The Society of Mind (1986) [pdf] | Hacker News Reader

The Society of Mind (1986) [pdf]

acad.bg

159 pointseigenvalue69 comments

I was a grad student in AI at the time this book came out so I can tell you a little bit about the historical context from my personal perspective with benefit of hindsight. The field at the time was dominated by two schools of thought, called the "neats" and the "scruffies". At the risk of significant oversimplification, the neats thought that the right way to do AI was using formal logic while the scruffies took an empirical approach: noodle around with code and see what works. Both approaches led to interesting results. The neat legacy is modern-day theorem provers while the scruffy legacy is chatbots, self-driving cars, and neural nets.

SoM didn't fit neatly (no pun intended) into either camp. It wasn't empirical and it wasn't formal. It was just a collection of random loosely-associated ideas and nothing ever came of it. It was too informal to lead to interesting theoretical results, and it was too vague to be implemented and so no one could test it experimentally. And both of those things are still true today. I think it's fair to say that if the author had been anyone but Minsky no one would have paid any attention to it at all.

Well it did spur the research into multi-agent systems (popular study area in 1990s) and interaction protocols. So there was a degree of influence.

But taken on its own it is indeed more a book of musings. I put GEB, ANKOS and the like into the same genre.

I'm with you on ANKOS, but GEB is an accessible and fun (if a bit wordy) introduction to formal systems and Godel's theorem, so I wouldn't put it in the same category. GEB also was not marketed as anything revolutionary (except in its pedagogy). ANKOS and SoM were.

Inspired me as an undergrad Industrial Design student in 1989ish that and The Media Lab: Inventing the Future at M.I.T by Stewart Brand were the two most influential technology books for me at that time.

Coincidentally enough it turns out my cousin was in the thick of it while the pre-media lab was still part of the architecture school. She would tell me stories of what she was up to in college... when I read that back I had to loop back and ask her about it.

This corroborates my experience.

Reading Society of Mind in undergrad is one of the things that led me to doubt AI progress and to stray away from the field [1]. It was handwavy, conceptual, and far removed from the research and progess at the time. If you held it up to Norvig's undergraduate level Artificial Intelligence: A Modern Approach, you could sense Minsky's book was as wishfully hypothetical as Kaku's pop-sci books on string theory.

[1] Recent progress has led me right back. There's no more exciting place to be right now than AI.

Even if it didn’t lead to empirical results I think most of the value of the book today is in the questions Minsky asked. How is intelligence organized in a distributed system like a neural net? ChatGPT may be able to do amazing things, but the mechanisms it uses are still very opaque. So even if the theory may not be “useful”, it is still worth pursuing IMO

It’s also pretty well written and written by someone who clearly spent a lot of mental energy on the problem

https://web.media.mit.edu/~minsky/papers/SymbolicVs.Connecti...

      Logical vs. Analogical
              or
     Symbolic vs. Connectionist
              or
         Neat vs. Scruffy

         Marvin Minsky

INTRODUCTION BY PATRICK WINSTON

Engineering and scientific education conditions us to expect everything, including intelligence, to have a simple, compact explanation. Accordingly, when people new to AI ask "What's AI all about," they seem to expect an answer that defines AI in terms of a few basic mathematical laws.

Today, some researchers who seek a simple, compact explanation hope that systems modeled on neural nets or some other connectionist idea will quickly overtake more traditional systems based on symbol manipulation. Others believe that symbol manipulation, with a history that goes back millennia, remains the only viable approach.

Minsky subscribes to neither of these extremist views. Instead, he argues that Artificial Intelligence must employ many approaches. Artificial Intelligence is not like circuit theory and electromagnetism. AI has nothing so wonderfully unifying like Kirchhoff's laws are to circuit theory or Maxwell's equations are to electromagnetism. Instead of looking for a "Right Way," Minsky believes that the time has come to build systems out of diverse components, some connectionist and some symbolic, each with its own diverse justification.

