Is mathematics invented or discovered?

 I was not trained as a mathematician so perhaps I should keep this to myself but I am going to answer in one word: YES. By that I mean that I think some of math is invented and some of it is discovered*. I suppose that one might first think about what mathematics is. Again, as a non expert my view has been that mathematics is: 1. A science of patterns. 2. A language. 3. A game. When I use mathematics to talk about patterns I see in nature I have discovered things about those patterns. Perhaps I describe mechanics using Newton's laws of motion and then discover recurrence phenomena. If, on the other hand, I am creative and extrapolate I may be able to create mathematics that has no parallel in the natural world. In that case I have invented something, perhaps new concepts** that I will later use in my languages.

* Wikipedia says most philosophers of mathematics believe that mathematics is discovered rather than invented.

** One of the longterm goals of A.s.a. H. is to invent new and useful concepts with which to think (new or old thoughts).

Python vs C

 I have not quantified the differences but for a couple of my A.I. robotics experiments Python proved to be too slow and I had to switch over to C.

A.s.a. hardware upgrade

 In the last week I have built up a new computer system with a clock speed of 5-6 GHz, 14 cores, 64 GB of RAM, 2 TB of internal SSD, and many TB of external memory. I did an A.s.a. H. extrapolation/creativity experiment to try it out.

Co-intelligences and disembodied A.I.s

 In my post of 25 August 2024 I describe how machinery augments my biological intelligence. Conversely, biological senses and intelligence can be used to supplement/augment a computer system. For example, we can replace one or more (all?) of the lowest layer elements of the concept hierarchy* in my post of 4 September 2024 with human input. If the lowest layer is completely replaced with human supplied inputs we then have a disembodied A.I. (This may or may not be desirable.)

* as well as some preprocessing, like the ANNs in my 4 September post.

dia de los muertos

A year ago for the day of the dead I listed a number of ways in which we might be immortal but I forgot to mention cloning (perhaps in combination with life logging across generations).

2024 US election

It should be impossible for Donald Trump to be elected US president. The fact that it isn't is proof once again that the human value system is inadequate. It is not necessary to "align" A.I. s' values with human values. It isn't even desirable.

Space "race"

Large projects must be broken up into pieces and the work distributed among many cooperating specialists. For these specialists to coordinate with one another there have to be schedules.

Work can be done either too fast or too slowly. Attempts to work fast push up costs and lower safety. The death on the first soviet manned Soyuz flight is an example. Work can also be done too slowly. The fixed overhead that must be covered year after year is one issue. Low launch rate also unfavorably impacts worker skill levels.

But it seems to me that the U.S. has no need to "race" at the current moment.

If the soviets had been first to land a crew vehicle on the moon that might well have led to a stronger American post-apollo effort. Similarly, today, if the Chinese land a crew on the moon before the U.S. that might lead to a stronger U.S. program going forward.

I also think that the payback from space telescopes and other unmanned space assets has exceeded that of the manned programs. Crew should only be included when they are actually needed for the mission. Also, see my blog of 3 November 2010. 

Musk

If you try to do too many things at once you end up doing none of them well. Some years ago I criticized Spacex for this. I would now suggest that Musk is trying to do too much: spacex, starlink,  electric cars, solar power, AI, robots, twitter, neuralink, politics, boring, ... I don't care how many good engineers he hires, this is too much. 

Robot skin/clothes

 My AI robots require a lot of sensors and, therefore, a lot of wiring:

One should, of course, first shorten all the wires as much as possible. It is possible to then cover the wiring (and the bot) with "skin" or "clothing." This can be "resistive rubber" sheeting (velostat, linqstat) or more conventional cloth with embedded flex/force sensing thin film ribbon sensors. Any tethers can have their wiring similarly covered.

