When I was a professor I always found it harder to debug code written by my students as compared with debugging code I had created myself. This is now even more of an issue when I try to debug code created by an LLM**.
* See my blog of 1 January 2025.
** No, coding is not "solved." That's the capitalists lying again. There are still lots of bugs to find and fix.
I am glad to see the effort and resources currently being spent on A.I., robotics, and space travel. I don't expect that these efforts will produce anywhere near what's being promised of course. Each of these efforts is currently too narrowly focused, they should all be more broadly based.
The "AI slowdown" is not caused by a coordinated reaction to safety concerns*. That's being used as an excuse. Rather, it is due to the terribly weak scaling law for LLMs. This was foreseen years ago; Ilya Sutskever's "end of the age of scaling" and return to another "age of research."**
* really they're doom trolling
** See also my blog of 1 January 2024.
It seems to me that there is very little we can really say on religious matters. We need to stick to what is evidence based.* Our current two "best" models** of the world, quantum mechanics and general relativity, both suggest that we are all immortal in certain senses.*** One can also identify/define "supreme beings"**** perhaps like the Buddhist "deities."***** This leaves behind an immense knowledge vacuum that is an open invitation to all kinds of crazy speculation.******
* Something like Spinoza's God is possible, see my blog of 3 January 2024.
** These are not "final theories." They may well be wrong. As our models change so must our religious notions.
*** See my blog of 1 November 2023.
**** See my blogs of 27 and 29 August 2018.
***** Though if we are alone in the cosmos the supreme beings could just be us apes.
****** I.e., present day "world religions." Not only are they not evidence based they are often times ruled out by a good deal of evidence.
To insure that a robot will be taking action based solely on ITS sensory input and memory alone I will frequently teleoperate "blind" from an adjacent room. This requires that I display the robot's sensory and state information as I go. I try to also avoid any influences coming from my own memory of past actions and consequences.
The learning curriculum* is a much under appreciated part of AI research and development. My current sequence of training for an AGI is something like:
1. First expose the agent to smells, wind, tastes, touch, wetness, temperature variations, light, sound, dust, IR, colors, and train the camera/ANN to learn shapes, numbers, letters, faces, people, plants, and animals.
2. Next teach grasping, motion, line detection, fire detection, acceleration, near and far, hunger, and hardness.
3. Then, approach/retreat, search, pain, carry, collision, hill, path, raise and lower, and object.
4. Then tool use and the self concept.
Training will differ for various specialist agents of course.
* See my blogs of 18 July 2014, 18 May 2015, and 29 January 2015
There is ample evidence that groups make better decisions than individuals do. See: F. Galton, Nature, 75, pp 450, 1907. J. Surowiecki, The Wisdom of Crowds, Random House, NY, 2004. R. Larrick and J. Soll, Management Science 52, pp 111, 2006. L. Hong and S. Page, Proc. of the Nat. Acad. of Sci., 101, pp 16385, 2004 and Economic Theory 79, pp 275, 2025. C. Davis-Stober et al, Decision, 1, pp 79, 2014. D. Prelec et al, Nature 541, pp 532, 2017. I. Douven et al, Artificial Intelligence, 275, pp 235, 2019 and Collective Intelligence 5, 2026
