AI Bodies & Brains: Solving A Problem
2023 June 3, July 14, Aug 2nd, Aug 13 G.G.Ashbrook
AI usually has no body. And, if more peculiarly, AI in a sense has no brain either. This may sound odd as it is probably not uncommon to think of AI as a literal, or proverbial, ‘brain in a jar’ or at least a brain isolated from a body.
If we are going to do projects involving AI-bodies-and-brains we will need some tractable approach for either dealing with the interconnected nature of bodies and brains, or some kind of clear framework for moving ahead in some direction despite that fact our sort-of-analogy gestures to body-ness and brain-ness tend to fall away into the mists of areas of the world that H.sapiens-humans either do not understand or have largely fantastically imagined to entertain themselves: We do not know what ‘bodies’ and ‘brains’ are, or what exactly we are invoking when using the terms. But perhaps ‘close enough’ will suffice for now. Fortunately the point to be made in this paper does not assume or rely on there being some absolute biology-STEM definition for ‘body’ and ‘brain’ (or ‘life’ or ‘intelligence’ or anything else).
What we will be talking about here is probably related to ~bodies and ~brains but in a more general sense that cuts across known organism body-plans, colonies of cells or organisms, possible ET forms of life, AI, and novel hybrids involving chimeras across categories, etc., and the unfamiliarity of all that may take a moment to digest. But we will focus on a functional and practical goal relating ‘more or less’ to something body-ish, with the practical goal being very important and the semantics being disposable when they cease to be helpful.
Brain
Let’s start with ‘brain.’ In a way, the generative intelligence of gpt-LLM deep learning model artificial neural networks is a containerized sort of hologram of millenia of H.sapien-human intelligence (by whatever name) gleaned from printed literature.
(see: Dumbledoor’s Portrait paper)
Though we may never understand ‘intelligence,’ [or “intelligence” may turn out to be a malapropism of meanings as was “phlogiston” (something which we now know does not exist)] a great approach is S.Raniwala’s hypothesis that ‘intelligence’ (as exhibited by GPT-LLM and by H.Sapiens-humans) exists not so much in a homunculus hiding somewhere inside a brain, but in the Firth-relationships between concepts in externalized literature: that something of, or all of, ‘human-intelligence’ exists ‘in’ language (being a set of connection between parts of language) itself flowing, transmitting, between (or in a network of) a substrate of “people-body-brains.” (a Firth-humuculus?) This may seem to be splitting hairs or getting stuck on arcane distinctions, but as you will see below there are very important function, system, design, details for what we mean operationally by ‘intelligence.’ What exactly can and can’t this ‘intelligence’ do? Or in this case, what is the general relationship between ‘intelligence’ and a ‘body-brain’?
As we will see, we may need to unpack what we mean by ‘intelligence’ in isolation vs. ‘a system that can employ intelligence to guide operations.’
Strengths & Weaknesses
We are in a sense still back in something like the same pre-2020 time when AI had no significant object-handling ability. (See full paper: https://github.com/lineality/object_relationship_spaces_ai_ml )
“AI is good at identifying things, but not very good at doing anything with that information.”
While this phrase originally came from describing the difference between a very narrow not-at-all-general AI and the kind of tasks we want AI to do, this still applies surprisingly well when we push ahead with seeing what even object-handling (see paper for explanation of term) gpt-LLM-AI can do in a context involving the following:
- projects
- participants
- interconnecting areas
- generalized STEM
etc. including self-checking, explaining what you mean, following process standards and protocols, and thorough reporting.
(See full paper: https://github.com/lineality/object_relationship_spaces_ai_ml )
Back to square one?