Attention again

Typically, each layer of the A.s.a. H. concept hierarchy broadcasts forward (upward) its N strongest outputs*.  It is also possible for layers to only accept the M strongest inputs that they see (from the one or more layers beneath it).  Feedback can be incorporated if layers are allowed to see outputs from above them.

* R. Jones, Trans. Kansas Academy of Science, vol. 109, pg 159, 2006.

"mind-only" Idealism

Why have I not used idealist models?* I think that my biggest problem with idealism is the following: one of the most fundamental distinctions we make is between "ourselves" (what we can immediately/directly control/influence) and "the external world" (that which we can't directly influence). "Inside" and "outside." "Me" and "not me." If idealism were true couldn't we control everything we "see?" 

Of course we could simply make idealism more complex. We could suppose there are other minds and they control those things we can't. But when I postulate the existence of other minds there still seems to be an "outside" that NONE of us can directly control. (If rocks, for example, were  other "minds" they don't seem to be like "me" at all. So could I consider them to BE "minds?")

Or perhaps there are "laws of thought" that limit what can be influenced? If so, what are these laws? Is idealism simply poorly developed, i.e. in a pre-theoretic state of development? That part of our experience that we can influence would be related to consciousness while things like rocks might be related to something like a subconscious.

I also find it hard to come up with idealist models and then try to make practical use of them. 

* See my blog of 1 April 2024.

teaching, attention, syllabus/curriculum

Teaching an AI agent to an idealized (noise and distraction free) syllabus is another way of imposing attention as well as speeding up and reducing the cost of learning.

Decomposing the AI into a society of specialists also helps with attention as well as helping to deal with complexity.

Concept structure

The approximate concept structure early in the life of one particular A.s.a. H. robot. This diagram can not display temporal sequencing.

Are expert systems obsolete?

 Knowledge-based expert systems remain a good choice for the control of machines and processes. See my paper in the Transactions of the Kansas Academy of Science, vol. 101, pg 132, 1998. Another specialist agent.

Co-intelligence

For more than three decades (the 1960s, 70s, and 80s) I spent a good deal of time in library stacks. When I went to other universities and conferences I'd spend time in their libraries. I also knew where all the bookstores were. Nowadays most of my library research is done with search engines. As Turing said, search is an important part of artificial intelligence.

In the 1960s I used slide rules and function tables. By the 1970s those got replaced by desktop and then pocket calculators. These, in turn, became programmable.

Beginning in the early 1970s by the end of the 1990s my plasma physics experiments were pretty much completely automated.* 

In the 1950s we'd tell the auto club where we wanted to go on vacation and they would send us maps with the routes marked out. Today Siri will direct you as you drive.

And then there's word-processing, email, spam filters, texting, spellcheckers, autocomplete, compressor/summarizers, code completion, self-parking, self-driving, GPT, etc. 

Each of these is a specialist.

Although a lot of my work has been taken over and done for me I've still found plenty more to do.

* See, for example, Trans. Kansas Academy of Sci., 101, pg 132, 1998 and 99, pg 85, 1997 and my blogs of 1 and 2 November 2011.

More robot pain sensors*

Brixo (Dakott) conducting "Lego" bricks can be used as pain (breakage) sensors. They complicate the robot design a bit but the final product looks cleaner**. 

To get greater sensitivity (detecting cracks rather than complete breakage) one can glue a conducting tab*** to conventional Lego and then butt that up against a Brixo brick. One can also space conventional insulating bricks**** between 2 Brixo bricks and then run a conducting paint line between them.*****.

* See, for example, my blogs of 29 Sept. 2015 and 31 March 2016.

** less wires dangling

*** cut from aluminum foil or thin copper tape

**** one or several

***** Between the Brixo bricks that is.

In order, in order to learn

One would not try to teach a student physics by starting with quantum field theory or general relativity. Rather, we begin with measuring time and distances and then work toward Newton's laws of motion. I have stressed the importance of the syllabus used when teaching A.s.a. H. I have also stressed the order in which the concepts are presented*. 