Minsky, whose seminal contributions in Artificial Intelligence are established worldwide, is one of the 1990 recipients of the prestigious Japan Prize---a prize recognizing original and outstanding achievements in science and technology.

https://en.wikipedia.org/wiki/Neats_and_scruffies

Neat and scruffy are two contrasting approaches to artificial intelligence (AI) research. The distinction was made in the 70s and was a subject of discussion until the middle 80s. In the 1990s and 21st century AI research adopted "neat" approaches almost exclusively and these have proven to be the most successful.[1][2]

"Neats" use algorithms based on formal paradigms such as logic, mathematical optimization or neural networks. Neat researchers and analysts have expressed the hope that a single formal paradigm can be extended and improved to achieve general intelligence and superintelligence.

"Scruffies" use any number of different algorithms and methods to achieve intelligent behavior. Scruffy programs may require large amounts of hand coding or knowledge engineering. Scruffies have argued that the general intelligence can only be implemented by solving a large number of essentially unrelated problems, and that there is no magic bullet that will allow programs to develop general intelligence autonomously.

The neat approach is similar to physics, in that it uses simple mathematical models as its foundation. The scruffy approach is more like biology, where much of the work involves studying and categorizing diverse phenomena.[a]

https://www.amazon.com/Made-Up-Minds-Constructivist-Artifici...

Made-Up Minds: A Constructivist Approach to Artificial Intelligence (Artificial Intelligence Series) Paperback – January 1, 2003

Made-Up Minds addresses fundamental questions of learning and concept invention by means of an innovative computer program that is based on the cognitive-developmental theory of psychologist Jean Piaget. Drescher uses Piaget's theory as a source of inspiration for the design of an artificial cognitive system called the schema mechanism, and then uses the system to elaborate and test Piaget's theory. The approach is original enough that readers need not have extensive knowledge of artificial intelligence, and a chapter summarizing Piaget assists readers who lack a background in developmental psychology. The schema mechanism learns from its experiences, expressing discoveries in its existing representational vocabulary, and extending that vocabulary with new concepts. A novel empirical learning technique, marginal attribution, can find results of an action that are obscure because each occurs rarely in general, although reliably under certain conditions. Drescher shows that several early milestones in the Piagetian infant's invention of the concept of persistent object can be replicated by the schema mechanism.

https://dl.acm.org/doi/10.1145/130700.1063243

Book review: Made-Up Minds: A Constructivist Approach to Artificial Intelligence By Gary Drescher (MIT Press, 1991)

> these have proven to be the most successful

Hindsight and all that...

Another thing to be aware of with SoM is that Minsky was reading in many fields, and trying to sketch out theories informed by that.

One time, before the DL explosion, during a lull in AI, I sent a colleague a Minsky pop-sci quote from the earlier AI years, before our time, asserting that, soon, a self-teaching machine will be able to increase in power exponentially. I was making a joke about how that was more than a little over-optimistic. My colleague responded something like, "What you fail to see is that modern-day Marvin is that machine."

By the time I was bumping into AI at Brown and MIT, the students (including Minsky's protege, Push Singh, who started tackling commonsense reasoning) described SoM various ways, including:

* Minsky sketching out spaces for investigation, where each page was at least one PhD thesis someone could tackle. I see some comments here about the book seeming light and hand-wavy, but I suppose it's possible there's more thinking behind what is there than is obvious, and that it wasn't intended to be the definitive answer, but progress on a framework, and very accessible.

* Suggestion (maybe half-serious) that the different theories of human mind or AI/robotics reflect how the particular brilliant person behind the theory thinks. I recall the person said it as "I can totally believe that Marvin is a society of mind, ___ thinks by ___ ..."

I don't know anyone who held it out as a bible, but at the time it seemed probably everyone in AI would do well to be aware of the history of thinking, and the current thinking of people who founded the field and who have spent many decades at the center of the action of a lot of people's work.

At the time it seemed to me that each page was, as you say, a thesis someone could tackle.

It never seemed to be an instruction manual or was presented as some kind of total blueprint of thinking Cough Wolfram

This made its way into pop culture via the X-Files, in an episode about A.I.: "Scruffy minds like me like puzzles. We enjoy walking down unpredictable avenues of thought, turning new corners but as a general rule, scruffy minds don't commit murder."

> It was just a collection of random loosely-associated ideas and nothing ever came of it.

I remember buying this in '89 and being completely underwhelmed by it. There is nothing there imo. I stopped paying attention to the name Minsky after this introduction to the 'great man'.

Thanks for this. I was going to post a comment asking how relevant SoM is to the form and structure of modern ML models. From just the title, not having read it, it seemed like SoM might have been prescient. Apparently not so much.

