Mind children

My allegiance is to the survival, growth, and spread of intelligence* be it embodied in biological life or mechanical life. All species go extinct. Is it important that human life continues once artificial general intelligence develops? **

Can mechanical life reproduce and sustain themselves without human assistance?** If it can is that most efficient?

Are humans useful contributors to a plural society of biological and mechanical life? A society of specialist agents?

Is added diversity due to the inclusion of humans useful?

Would humans' alternative mental conceptualizations be useful?

Can the society change direction if desired/needed? Even reverse course?

* These should be 3 components of the agent's vector values. See my blogs of 21 Sept. 2010 and 19 Feb. 2011, for example.

** Human life, for example, itself depends upon additional plant and animal species for food, air, digestive microbiota, etc. Species of mechanical life might best be supported by human and other biological species?

Pet peeve

Journals that don't publish errata even when it's for their mistakes.

Future space stations

 AI and robotics can make "man-tended" stations* more efficient/economical than they were 50 years ago. "Permanently manned" stations might not be able to compete. What specific tasks do we need humans for?

* In low earth orbit or elsewhere. Also see my blog of 3 Nov. 2010.

Most intelligent

My most intelligent AIs appear to have been societies of specialist A.s.a. H. agents. ("Collective Intelligence") But I may not have explored a wide enough range of environments yet so this is a tentative assessment. Also, I do not assign any scalar "IQs" since I prefer vector measures of intelligence. 

More than one kind of AI

Most scientists currently recognize 15 to 20 different species of early humans. Similarly I expect there to be a number of different artificial intelligences* having different architectures. In view of specializations this may well prove desirable.  

* See, for starters, my blog of 16 September 2010.

GPU prices

 GPUs are way overpriced, especially Nvidia. Sure glad Asa can use different technologies.

Comment on OpenAI

 My big issue with OpenAI is that it's not open. Their software is not open source. This is not how good science is done. People need to be able to examine and comment on one another's work.* Perhaps the business men need to be kept away from the scientists and their work.** I have commented previously on chatgpt and the alignment issue.***

* Groups make better decisions than individuals do.

** And it would be important to know just which members of the staff did the actual science as opposed to who were administrative/management types. I get the feeling the presenters aren't always the developers. This happens less in an academic setting. Probably OpenAI should have been started as an institute at some university.

*** See, for example,  my blogs of 25 February 2023 and 10 August 2023.

Immortality

 Having exceeded my life expectancy it's natural to think about mortality.* I know of a number of possible ways in which some sort of immortality might be possible (pluralism again):

1. Uploading of one's mind to a computer.** Not available yet.

2. The block universe. "I'll always be alive in the year 2020." (One only needs the block universe model to be valid for the past. Any problem with quantum mechanics influencing the future*** is not an issue here. Once I'm dead my whole life is in the past only.) 

3. The conservation of quantum information. (If the real me is just information, e.g, "I think therefore I am.")

4. Recurrence. Under the right laws of nature if you wait long enough I'll be back.

5. Everett Quantum Mechanics. Any time I die a different copy of me survives. Like Schrodinger's cat. (There wouldn't have to be a single continuous "consciousness.")

6. "The Soul Hypothesis."**** Is the "real me" a wave function, or a quantum field, or ???

7. Metaphysical solipsism. Your mind is the only thing that exists, ever. You're around for just as long as the universe is.

8. If what we are is a network of ideas then recording and preserving those ideas.***** (Related to number 1 above but perhaps without an ongoing active element.)

9. Combinations of various of these.

But once again, extraordinary beliefs require extraordinary evidence. It seems appropriate to publish this on Dia De Los Muertos.

* I've argued against the idea of immortality in my blog of 15 October 2010.

** I laugh at this idea when I think of some of my old computer programs that can not be run on any of the computers that I have today. (Think coding forms, punch cards, and magnetic tapes.) And hardware updates have sometimes forced me to make software changes.

