10x

Elon said the next version of FSD— presumed to be version 14— will have 10x the number of parameters as version 13 (including the version 13 builds used in Robotaxi).

What does this mean?

Architecturally, a neural network is made up of layers, and those layers contain nodes (also called neurons). Each node in one layer is (typically) connected to every node in the next layer. These connections have numerical values called weights, which determine how influential each connection is. Each node also has a bias, which adjusts the node’s output independently of its inputs.

The parameters in a neural network are the weights plus the biases.

So when Elon says version 14 has 10x more parameters, he means the total number of weights and biases has increased— implying a larger, more expressive model.

In a sense it means the model is more nuanced, can contain more "decision" points.

In general a bigger model is better, but also requires more compute. More compute to train and more compute to run (inference).

Deciding that version 14 will be 10x bigger than version 13 implies the team is confident they can run it on HW4 Tesla's. In other words, they know the performance can be achieved for inference.

Part 2 (https://t.co/yjqIkLRHvF) of this series does a good job explaining this.