The fact that an often powerful ‘intelligence without body or brain’ can nevertheless often fail to manage math, analytical, system 2
https://en.wikipedia.org/wiki/Thinking,_Fast_and_Slow tasks
may be related to our body-brain problem at hand. Perhaps like the genie scenario where you ask for something ‘intelligent’ and then realize what you got was not what you thought you were asking for, to rehash the immortal story again: back in the 1970’s Douglas Hofdstedter (in GEB) prepared us for the likelihood that even powerful human-level AI would (and perhaps should) have difficulty with doing math, analytical, system 2 type processes quickly (note: while I am not aware of Hofstedter was working with Kahnaman and Tversky, they both emphasize a similar problem space and in particular fast vs. slow (it was literally the title of Kahnaman’s summary book). Kahnaman and Tversky have helped us to better understand how H.sapiens-human brain-bodies work and quite often fail. And in general most highly intelligent STEM professionals are oblivious to proper project management (perhaps an echo of the pre C.P.Snow pre Sir Eric Ashby years when serious people could seriously assert that a scientist, an engineer, and a mathematician, would never have any business speaking with one another [for a brilliant portrait of this utter tragedy, read Andrew Hodges’ “Alan Turing : The Enigma” 1983, for the backstory of how post-war US and UK fumbled for years to build the first generation of computers because they could not overcome this exact social barrier and “leap of understanding”]. Indeed the problem was much worse, with ‘science’ having no single conception and the various factions comically (in retrospect) fanatical about not recognizing the others (see Ashby, back in print thank goodness). Note, even core events later in computer science and AI are enmeshed with this problem, where none other than John McCarthy (creator or co-creator of the term and current form of AI itself) tried to stop the implementation of none other than the LISP language (his own invention!) from being implemented in reality, not because it was dangerous, but because he saw no conceivable connection between philosophy-thought space (STEM) and reality-engineer-maker-space (also STEM). https://en.wikipedia.org/wiki/Lisp_(programming_language)
History is full of this kind of paradigmatic disparity that makes it so difficult for a given person (or group) at a given time to intellectually-empathize with how the world looks in both the past and future over gulfs of time.
Problem & Solutions
Solving a problem, or doing a project, such as a child’s math-word-problem, is not accomplished by a fantasy of ‘symbol manipulation’ or an invocation of some wooly essence like ‘intelligent-ness’ alone, however much our please-sensors light up at hearing and retelling such ‘just so’ mythic tales.
Our challenge is to help the kabbalah-shard-fragment of an Indra’s-web of notions that is the GPT-LLM containers of human intelligence to be able to do ‘projects’ like 1+1=2 reliably, by building a sort of ‘brain-body’ for that AI: an architecture, an OS. a programming language: a debugger, a compiler, a linker. Pick whatever mascots or analogies you like, but we need some framework that can get the job done.
Across disciplines (with a nice recursion around the teaching of language and STEM using language and STEM to teach AI to help teacher’s of language and STEM using Language and STEM…so that children can learn Language and STEM…to help etc.) from basic teaching best-practice methodology to Francois Chollet’s (he is the creator of Keras the primary AI framework) exhortations of breaking things down into modular recombinant parts that can be assembled and used and re-assmbled and re-used, to how many computer languages from C to Rust work, there is a value in breaking down a project-task into parts, solving those, then linking the whole solution back together. And, again see the full paper, when this is a project that involves multiple participants and other project aspects there are many more important parts to this breakdown-build-up pattern that do not seem readily apparent when it is considered a mere flourish to adorn a single person solving a hypothetical problem in their mind (or to imagine and believe they have solved it…).
Probably not for the last time, the system by which we might furnish and AI’s echos of literary intelligence with a ‘brain’ are strikingly similar to how we teach…or try to teach…children how to think, write, and solve problems in an organized effective way (with ‘externalization’ being a stubborn theme). And interestingly even just the disembodied AI-echo that is GPT-LLM shares many similarities with the strength and weaknesses of corporeal H.sapiens-humans. See:
https://medium.com/@GeoffreyGordonAshbrook/ai-counting-problems-8cb9f66e4c7f
And there may be more to this than merely saying the former is an echo of the mind of the latter: there may be a more general set of process dynamics going on here, beyond contagion of idiosyncrasies.
Does this sound familiar?
- Use a structured framework.
- Make a plan for what you are doing.
- Understand the focus and scope of your task.
- Follow your plan.
- Show your work!
- Do not skip steps!
- Write everything out clearly!
- Explain what you mean.
- Double check your answer!
- Proofread your work!
- Revise your work!
- Publish, share, or submit your work.