It looks to me like knowledge is rather disorganized when it is dumped into LLMs. I believe that is a mistake. The output of a highly nonlinear system not only depends upon its inputs, it also depends upon the sequence/timing of those inputs.

* See, for example, my blog of 12 September 2020.

Model decomposition efforts

 An (isolated) world model is required for a number of AI applications. So we would like to separate world model* from  agent** model, action model, any pre and post processors, etc. Specialist A.s.a. agents do this to some (limited) degree and even (partially) decompose the world model into specialized submodels*** (i.e., knowledge organization). The mechanisms at play in that case are the specialized (and limited) training curriculum plus the subsequent tasks and work environment selection.

* "laws of nature"?

** observer?

*** scientific fields?

GPT passes minimum Turing test

Although I have been critical of both LLMs and Turing tests it is still interesting to see Jones and Bergen's paper.* I generally do not think science is about tests, benchmarks, and metrics. A number of specialists and organizations have been rather premature in switching from doing science to doing engineering. Work done in industry tends to be rushed, seeking a product and a profit. Pure science is then best done in academic institutions. Of course the capitalists are trying to turn everything into a business including universities. This is not helpful.

* Does GPT-4 pass the Turing test?, arXiv:2310.20216, v2, 20 April 2024.

Tingling sensation

Vibration can cause Lego bricks to separate enough* to allow pain sensor contacts to lose connection. Continued vibration then causes bouncing contacts, a "tingling" sensation is recorded. 

* a "crack" forms

US Supreme Court opinion

 So let me get this straight. Aren't they saying that Joe Biden can have seal team 6 assassinate Donald Trump tomorrow and he (Biden) will be insulated from criminal prosecution because he is the current president? That's nuts. The GOP has a lot to answer for. I fear for the future of my children and grandchildren.

Is this a new Gesetz zur Behebung der Not von Volk und Reich?

Concept grounding*

What sensors are needed to ground the core concepts learned by an intelligent robotic agent**? For some time I simply added sensor types as soon as they were available for use***. But in retrospect I believe that essential ones were:

force sensors, accelerometers, temperature sensors, light sensors, color sensors, gas sensors, anemometer, gps sensor, rain/water sensor, microphone and speaker, humidity sensor, voltage sensors, current sensors, ultrasonic and/or IR range finders, line sensor, servo/motor encoders, and neural network preprocessors****.

Combinations and sequences of these raw input signals together form primitive core concepts.

* Rather than symbol grounding.

** Or society of agents. 

*** For example, my blog of 1 October 2015.

****Possibly classifying input from cameras.

A.s.a.'s values

In our early publication* A.s.a.'s value module was placed at the top of the case/concept hierarchy. This can be true early in training and/or for small agents**. But as training proceeds more and more concepts are acquired/created and these are chained and the hierarchy grows upward. I.e., our values need not be our deepest/most abstracted concepts**.

Furthermore, in the light version of A.s.a. H. 2.0 published in our 10 February 2011 blog we value pattern occurrence frequency and pattern size on each level of the hierarchy.   

* Trans. Kan. Acad. Sci., vol. 109, pg 159, 2006.

** Or by use of pass through across layers.

Another argument for vector utility

Diversity. There are multiple niches, multiple ways to improve a creature's fitness and increase its chances of survival and spread. Beneficial traits, different adaptions, like protective coloration, larger brains,  etc.  Different creatures may find very different ways to improve.

AI alignment/safety

If humans and AIs are to peacefully co-exist it is not a matter of giving AIs our (rather poor) set of values. Rather, human values will also have to change (improve).  

The further evolution of intelligence

 There is an upper bound placed on human brain size by what can be pushed through the birth canal. There is a lower bound placed upon the human computing elements by minimum cell size. There is a speed limit placed upon human computational processing by neural electrochemistry. Nonhuman processors relax all of these limitations. This all favors mechanical life.

