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
Some specialist AI agents that I collaborate with:
1. natural language knowledge base*
2. machine/process control agent**
3. recommenders***
4. planner/scheduling agent
5. calculators****
6. summarizer
7. A.s.a. H.
* which I use every day.
** see, as an early example, Transactions of the Kansas Academy of Science, vol. 101, pg 132, 1998.
*** which know my wishes/likes/activities.
**** machines have been better than humans at arithmetic for a long time now.
A.s.a. H. does not "stand-by" or "idle", it's always thinking*, even when its sensory input is minimal. I'd be interested then to see A.s.a. ask any questions "unprovoked."
* Unlike LLMs and many other AIs.
I got my PhD in physics 50 years ago. At that time we thought commercial fusion reactors were "right around the corner."
