It is both funny and sad to see 15 and 30 minute videos claiming to show "full self driving.". Currently Tesla is level 2, Mercedes has a level 3, and Waymo is level 4 autonomy. Any fully self driving car needs to operate routinely day after day for YEARS without causing an accident.* Otherwise who would even insure it? Even a successful unmanned drive from coast to coast would not prove much compared to what's needed.
* as many humans do
I want to tryout Arduino's new SBC. It should be able to run a light version of A.s.a. H. They were not immediately listed by Sparkfun or Adafruit when I looked. Perhaps I should wait for the 4 GB/32 GB version.
As A.s.a. H's concept grounding* is built up (from something like Wierzbicka's** "alphabet of thought," through Lang's Toki Pona, and then minimal English vocabulary) I find that my use of camera(s) and neural network classifier(s)*** has grown. The extent of this depending upon the environment(s) and task(s)/specialties, of course.
* See, for example, my blog of 4 September 2024.
** And Leibniz.
*** Marked "ANN" in the diagram of my 4 September 2024 blog.
My XAI involves producing concept maps like that of my 4 September 2024 blog* and printing out those concepts that A.s.a. H. has active as a function of time as it runs. Of course Asa can form concepts that are not shared by humans.
* and my Trans. Kansas Acad. Sci., vol. 120, 2017, pg. 108 and blog of 14 November 2017
When you deploy an Asa H AI agent its performance will usually improve over time.*
The performance of a deployed LLM, on the other hand, tends to degrade over time.
* Due to learning new experiences, extrapolation, etc.
I'm thinking of something like Shuttle-C (disposable external tank but recoverable/reusable main engines, pumps, avionics, etc. and heat shield)** but with something like the reusable Falcon 9 first stage replacing each of the solid rocket boosters.
* see my blogs of 1 January 2025 and 1 June 2025
** something like this:
I am not a specialist in transformer neural networks or in natural language processing but it is my impression that LLM progress has plateaued.
Is this serious enough to block Tesla's approach* to controlling robotaxis, robots, and fsd?
* end-to-end neural network