For years this has been the nagging guidance of H.sapiens-human adults to H.sapiens-human children, and now it is the nagging guidance of H.sapiens-humans (of all ages) to an AI (which is a fragment of their own minds from ancestral texts).
But how is AI, having no body and no brain, going to do this?
We will build something with which the AI can do this, and this something is, or may be akin to, a brain in which this AI, which is so good at making observations, can then actually do something with that information.
Intuition vs. Code … “Show your work…”
Another important strength-weakness node in this approach is the observation that while gpt-LLM models are often very slapdash about estimating an overall answer, they can be extremely good at writing computer code to do the same exact thing properly. On the one hand this may seem strange, or seem like a deficiency, but this is remarkably similar to how H.sapiens-humans behave and it may be part of systems in general, (may be) for example if the ‘system 1’(non-analytical or intuitive-guesstimating) and ‘system 2’(analytical) of Khanman and Tversky are a more general part of the world and not just an accident of evolutionary history for one branch of mysteriously hairless ape.
Another possible context for this distinction (and another excuse to possibly merge chess mysteries into an AI narrative) is strategy vs. tactics. Or is that a different subject altogether? All in all we are still rather new to this whole space of general-STEM, game-theory, problem solving, nature-of-language, nature-of-thought, so hopefully a few hundred years from now people will have a better idea of what’s going on and they can safely look back at our confusion with smug disdain. Bless them.
Part 2 : Solving A Problem
As a concrete example, let’s look at the process of solving a multi-step problem, or solving a problem with an externalized project-share-able and scrutinize-able, error-check-able, process. This description could still refer to an unhelpfully large number of problems and kinds of problems (or a very large problem-space), so let’s start with some concrete first samples (however they are pigeonholed):
A math problem:
A math word-problem: A conventional k-12 school question.
A logical-puzzle problem: Who is coming to the party?
Note: (See a later paper on modularizing problem space with an example ‘counting problem’ in this same overall context.)
Part 3 : Math, Language, STEM, Computer Science
Riddle Me This
A quote by Francois Chollet:
Page 26, Chapter 2, opening paragraph,
“The most precise, unambiguous description of a mathematical operation is its executable code.” — F. Chollet
https://www.manning.com/books/deep-learning-with-python-second-edition
This perhaps simple seeming statement is somewhere between nuanced, subtle, bold, and profound, as both the long history of debates over whether mathematics and logic has any relationship to STEM and practical problem solving is tumultuous and largely unresolved, and computer science (discrete or nonlinear) is rarely mentioned as part of STEM at all: so while the time may have come to postulate an authoritative seat for computer science to best define a very real ability for math-logic to solve practical STEM problems…most of the rest of the world is still bickering over preclusive component objections.
https://www.amazon.com/Technology-Academics-Universities-Scientific-Revolution/dp/1015312659/
A truly fascinating lineage of books in and about the history of thinking about STEM and science is the progression from Sir Eric Ashby’s “Technology and the Academics” to C.P. Snow’s “The Two Cultures,” to Stephen Pinker’s “Enlightenment Now.” Though each of which pays homage to the previous and states their goal to re-write, update, and build upon the past, there is a ‘arc of history’ of sorts from the first two books that trace out a bewildering diversity of disagreements over whether there is or could be a generalized or unified STEM and how that relates to the world, with the perhaps a-touch-too-combatively glib Pinker’s work making his cut-to-the-chase case that everything is reason and anyone who disagrees is a dangerous moron.
“someone else’s mind”
Along these lines, unassuming books such as “A Programmer’s Introduction to Mathematics: Second Edition” by Jeremy Kun (Author) https://www.amazon.com/dp/B088N68LTJ
may contain important practical insights for how ‘tricks of the trade’ for both math-logic and programming can come together for systematic problem solving.
From chapter 2, page 12:
“This theme, that most mathematics is designed for human-to-human communication, will follow us throughout this book. Mathematical discourse is about getting a vivid and rigorous idea from your mind into someone else’s mind.”
