Creating small, specialized models for specific tasks. Being able to leverage the up front training/data as a generalized base allows you to quickly create a small local model that can generate outputs for that task that can come close to or match the same you would see in a large/hosted model.
Locked down meaning the storage devices are encrypted and decrypted on-the-fly via the SoC/CPU using a key programmed into the Fuses/OTP (this is usually per device keys), bootrom/loader requiring signed firmware images, limited exposure of external interfaces (attack surface) - from my memory even the Uart interface attached to the SoC was disabled very early on in the boot loader, exposing only one or two messages. I would not expect that ram is encrypted - I cannot think of a single time I have seen that implemented in a device. Maybe it’s time to dig that board out of storage and poke at it a bit more invasively, my understanding is they are not very robust when faced with fault injection :>
Following up on this - the actual 'self driving' part of the HW stack is an entirely separate board with 2x custom ARM chips on it. The HW/SW is much more locked down and the OS/Data is not accessible. I believe a lot of the self-driving info gleaned by types like green were built up from the first generation of Model S cars where the 'self driving' HW was much less defensible and it was much easier to gain access to it.
HW wise, the older units were intel atom based cpu (latest gen is amd I believe?) - the hardware is typical embedded stuff - cpu + eMMC + bt/wifi mcu + cellular daughter card. OS is linux + QT UI stuff. I would expect things have changed for newer HW revisions, but the previous gen did not utilize encryption (dmcrypt) so all data was unprotected at rest.
Years back I bought a model3 infotainment unit on eBay to hack on - it’s absolutely insane at the amount of data contained on them. After gaining access to the system I was able to get the VIN of the car and find the salvage auction from the car it came out of - it had been wrecked. I then was able to get all the location data that gets logged, showing a glimpse of the previous owners life (house, work, stores they went to, etc) as well as the final resting place of the car. The last gps locations logged were at the end of a “T” intersection in North Carolina - google street view gave a nice look at the trees the car most likely hit :>
https://www.atredis.com/blog/2024/6/3/how-to-train-your-larg...