Specialization as attention

 For a long time I have stressed the importance of attention and explored various attention mechanisms*. Agent specialization is an attention mechanism as well as an approach to reducing complexity/dimensionality**. There are some senses/sensors a given specialist may have no need for at all. Specialization is attention to some specific set of skills/tasks in a particular environment.

* See, for example, my blogs of 1 June 2011 and 3 January 2018.

** See, for example, my blog of 1 July 2023.

Fair use or plagiarism?

Is ChatGPT's use of massive amounts of human generated data "fair use" or plagiarism? The problem is ChatGPT regurgitates chunks of input that are too big.

Co-intelligence

Existing AI agents have gotten good at searching and answering questions we put to them but they need to start asking questions of us unprompted.

Ion ramjets

I see that Andrea Lucca Fabris and Mansur Tisaev of the University of Surrey Space Centre are now working on electric ion ramjets for satellite drag makeup. My friend James E. Cox was developing that idea at TRW in the 1970s. I have his TRW report* and he and I discussed the idea at the time as it related to my own work** on ion electric space propulsion.   

* TRW Systems Group 4753.75.3-003, Applications & Feasibility of An Electric Ion Ramjet, 13 January 1975

** See, for example, I.E.E.E. Transactions on Plasma Science,  vol. 10, pg. 8, 1982

Humanoid AI robots

 For the reasons that I have discussed previously* I do not think that the majority of AI robotic agents will have a humanoid form.

* See, for example, my blogs of 1 Oct. 2021, 1 July 2023 and 10 April 2024.

AI power demand

 The time, computational resources, and power required by any given AI specialist agent depends upon the tasks being performed, the environment, the algorithms employed, and the agent architecture and will vary from training to performance. Teaching a common robotic pick and place task via one-shot learning from demonstration requires roughly equal computational power for training and for performance. Traditional backprop artificial neural networks require much more computation during training* and much less during performance. On the other hand for agent specialties where performance requires substantial inferencing then that may actually be more costly than initial training was.

* I've typically been training A.s.a. with handcrafted curricula rather than using the vast amount of data that systems like GPTs require.

Human-robot interaction

 It is usually argued that robots with a humanoid form factor will more readily adapt to the existing human environment. But I find that in most applications we modify the environment to be more robot friendly*. In making use of Roomba, for example, I reduced the clutter and set up virtual walls. Robot lawn mowers typically have a boundary wire. Similarly, factories and warehouses are modified to accommodate the robots that work there. As robots get smarter they may do more of the adapting and we may do less. But if we cohabitate with robots I expect the result will be a compromise, not our current environment, just as factories have long been tailored to both men and machines.

* I think we also modify our activities/behavior in adapting to the use and limitations of software.

Do you use what you studied in college?

For some careers you will use what you learned in college. For others you may not. In becoming a plasma fusion research scientist I certainly used my undergraduate physics major. When I taught at the National University of Singapore I saw that businesses were eager to hire physics majors because they knew our graduates must be smart, worked hard, and were 4 years more mature than high school graduates. The college degree filtered out the better candidates for the employers.

Pluralism

 Kant argued that one can not know the thing in itself (das ding an sich). Rather, we all create models of the world. I have previously noted* 5 such alternate realities.** 

1. materialism  

2. idealism

3. dualism

4. theism

5. multiverses

In my pluralism I most frequently employ #1. Next most frequently I employ #5. I infrequently use #3 and hardly ever use #4. I never seem to employ #2. I could probably keep detailed statistics on this but it's surely dependent upon what's being studied.

* See my blog of 15 June 2022.

** There are, of course, variations*** and combinations**** of these 5.  (Neither do I insist upon keeping the number to 5.)

*** If ultimate reality was composed of simply quantum fields or the wave function that would be a variation of #1 (provided "materialism" refers to any of the concepts defined and employed by modern physics).  There are, of course, many variations of #4.