Funnily enough, I'm currently trying to make my way through a preprint showing that models of dense associate memory with bipartite structure, including Transformers (!!!), are a special case of a more general routing algorithm that implements a "block of agents" in a differentiable model of Minsky's Society of Mind: https://arxiv.org/abs/2211.11754. Maybe "symbolic" and "connectionist" AI are two sides of the same coin?

EDIT: I feel compelled to mention that the efficient implementation of that more general routing algorithm can handle input sequences with more than 1M token embeddings in a single GPU, which quite frankly seems like it should be impossible but somehow it works: https://github.com/glassroom/heinsen_routing#routing-very-lo....

Thank you. If you gain any insights while reading, please share!

https://news.ycombinator.com/item?id=34127486

Wow, this sounds exactly like what I was talking about. Thanks for the reference.

:-)

For those reading this after-the-fact, this thread started as a response to eigenvalue's original headline, a question, which was subsequently moved to a separate comment: https://news.ycombinator.com/item?id=34110868

How does his routing algorithm compare to attention? I saw this question in the repo faq, but no satisfactory answer is given.

I think the next-to-last faq ("Is it true that EfficientVectorRouting implements a model of associative memory?") answers that. Did you see it?

Oh I see, thanks! Interesting. It sounds like this is some kind of dynamic attention, as opposed to static attention in transformers, where queries and key don’t change during the calculation of their similarity. His routing algorithm computes the similarity iteratively.

Is this your assessment as well?

I'm still trying to make my way through the preprint :-)

EDIT: According to https://ml-jku.github.io/hopfield-layers/#update , attention is the update rule for an (iterative) "dense associate memory," even though in practice it seems that one update works really, really well for attention if you train it with SGD.

OK, here's a preliminary, hand-wavy, high-level summary of my understanding so far:

It seems we could repeatedly apply attention as an iterative update rule, feeding the output sequence as the new queries: new_Q = Attention(old_Q, K, V). Obviously, no one ever does that. But in theory we could, and apparently it would work just as well.

In self-attention, we apply dense layers to the input sequence to get K and V, so we can rewrite the update rule like this: new_Q = Attention(old_Q, input_seq). Internally, Attention would have to compute and cache K and V from input_seq. The input sequence is given and remains fixed, so we can rewrite the update rule again to make that explicit, like this: new_Q = Attention(old_Q | input_seq).

According to the preprint, everything the routing algorithm does can be written as an update rule U that looks just like that, except that the preprint calls the queries "x_out" and the input sequence "x_inp". Using their notation, the update rule looks like this (section 3.3 and appendix B): x_out <-- U(x_out | x_inp).

One more thing that seems important: In self-attention, we get the initial queries by applying a dense layer to the input sequence. In the routing algorithm, the initial queries come instead from the same update rule, U, by assuming all possible output states have equal probability. Interestingly, the code in the github repo executes only two updates by default: one to compute the initial queries, and another one to get updated queries, which are returned as the output sequence. This isn't too different from what we normally do with attention: first we compute the initial queries, and then we apply one update to get updated queries, which are returned as the output sequence.

EDITS: Many.

MORE EDITS: There are quite a few other differences vs attention that look significant. Preliminary and hand-wavy: Like other capsule-routing networks, this one gives us credit assignments that tell us how much each capsule (token embedding) in the output sequence depends on each capsule (token embedding)in the input sequence. Unlike other capsule-routing networks, this one seems to scale up really well, even though V has n x m x d elements, vs. n x d in attention.

Minsky wrote this book in 1986, towards the end of his very long career thinking about how to build intelligent machines. For a basic overview, see:

https://en.wikipedia.org/wiki/Society_of_Mind

You can find a complete pdf of the book here:

http://www.acad.bg/ebook/ml/Society%20of%20Mind.pdf

My question to the HN community is, has all this work become irrelevant given recent progress in machine learning, particularly with Transformer based models such as GPT-3 or "mixed modality" models such as Gato?