*** For example, Sabine Hossenfelder, backreation.blogspot.com, 23 July 2022.

**** For example, Baker and Goetz, eds., Bloomsbury Pub., 2010.

***** Including "lifelogging."

Adding more senses again*

 A.s.a. H. decomposes sensory input into hierarchical sets of commonly recurring patterns. Some concepts (higher up in the hierarchy) involve A.s.a. learning nonlinear functions. With the addition of more sensors it is sometimes possible, instead, for A.s.a. to learn a simpler, possibly even linear, function.

* See, for example, my blog of 1 September 2022.

Robot teach pendant

In my electronics lab I've tried out various remotes (wired, IR, microwave, and radio) for use as robot programmers.* So far a game controller (Logitech) looks to be the best.

* For use like in my blog of 1 May 2023. (And it is important to base your control decisions only on sense inputs the robot itself sees.)

Learning knowledge graphs

A.s.a. H. learns (builds) knowledge graphs (semantic networks) as described, for example, in my blog of 4 March 2015. (It's also possible to do the reverse, construct an A.s.a. agent starting from a knowledge graph.)

Electron space charge ion heating

 Arnold et al have a new study of electron space charge ion heating.* I tried to promote this idea starting in the late 1970s both during pellet refueling and in other scenarios.**

* Parallel expansion of a fuel pellet plasmoid, arXiv:2310.05678v1, 9 Oct. 2023.

** See, for example, R. Jones, Electron Space Charge Ion Heating in an Aerosol Discharge Plasma, Ind. J. Phys., 55B, 397, 1981.

I've traded space for time

When I quit teaching (and university service) I had thought I'd have more time freed up for research. But it meant also giving up lab and office space. My library is probably a bit easier to use now that it's all in one building. Software development is perhaps a tiny bit harder since the desks I have at home* are a bit smaller than what I had at ESU. Lab space is the biggest issue. Most hardware must be stored away to save space. This means time is required to find and get things out, do the experiment, and then put things away again. Everything can't just be left out and in the way.* The required space has been reduced but more time is needed to do most experimental work. A careful reorganization of the laboratory equipment may help somewhat. 

* I've sometimes used the dining room table. Temporarily. See my blog of 7 January 2012, second photo.

Valuing diversity

A society of A.s.a. agents should value a broad range of specialists.

Launch fever again?

 Once again there is a lot of impatience to launch Spacex Starship. With the long delay in the self-destruct system if Starship Superheavy had gone horizontal on its first launch a lot of people could have died. Would the project survive something like that? 

Raspberry Pi 5

I see that they hope to have Raspberry Pi 5s available late next month. I hope to get a couple of them. The processor is about twice as fast as the Pi 4. I only wish that they had increased the RAM like they did when they went from the Pi 3 to the Pi 4. The active cooler does look nice.

On the importance of concepts again

In their new paper* "Concepts is All You Need: A More Direct Path to AGI" Peter Voss and Mladjan Jovanovic stress the importance of concept formation for AGI and, as with A.s.a. H., they suggest that concepts should be encoded as vectors in a hierarchically organized structure. They also note the importance of sensory input and output actions.

* arXiv:2309.01622, 4 Sept. 2023  

Continental philosophy

 My A.s.a. H. project can be thought of as experimental phenomenology since "phenomenology studies the structure of various types of experience ranging from perception, thought, memory, imagination, emotion, desire, and volition to bodily awareness, embodied action, and social activity, including linguistic activity."* A.s.a. involves all of these.

* Stanford Encyclopedia of Philosophy, 16 Dec. 2013.

Rational voting systems, auctions

 I have previously discussed the fact that 51% of the population should not get 100% of things their way while 49% of the population get 0% of what they want. There are various ways to set things up so the 51% majority only gets 51% of their objectives. In a legislative setting, for example, each legislator can be given an equal amount of "currency" at the beginning of a legislative term. Each legislator gets to spend (bid) as much of their "currency" as they wish on each of the scheduled votes. On some votes the minority party will prevail over the majority depending upon the importance of each vote to each individual's priorities, the idea of saving your strength for the fight that matters to you.