Whether or not these aspirational statements by Chollet and Kun are true-enough remains to be seen, but the game is afoot as we set out to transfer ideas not only between ‘human’ minds (whatever that means), but between any mind or intelligence in general, and even using decidedly non-mind tools: (perhaps similar to the Raniwala Hypothesis that conceptual-intelligence exists in language space (in the measurable Firth-space between n-grams) not in homunculi hiding in brains,) ‘mathematics’ (which Skip Ellis believed was a natural language (another debate that I cannot resolve)) is contextualized as being able to contain and convey ideas (and/or that computer-code is contextualized as being able to convey mathematical ideas between humans and machines) and in language-space. (To shorten that: The claim that ‘mathematics’ is able to contain and convey ideas in language-space.) This is testable and falsifiable, and we shall see what works and what fails.
Examples and Possibilities:
Again, in practicality, there will probably be a slowly or quickly growing set of types of problems that can be systematically broken down and solution-module-compiled in a general STEM etc. approach, with the open-ended “any problem” being too open-ended to resolve.
As with any area in computing, ‘programming language’ ‘compiler’ ‘operating system’ ‘network architecture’ ‘processor architecture’ ‘gpu’ ‘tpu’ ‘memory storage’ are categories with perhaps infinite possible solutions and variations (either with pluses and minuses) usually for specific implementations and uses, these not individual specific things of which there is only one example.
Part 4 : Context
On the one hand it may sound risky to ‘give an AI the ability to do anything!’ but that is perhaps an overly vague, hyperbolic and inaccurate way of describing this small step into a larger understanding of the nature of STEM, projects, and our own intelligence in a jar: this has not, per se, solved or even addressed questions like ‘volition’ and ‘independent action’ and ‘appetition (or desire)’ etc.
This proposed architecture is an additional layer to the same conventional ‘chat-bot’ interface so that some narrowly-defined types of questions asked of the chat-bot can involve an incrementally-less superficial ability to interface with problems, and projects, and object-relationships. But, back to the title, there still is no mature body or brain. And, if frustratingly, this is not a ‘do everything generally’ theory-of-everything.
It might be interesting to experiment and say to such a semi-embodied-AI, should it be possible to build and craft one using such a mini-brain architecture: “Make a list of things you would like to do today, do the item at the top of your list, and report back! How did it go?” And this may pre-suppose that the AI can ‘do’ things in ‘the world.’ The version of gpt4 I have now still says it cannot access the internet in any way and its knowledge is limited data from two years ago. Politics aside, adding in more abilities for the ai-fragment use, for example, to look up information on wikipedia, will eventually be created (and permitted).
But even this example, which seems like the AI “deciding to do something on it’s own volition” would still be a single process, run in isolation, that would terminate at the end of the instruction, and the AI would return to its amnesiac state, passively inert until something bumps into it that it reacts to; still an image of intelligence (however high resolution) that can only reflect and refract back from a prompt shown into it: There is no body.
The Mirror Analogy: “Mirror Mirror On the Wall…”
We might look at GPT-LLM AI as a kind of intelligent mirror; it can reflect intelligence; it can reflect articulately. But that one, passive, reactive, involuntary, reaction is all this fragment can do. For example, and hopefully this will be done some day,
you can imagine an intelligent mirror that has been trained with all of the works and words (and whatever other data) of your wise grandmother, or some benevolent great thinker and teacher from the past like Niels Bohr. You could hang that mirror on your wall, and give it things to ‘reflect on.’ And, if well-made, it could reflect with great wisdom, subtlety, nuance, insight, depth of experience, and articulation. But it would not be a person, an organism, an ecosystem, it would be a fragmentary reflection no matter how profoundly articulate.
For example a mirror that can only reflect cannot solve a multi-step problem. You may visualize this as a contraption with laser light sources and mirrors and lenses. This ai-mirror may be able to give (reflect) an all-in-one guesstimation as to what the outcome might be, but that would be (as Hofestedter predicted) a lagging ability (a fast-fuzzy-guess at what the outcome of a longer slower process probably might be), or an an internal guesstimation ability not to be confused with the outcome of an externalized process. The mirror can reflect on each step, but it cannot alone do anything with the steps for a problem. It cannot feed a step to itself. It does not know what step (in or out of sequence) it will reflect on next. It can be fed a jumble of randomly ordered instructions taken from unrelated projects and forever patiently reflect brilliantly on each one in term, with no memory or concern about what came before or what comes next (perhaps a Hume-moment-sequence nightmare…Hume’s Daemon if he does not have one named after him yet). In some ways this is like a CPU or ALU in a Turing/Von-neuman type computer architecture, processing (or reflecting on) one operation at a time, passively (and not fussed about whether the concurrency and parallelism and swapping of instruction sets is being managed properly). It cannot refuse to react. It cannot ‘react’ unless there is something to react to (or reflect upon). A CPU is not a computer. An intelligent mirror is not a person. A brilliant gpt-ai fragment is not a brain.
