The Tesla autonomy show had a lot of nuggets about what the typical self driving car startup does. There's a chip which is a scaled up version of a GPU, using stock IP that doesn't use very current feature sizes, but marketed like crazy as a revolution. Something never mentioned before, but which slowly oozed out over the last 5 years is how training the networks still entails manually annotating thousands of images. They have some interpolation, some LIDAR to generate annotations, but most of it is mass Bangalore labor.

A more telling feature was the use of neural networks in place of all conditional statements. Determining if a car was changing lanes was an exercise in thousands of sets of car coordinates manually classified as lane changes. They couldn't just classify it based on x > threshold. Enormous effort was spent doing it with a neural network. How much more effort would have been required to make a pose classifier based on neural networks instead of if statements?


There's still no magic in using a neural network to detect a path. It has to be trained with thousands of images of every section of path in every lighting condition. They all have to be manually annotated or interpolated with path coordinates. All the network does is compress the training set & replay it. The network has to be more like millions of neurons instead of hundreds.

It was $400 all in, without the electronicals. After 20 miles, the transmission was extremely noisy, despite lubrication & the sound damper. Despite the noise, the efficiency was a record high of 240mAh/mile empty. 2200kV is a better speed than 2400kV. The hall effect sensor worked without any firmware changes. It actually manetaned the same speed going downhill. There's no brake modulation, so it would have to always deliver the minimum power....Continue Reading

These are quite good, by lion standards. After running 20 miles with the crummy speaker, the lion kingdom was willing to pay anything for a better speaker. They have all sounded good over $20 & in the narrow form factor. Below $20, they sounded terrible. Never tried the pancake style ones. The lion kingdom is continually amazed by how good any bluetooth speaker over $20 sounds, compared to anything that small from 30 years ago.

The old speaker was a TV speaker modified to be very light, for a smaller vehicle. The smaller vehicles are no more, so the old speaker may end up converted into a headset.


Millenials would freak out over how bad it is, compared to their $200 Bose speakers. A generation exists which has never heard sound as bad as a 1970's transistor radio. The lion kingdom manely grew up with the sounds of 1970's transistor radios & clock radios. You wouldn't think something mechanical like a speaker could be improved.