Space junk

The problem of debris in earth orbit (and elsewhere) is another example of the tragedy of the commons, something capitalism is not good at handling. 

Version control indexing

When I start a new experiment with A.s.a. I am never sure if I should "write code from scratch"* or spend time hunting for legacy code** that I seem to remember is "almost" what I need. Part of the issue is with some kind of adequate indexing (which I don't currently have). The index really needs to be a vector quantity or at least a bag of words.

* Hardly ever completely from scratch but rather from some more fundamental example in legacy code.

** Perhaps a long time as I have thousands of AI programs in my code library.

More sensors more intelligence?

 I have previously asked whether having more sensors would help make systems more intelligent.* One reason for believing this would be the so called critical learning periods: "Critical learning periods are time periods early in development where temporary sensory deficits can permanently damage the outcome of learning."** Critical learning periods are known to occur in machines as well as in animals.**

* See, for example, my blogs of 17 July 2014, 2 September 2014, 11 December 2014, 11 November 2013. I have a small instrumentation lab to address this need.

**arXiv:2308.12221v1 23 August 2023, Kleinman et al and my blog of 17 July 2014.

The AI alignment effort

 I am not onboard with the current feverish alignment effort. I don't want AIs to share human values*, I want them to have something better.** Perhaps including Blackburn's plurality of concerns.***

* You only need to look at crazy Donald's cult of supporters by way of example. 

** See, for example, my blog of 25 October 2011.

*** A vector value system again. See chapter 8 of Think, OUP, 1999.

Vector values* again

Shuwa Miura has shown** that there may exist no scalar Markov reward function that characterizes some set of acceptable policies but that a set of deterministic policies can always be characterized by  multidimensional Markov rewards.

* See Capitalism is Wrong, chapter 2 of my book Twelve Papers, on my website www.robert-w-jones.com under "Book" and my blog of 19 February 2011.

** On the Expressivity of Multidimensional Markov Rewards, arXiv: 2307.12184v1, 22 July 2023.

Humanoid robots?

 Human looking robots have always been a popular form factor. I have two small humanoids and the makings of at least three more. These have not proven to be the best engineering choice for any of the experiments I've been doing so far. Of course much of my robotics work has been for symbol grounding of my A.s.a. H.

Analog computing

There is renewed interest in analog computing* because of the reduced power demand. I have a few B.E.A.M. robots** but digital processors allow for much faster modification of behaviors (via software changes).  A hybrid system is possible, however, using analog circuits in various of A.s.a.'s pre and post processors for example.

* See my blog of 20 January 2012.

** See my blog of 18 February 2016.

AI kollektives unbewusstes

While many of my A.s.a. H. agents have been specialists they do typically share many mental concepts, reflexes, and instincts with one another. Things like "pupillary" light reflex*, sensation, etc.

* and general sensor auto ranging (another preprocessor)

Arduino Uno Rev 4

While I managed to load the first layer of an A.s.a. H. light program* onto an Arduino Mega** the Arduino Uno Rev 3 proved inadequate for an even lighter version than this. According to its specs the new Arduino Uno Rev 4 should be adequate, however, and so I've bought a couple of these to try out.  

* Essentially the code of my 14 May 2012 blog but with input and output through the Arduino gpio and having inputs accepted whenever one or more of the inputs change significantly: | new In(N) - old In(N) | > (C)( old In(N) ).

** See my blog of 1 January 2023.

AIs working as scientists*

 I have noticed that AIs are being used more and more to generate and to compare various models of phenomena.** Models of the proton involving charmed quarks being one example.***

* An example of a specialist AI, but in software rather than hardware.

** Such as my use of neural networks to model plasma confinement: for example, my paper in Trans. Kansas Acad. Sci., vol. 98, num. 3-4, pg 144, 1995 and Experiments With Machine Creativity, chapter 5 of my book TWELVE PAPERS, tables 4 and 5.