A computer, on the other hand, with an architecture and operating system of some kind, and many peripheral components, is more of an ‘organism’ that can knit and link steps together. A CPU or ALU cannot compile a long program into a finished executable binary in a single boolean operation (with a running theme here so far unmentioned of how much information can or should be squeezed into a single operation…a single reflection, a single ‘prompt’), but a full computer body-brain can link and compile the parts of a program. A mere reflection of a person could not participate in a project, with bits and parts of schedules and instructions being crafted and passed back and forth, multi-step tasks accomplished, plans followed and reported on, plans changing; but a “person,” or participant, with a body-brain can do these things. (Or can, in an ecosystem of other participants?)
Using this analogy to frame our problem-problem, how can we build a ‘brain’ around this intelligent mirror so that the intelligence that is only passively contained can be channeled and embodied into an architecture that can, if only one small step, solve a problem. The wise-mirror can do each step, and knows how to link the steps, but it needs a body to do so.
There is also perhaps an interesting twist or fold, where on the one hand we think of AI needing extra furniture to be more like ‘the solitary and sovereign human brain,’ but H.sapiens-humans are exemplars of the gravity and tendency to not show their work and to jump to opaque emotive reactions, making suspiciously similar mistakes as compared with AI, which on the one hand is not surprising because the AI is literally a reflection of human behavior, put perhaps a bit more surprisingly this is an architecture question not an intelligence question, perhaps suggesting that H.sapiens-humans as well have an intelligence-vs-architectures set of problems in themselves, which in that light may make sense of the fraught relationship between H.sapiens-humans and their own bodies and minds, societies, projects, etc.
“The Spitting Image”
Probably most people have often heard people talk about a child in the family as the ‘mirror image’ of some, often long departed, relative. And indeed, on a large enough display of family photos without dates on the photos, it can be impossible to tell whether a picture is of one person or of their grandmother at the same age. Homo-sapiens humans are in some ways mirrors of past iterations of their amnesiac-phylogenetic-body, and in a way that is curiously complimentary to the AI mirrors of our ancestors. Biological reflections have volition and bodies and minds and interact with the world, but they have (at least not superficially) none of the memories, experiences, and ideas, of their ancestors. The AI-mirror on the other hand has only those things: memories, knowledge, wisdom, perspectives, etc. Is this a match made in…heaven?
Part 5: Process
An interesting if oddly invisible part of the history of computer science is the ‘invention’ of the idea that computation can be done by a process, in particular an externalized process.
And in assuming that a generative AI is ‘internally using an externalized process’ while also not addressing generalized STEM or how people solve problems is a fascinating mixed-up half-amnesia where we both assume that problems are solved by analytical steps, forgetting that this was ever not-obvious, but also forget that this distinction exists assuming that generative internal type 1 imagination is type 2 systematic process. What could be a more ‘human’ assertion?
This may even go back, in a way, at least to the not very clear ‘origin’ of algebra as a formal process of defining variables, which only very slowly became formalized and adopted (perhaps too slowly for anyone to really notice that it happened) between Al-Khwārizmī and Girolamo Cardano. (see “Significant Figures” by Ian Stewart)
https://www.amazon.com/Significant-Figures-Lives-Great-Mathematicians/dp/0465096123
Also see:
- https://en.m.wikipedia.org/wiki/Prompt_engineering
About The Series
This mini-article is part of a series to support clear discussions about Artificial Intelligence (AI-ML). A more in-depth discussion and framework proposal is available in this github repo:
https://github.com/lineality/object_relationship_spaces_ai_ml
