Comment Re:Tech inversion (Score 1) 163
These two statements are wrong.
The AI object recognition algorithm accuracy is known (in the AI field) to scale linearly with the amount of input data.
This cannot be true. Accuracy is the fraction of correct predictions. It must lie between 0 and 1. Thus it cannot scale linearly with input data which is unbounded.
Furthermore, it is empirically not true. Look at scientific papers for measurements of input data quantity versus accuracy for a variety of tasks. Those measurements do not show a linear relationship.
Tesla is currently switching from a 2-d representation model to a 3-d model, a change which requires a complete rewrite of their autodrive system from the ground up. They expect to be done sometime in August, and they expect this to improve the object recognition accuracy by a factor of 3 to 5 *times* over the current algorithm
This is a misleading statement that can only confuse non-experts. There is no clear way to interprete multiplicative factors of improvement in accuracy. A factor of 2 improvement when accuracy rate is 0.01 is very different from the same improvement when the rate is 0.5. Scientific papers do not measure progress like this.