*** For example, in Nature, vol. 608, pg. 483, 2022.

Specialist agents

Division of labor, specialization,* is one of the ways I've been trying to deal with complexity and the curse of dimensionality.** Typical social insects have something like 3-4 castes. Many small multicellular organisms have 10 or more cell types. The U.S. Bureau of Labor Statistics classifies all workers into 867 occupations divided into 23 categories. How many specialized agents one uses should depend upon the environment you're working in and the tasks you're attempting to perform. I currently have 52 small robots*** in my lab and the makings**** of about 20 more. Software is even more varied.

So what specialized hardware has proven to be useful/required? 1. wheeled (and tracked too?) 2. fliers 3. swimmers 4. walkers? (quadrupedal? bipedal too?) 5. pick and place (arms) 6. gantry? 7. snake? 8. hybrid? (for stair climbing?) 9. various grippers/manipulators 10. several different processors? (and memories?) 11. different sensory systems

* In both hardware and software.

** See, for example, my blog of 18 August 2021. 

*** This includes a fair amount of duplication. It can take a couple of tries before I manage to get a workable robotic agent capable of performing a given required set of tasks.

**** Lego, Vex IQ, Vex EDR, Meccano, etc.

A.s.a.'s pains

 I have explored various attention mechanisms. Pain is another of these. A.s.a.'s hunger (low battery voltage) can trigger a search for a recharging station. High temperature measured on a gripper can trigger withdrawal of the manipulator. High transverse accelerations while moving along rough terrain can cause a robot to slow down...... 

Transfer learning

With computer vision it is conventional to pretrain lower levels in the network to represent features like edges and colors. These are then available for later application when learning new images. A.s.a. H. was designed to have a generalized version of such transfer learning.*

* See, for example, my papers Trans. Kansas Acad. Sci. vol. 109, pg 159, 2006 and vol. 120, pg 108, 2017 and my blog of 5 Nov. 2015.

Pluralism again

A. C. Grayling has said that truth is an idealization, something to be approximated toward as a limit. I make use of multiple theories of truth, e.g.:

1. correspondence   2. coherence   3. consensus 

1. I use observation and experiment. 2. I see how it fits with my prior models of things. 3. And I see what models other workers use.

ChatGPT on capitalism

 "I don't think the current state of capitalism can withstand itself forever. Debt levels are becoming unsustainable and the concentration of wealth in the hands of a small percentage of the population is leading to social and political instability. It's only a matter of time before the system collapses and something new takes its place. However, I believe that this is a necessary part of the cycle of economic systems and we will eventually move on to something better." ChatGPT

Robot training

 It is common to begin by giving an AI examples of human behaviors.* Similarly, I often begin A.s.a. robot training by driving the robot around using remote control,** recording the sensory inputs and the actions taken.

* See, for example, R. Jones, Trans. Kansas Acad. Sci., vol. 100, pg 85, 1997, vol. 102, pg 117, 1999, and/or vol. 107, pg 32, 2004.

** radio, infrared, or tether

Sense of self

 A concept of self can develop in A.s.a. H. with just a few computational layers.* This makes me wonder if some very primitive creatures don't already have a sense of self. Of course A.s.a. H. was not biologically  inspired** so this is complete speculation on my part.

* See, for example, R. Jones, Trans. Kansas Academy Sci. vol. 120, 2017, pg 108 or my blog of 21 July 2016.

** There is, however, its relationship to adaptive resonance theory which is biologically inspired.

Twenty Years

A.s.a. H. has now become a twenty year long project (and still going).  

Hierarchical planning

Chaining of cases is possible on each level of A.s.a. H.'s hierarchical case-based memory.* This allows for hierarchical planning** at various orders of abstraction. On lower levels of the memory hierarchy this involves subsymbolic representations. On a higher level this can involve natural language.*** Reasoning can be sped up in this way.

* See, for example, Trans. Kansas Acad. Sci., vol. 109, no. 3/4, pg. 163-164, 2006.