**** #3 and #4 are frequently combined, for example.                                                                                                                                                      

Robot disconnecting randomly

I was trying to transfer files from one computer (onboard the robot) to another (external, fixed) computer. My system was disconnecting but apparently only when the robot is moving. Seems to be (USB) connectors. I could hard-wire everything but that makes (subsequent) reconfiguring of things so much harder. Maybe some kind of clips/clamps on the connections to reinforce them, hold them tight? Or something like the old RS232 screw connectors.

Physical theories of everything

 Each of our concepts typically has some limited range of applicability.* Are there suitable** concepts that are universally valid? If not, then it might not be possible to create a single physical theory of everything. Rather, one might need a network of more limited theories, each of which has its own range of applicability. Perhaps a concept hierarchy (like the one A.s.a. H. creates) with theories summarizing the patterns that are seen in various portions of the concept (knowledge) web.

* See, for example, my blog of 30 May 2018.

** Physical? (As opposed to mathematical for example.)

Why should you learn to code?

Jensen Huang, CEO of Nvidia, is claiming that we should no longer learn to code. Suppose he is right. Is writing a specification easier than coding? Some computer scientists will tell you that "Formal specifications are often just as hard to read, and almost as hard to write, as code." I then must ask, is writing prompts easier than coding? Do you get what you intended in each case?

As for me, I needed to know some coding for the same reason I needed to know some digital electronics. I wanted to better understand how the computation actually occurs and in some adequate degree of detail.

Understanding occurs over various levels of abstraction. For some purposes it is sufficient to have a broad overview. For other purposes one needs to deal with finer details on a less abstracted level. I.e., there is a concept hierarchy. Some things can only be understood by working on/at the right level of abstraction using the right concepts.  And concepts that work on one level of abstraction may not be valid on another level.

Embodied AI

 Some researchers believe that embodiment is essential for AGI while others believe it imposes a bottleneck. In providing an AI with some specific sense or a particular concept I have sometimes found simulation to be easier.* In other cases I have found embodiment to be easier.**

Simulations are always an imperfect model of reality. Ultimately we want an AI to have contact with and operate in the real world. It can sometimes be faster to begin AI training on simulations, however.

* Recharging ("feeding") was one example.

** For examples see my blog of  1 Oct. 2015.

Conscious of

 My A.s.a. embodied robots learn concepts like "touch", "smell", "hunger", "pain", "bump", etc.* When one or more of these concepts becomes sufficiently activated** A.s.a. is "aware of"*** their presence. The robot is "conscious of" the sensation. More complex concepts become activated higher up in the concept hierarchy.

* See my blogs of 1 Oct. 2015 and 5 Nov. 2015 for examples.

** By sufficiently strong sensory input.

*** i.e., May react to as appropriate.

Some A.s.a. H. Societies

 I have tried various societies of A.s.a. H. agents. In a pandemonium-like configuration* all agents see the input and the one with the strongest response takes control. Another configuration has a single ("supervisory") agent receive the input and then select which (one or more) specialist agent(s) the input is sent to for response.** A third configuration is similar to Rod Brooks' subsumption architecture but the levels/behaviors and an arbiter are each replaced with an A.s.a. H. specialist agent. The behavior specialist agents all receive the input and propose output responses. The "arbiter" receives the original input and all the proposed outputs from the behavior agents and then generates an eventual output for the society. Details have varied depending on the problems/tasks being attempted.

* Due to Oliver Selfridge in 1959.

** For a large enough society with many specialists the signal can instead go into a tree of supervisory agents which route the input down to the specialist(s).

Even principles of logic are subject to change

Our human experience is very limited.* All of our concepts and models, even our logics/maths, abstract, idealize, and simplify. As our experience grows we are likely to need new logics,** new maths.

As a simple example, A.s.a. H. can be thought of as employing a kind of approximate vector logic where various sensations constitute an input vector which is processed to generate a vector output composed of actions.

