faultbugs

Shoot a night street on daytime settings and you get a nearly black photo. The scene still has differences in brightness; the exposure is just tuned for daylight, so the camera can't capture them. @Tesla's car AI chips face a similar problem. To compute fast, they store numbers as INT8, or 8-bit integers. Eight bits give only 2⁸ = 256 values, so every number has to land on an integer from −128 to 127. When large and small numbers share one ruler, the small ones crowd into a few marks near zero, and their differences vanish. A Tesla patent application published on October 1 gives every token, each chunk of data the AI reads, its own ruler. It finds the largest value in each token and maps it to 127, the top of INT8. A small value like 0.07 rounds to zero on a shared ruler but survives when each token gets its own. Yet the drawings show Tesla doesn't do this at every step. The full post below walks through where it does, and why some steps keep a fixed ruler.