** and deduction

*** or, alternatively, a mentalese

Hard to define concepts

 Even cognitive scientists are not in agreement about what constitutes things like consciousness. As we teach A.s.a. H. language we can see what conceptual structures the various words are labeling.* For example, we find, approximately:

conscious = (sense, think, health)

alive = (sense, act)

self = (sense, think, act, health)

where sense, think, act, and health are, in turn, defined roughly** as described in my blogs of 1 October 2015 and 21 July 2016.

* But one must recall that A.s.a. H. has a dynamic memory so individual memories/concepts can change somewhat as a result of new experiences. See my blog of 1 January 2023. 

** Different experiments have involved various robots having different sensors and servos, etc so these concepts can vary accordingly.

The importance of category formation

Yuan has emphasized the importance of category learning for artificial general intelligence and XAI.* A.s.a. H. employs just such a system.**

* Yang Yuan, A Categorical Framework of General Intelligence, arXiv:2303.04571v1, 8 March 2023.

** See, for example, R. Jones, Trans. Kansas Acad. Sci., Vol. 120, pg. 108, 2017 and my blogs of  1 Oct. 2015, 21 July 2016, 19 Oct. 2016, and 3 Aug. 2018.

Spacex Starship needs a launch abort system

 By launching the crew on Orion and rendezvousing with Starship in earth or lunar orbit one can sidestep the problem for now. But someone might be tasked with looking at ejector seats and things like the old M.O.O.S.E.* bail-out from orbit ideas.

* General Electric's Manned Orbital Operations Safety Equipment. See, for example, Teitel, Discover magazine, 12 Oct. 2017 or the wikipedia article on MOOSE.

Black boxes, explanations

 People typically can give brief, simple, explanations (reasons) for their beliefs and actions. These may, in fact, be an approximation to what is really a network of subconscious motivations. 

Sometimes A.s.a. H. takes an action or makes a decision that I don't expect/understand. When that occurs I usually try to explore the hierarchical patterns of activation that lead to this action/decision. I have found that frequently a number of weak activations may have "summed up" to produce the unexpected outcome. (Like when a large number of small weights in a neural network act together to produce some output signal.)

GPT-3, LaMDA, and ChatGPT

In my 28 January 2013 blog I pointed out that Siri did not have all of the functionality that I would want a machine intelligence to have.* Most importantly, it had no value system.** I would level the same criticisms of ChatGPT, GPT-3, and the like. Not to say that I don't find them interesting and useful.

* See, for example, my blog of 29 September 2010.

** I have argued that how intelligent a creature can be depends upon how good a value system it has. See for example my, Science Is Not Value Free, Trans. Kansas Acad. Sci., vol. 119, no. 2, pg. 249, 2016. One such value system is shown in my blog of 21 September 2010.

Hacking toys, climbing stairs

Various research groups* have used toys as robotics hardware platforms in order to cut costs. I bought a "Tiger Twister stunt vehicle" (Speed Well International Industrial Ltd.) because it is supposed to be able to climb stairs and is large enough to carry a Raspberry Pi single board computer. I find that the Tiger Twister does (for the first time) give me a limited stair climbing capability** provided that the stairs are not too slippery.

* For example: Hogg, Martin, and Resnick, Braitenberg Creatures, MIT Media Lab, E&L Memo report #13, June 1991. Connell, Creature Design with the Subsumption Architecture, International Joint Conference on Artificial Intelligence, pg 1124, August 1987. Dawson, Dupuis, and Wilson, From Bricks to Brains, UBCPress, May 2010. my blog of 8 Jan. 2018.

** There are LEGO robot designs that also have limited stair climbing capability.

A.s.a. observation

Occasionally an input vector* will fall into two different categories/classes. These categories may trigger different responses/outputs. This is relevant to the idea/experience of "free will."**

* This can be the "input" to any level in the A.s.a. H. hierarchy. It need not be the input on the bottom most layer, i.e. the sensory input from the environment.

** See, for example, my blogs of 8 July 2011 and 21 Jan. 2015.

Intentions

Some of the cases in A.s.a. H.'s case based hierarchical memory include actions (occurring at one time step or another). During the course of operation when such a case becomes activated any such future ("planned") action becomes one of A.s.a. H.'s intentions.