* Although, of course, telescopes (and microscopes) do allow us to look away some distance as well as some way back in time.

** See, for example, Richard Epstein, Propositional Logics, Wadsworth, 2001, especially chapter XI and Paul Weingartner, ed., Alternative Logics: Do Sciences Need Them?, Springer, 2004.

There is not one single right way of doing science

 It has been said that "There are as many ways of doing science as there are scientists." That is surely an exaggeration, but there is definitely more than one way of doing science.* I have argued that being a scientist is just being intelligent (plus careful hard work).** I have also argued that intelligence is a vector quantity, one can be intelligent in various different ways.***

A given "scientific method" may also be subdivided and a scientist may specialize in just one of the resulting pieces.**** The scientific endeavor as a whole becomes a group effort. Publications may also be quite specialized/focused. I'll give an example from my plasma confinement work. We had calculated the curves describing particle and energy balance for a linear solenoid dominated by endloss. These had been verified by comparison with experiment. We then calculated (and published) the curves that resulted when other different loss mechanisms/scalings were active. These might possibly describe other confinement devices. 

Not everyone is going to be following "THE scientific method" in every one of their publications. There are different ways to do science as there are different ways in which to think.*****

* Theories of Scientific Method, Nola and Sankey, McGill-Queen's U. Press, 2007.

**See, for example, my blog of 1 September 2012.

*** See, for example, my blogs of 23 August 2010 and 8 September 2011.

**** For example: theory, computation, experiment.

***** For example: induction, deduction, abduction.

Limited reusability in A.s.a. H.

Many of my A.s.a. H. agents are specialists. They can not be reused (without modification/retraining) for some different specialty. It may be possible, however, to reuse some of the various layers (if the concepts used by the two specialists are the same and defined in the same way).   

A pantheism

With the universe as the body of god.

Evolution as the mind of god.

And evolved agents as children of god. 

Will there be yet another AI winter?

 Fully self driving cars that are always a year away.

AI "experts" that have no academic credentials.

Overpriced hardware.

Closed source software.

"Hallucinations"

Fads, hype, and fanboys.

etc., etc..

Why pluralism?

"...absolute certainty is not achievable...the depth of understanding in any field, including philosophy, often comes from exploring multiple perspectives...there are multiple ways of understanding and interpreting the world." ChatGPT 4.0

See, for example, my blogs of 26 September 2010 and 15 June 2022. 

AI curricula again

An AI should not just uncritically read from (or be trained on) the web.* Garbage in, garbage out. A mature intelligent agent would choose what to read/learn depending upon the quality of the reading material and the agent's goals and values. The early learning curriculum for an AI should be carefully chosen. Once it's matured the AI can choose for itself. 

This is one of the problems with gpt. Mainline AI research has not spent enough time/effort on curricula.

* See, for example, my blog of 6 June 2020. 

AIs learning from human publications

Humans do not share a single common world model and ontology.* The concepts that they employ differ from one another even when the same names are being used to label them. An AI model trained off human publications would, at best, learn some kind of fuzzy averaged world model and concepts.

* See, for starters, my blog of 15 June 2022. 

Realistic space launch economics

I do not agree with the optimistic estimates out of Spacex. Back of the envelope calculations, more detailed studies specifically for the Vulcan rocket*, and similar studies done by the European Space Agency are all much more pessimistic (realistic?). (But at least some of our oligarchs are putting some of their money to good use. I'll give them that much.)

* See, for example, Launch Vehicle Recovery and Reuse, M. M. Ragab, et al, American Institute of Aeronautics and Astronautics, 2015.

Adding more actuators

 We have previously noted some of the advantages to giving A.s.a. H. more sensors.* We think that there is also some advantage to its having more actuators.** Giving it two or more mobile arms, for example, rather than just one.

* See my blogs of 1 Sept. 2022 and 1 Nov. 2023.

** And actuators of various different sorts and sizes.