* Internal feature space accuracy increased tenfold during training, this makes the model dramatically more accurate.
* Internal feature space information content increased tenfold during training (to ~700 bits), this makes the model dramatically more accurate
* Less reliance on previous frames makes model more reactive and snappy
* Trained in new reprojective simulator
* Model trained in openpilot was trained in 36hrs from scratch, compared to around 1 week of previous releases
* Model training now simulates lateral and longitudinal behavior, this allows openpilot to slow down for turns, stop at traffic lights, etc,... in experimental mode