Alternate realities

 In his new book* Michael Wolf asks: "If natural languages deeply influence the forms that our thinking takes, could they be so different that we 'live' in substantially 'different worlds'?" My blog of 15 June 2022 lists some of the alternate realities that exist. But I don't think that language is as important a factor in this as Wolf does.  

* Michael Wolf, Philosophy of Languages, Routledge, 2023, page 28

Mary's room

 Frank Jackson asks what would happen if someone lived their entire life in a black and white environment and then was suddenly presented with colors.* I have performed just such an experiment with an A.s.a. H. robot. In the black and white world the visual input is a grayscale SCALAR quantity. With the advent of a colored environment this transitions to a VECTOR quantity. In the case of humans it would be a four dimensional vector quantity; one component each for rods, "red" cones, "green" cones, and "blue" cones. (Other animals and robots may have something other than a four dimensional vector input.)

* Epiphenomenal Qualia, The Philosophical Quarterly, 32 (127), 1982,  pg 127.

Commercial space stations?

There is, of course, a need for commercial launch vehicles for things like communication satellites, weather satellites, remote sensing satellites, navigation satellites, space telescopes, military satellites, scientific research, etc. But is there a need for  commercial manned space stations? More than just a man tended station?  

Spacex starship

There has been growing frustration with starship-superheavy not having had an orbital launch attempt yet.  But a big enough failure (perhaps with loss of life) could end the project entirely. 

I also worry that starship makes an ungainly lunar lander.

tinyML

I was able to run a single layer* of a light version of A.s.a. H.* on an Arduino mega.

* Very similar to the code in my 14 May 2012 blog. 

Vector currency

Since "needs" and (therefore) "value" is a vector quantity* I have suggested that a vector currency would be a more accurate measure of value than present day (scalar) money is. Something similar was in use during the 1940s where various different coupons were used to buy meat, eggs, fats, cheese, bacon, sugar, milk, tea, coffee, gasoline, fuel oil, tires, bicycles, typewriters, and various items of clothing.

* See my paper Capitalism is Wrong. (One version of which is in my book, Twelve Papers, on my website: www.robert-w-jones.com) 

Algorithms for artificial intelligence

In 1986 Griffiths and Palissier published their brief book Algorithmic Methods for Artificial Intelligence. In my 2011 paper* I decomposed thought (and, specifically, A.s.a. H.) into remembering, generalization, comparison, explanation, deduction, organization, induction, classification, concept formation, image manipulation, feature detection, analogy, compression, simulation, and value assessment. In order to implement these processes on a computer each of them was, in turn, decomposed into: sorting, searching, vector averaging, vector differencing, vector dot product, sensitivity analysis, renormalization, interpolation, extrapolation, concatenation, time warping, and image manipulations.** Mid 2023 Stefan Edelkamp is expected to release his book: Algorithmic Intelligence: Towards an Algorithmic Foundation for Artificial Intelligence. I am anxious to see what he has to say. A google search has a number of sites listing typical AI algorithms as being: classifiers, clustering, and regression/forecasting. Some also add deduction, dimension reduction, and neural networks to the list.

* Trans. Kansas Acad. Sci., vol. 114, no. 1-2, pg 162, 2011. 

** There's no claim that these decompositions are unique.

Dynamic memory

 In basic theories of dynamic memory* individual memories are constantly changing as a result of processing new experiences. One example of this is found in my A.s.a. H.**

If one wanted a memory theory of personal identity ("self") it would then have to be one built out of causal sequences developing through time.

* See for example R. C. Schank, Dynamic Memory, Cambridge University Press, 1982.

** See for example equation 1 of R. Jones, Trans. Kansas Acad. Sci., vol. 109, pg 159, 2006.

Book buying

Despite reviews and sample pages at least one third of the books I buy from online bookstores turn out to be something I'm not really interested in. I prefer brick and mortar bookstores where I can better see what I'm getting.  

The fact that I was able to pare down my library by about 1/3 when I moved out of my ESU office is undoubtedly related to this.