware. — Christian Keil's "read access" framework > [!example] The water-injected chickens > A buyer of 45M chickens per week saw up to 20% weekly variance in saleable product, with the same spec, suppliers and machines. Deep measurement found the cause: one supplier injected water. Intake weights passed. Water is not meat. ## Five reasons 1. **Non-uniform and changing.** Parts have defects. Properties shift with temperature. Machines and gauges drift. You see drift only if you invest in measuring inputs and outputs. 2. **Sparse.** You cannot sense everything. On satellites, the thermal team wanted thermistors everywhere; other teams pushed back. After a failure, you wish you had the one data point you skipped. 3. **Expensive.** Precise measurement needs even more precise machines. CMMs cost a lot and are slow to buy and run. 4. **Latent.** Thermal cycling, propellant mixing and fatigue tests take time. A HALT ("shake and bake") campaign reruns in full after every redesign. 5. **Qualitative.** The least compressible. Do users like it? Is it comfortable to hold? This barely fits a digital learning loop. ## Consequence - Machine-readable, quantified tests → fast model progress. - Sparse, slow, qualitative signals → slow model progress. - AI capability in hardware stays spiky until this changes. How long the barriers hold against the Bitter Lesson is open. ---- tags: #boat #deeptech links: [[Superhuman Hardware Design Is the Easy Part]] [[Industrial AI MOC]] [[Sensor Data and Historian Systems]] [[Digital Twins]]