It seems to me that some of these ideas could make a comeback in the context of a group of interacting models/agents that can pass each other messages. You could have a kind of "top level" master model that responds to a request from a human (e.g., "I just spilled soda on my desk, please help me") and then figures out a reasonable course of action. Then the master model issues requests to various "specialist models" that are trained on particular kinds of tasks, such as an image based model for exploring an area to look for a sponge, or a feedback control model that is trained to grasp the sponge, etc. Or in a more relevant scenario to how this tech is being widely used today, a GitHub Copilot type agent might have an embedded REPL and then could recruit an "expert debugging" agent which is particularly good at figuring out what caused an error and how to modify the code to avoid the error and fix the bug.

I suppose the alternative is that we skip this altogether and just train a single enormous Transformer model that does all of this stuff internally, so that it's all hidden from the user, and everything is learned at the same time during end-to-end training.

This post is fucking catnip for me. I still believe society of mind is due for a huge resurgence, due to exactly what you're saying. Bottom-up composable skills will be a huge step forward, and free up GPT etc to be creative while getting the low-level stuff 100% correct (instead of screwing up arithmetic half the time)

The inverse side of the coin is emotions, meaningful relationships, and wisdom, which I think work in similar way but more diffuse. There can be a horny submodule that analyzes incoming content for sexual context, one for anger, fear, gratitude, etc. The same way an image processor convolves over pixels looking for edges and features, an emotional processor will operate over data looking for changes in relationships.

Feelings act like filters on incoming data, and are composed out of base physiological reactions. Like anger involves adrenaline, which increases pain tolerance and dampens conscious thought in favor of subconscious and instant reactions.

This makes so much intuitive sense to me. I love society of mind, I wonder if it bears a relationship to how the human mind works

Absolutely, see Internal Family Systems. With a little practice, you can empirically determine that this is how your mind and emotions work.

https://ifs-institute.com/

Edit: society of mind is a great launching point for AI alignment as well. If there are internal processes that "care" about/sense different things, how do they work together?

1. Bidding system. Subsystems parse incoming info, and pay to bubble their need/plan into consciousness/action

2. Nested levels of concern, ex. maslow's hierarchy

3. "Recipes" that combine output from multiple subsystems. Learning recipes via subconscious communication / mirroring.

I get the sense AI alignment is too focused on the fantastical end-of-the-world catastrophe scenario. It's impossible to get anything actually done vs a vision of apocalypse. The real strides are going to come from operational safety and "open enough" reasoning, not "how do I know that its not lying about lying about..." paranoid recursing.

Sounds similar to what Google's SayCan is doing. https://say-can.github.io/

They taught it separate skills. When a situation arises, the skills (which you could almost consider sub-agents) compete to decide who is most likely to be relevant here. Then that skill takes over for a bit.

They also have a version called "Inner Monologue" in which the different parts "talk to each other" in the sense of collaboratively creating a single inner monologue, allowing for reactiveness/closed loop behaviour.

I interviewed 2 authors of SayCan/Inner Monologue here: https://podcasts.apple.com/us/podcast/karol-hausman-and-fei-...

The Society of Mind (2011) - https://news.ycombinator.com/item?id=30586391 - March 2022 (37 comments)

The Society of Mind - https://news.ycombinator.com/item?id=12050936 - July 2016 (2 comments)

Marvin Minsky's Society of Mind Lectures - https://news.ycombinator.com/item?id=10971310 - Jan 2016 (6 comments)

The Society of Mind (1988) - https://news.ycombinator.com/item?id=8877144 - Jan 2015 (6 comments)

The Society of Mind Video Lectures - https://news.ycombinator.com/item?id=8668750 - Nov 2014 (10 comments)

Marvin Minsky's "The Society of Mind" now CC licensed - https://news.ycombinator.com/item?id=6846505 - Dec 2013 (2 comments)

MIT OCW:The Society of Mind (Graduate Course by Minsky) - https://news.ycombinator.com/item?id=856714 - Oct 2009 (2 comments)

In a way, any GAN (https://en.wikipedia.org/wiki/Generative_adversarial_network) has aspects of a Society of Mind: two different networks communicating with each other, with the discriminator attempting to find flaws with the generator's ongoing output.

And https://scholar.google.com/scholar?hl=en&as_sdt=0%2C31&as_vi... shows many attempts to generalize this to multiple adversarial agents specializing in different types of critique.

One of the challenges, I think, is that while some of these agents could interact with the world, it's just far more rapid for training if they just use their own (imperfect) models of the relevant subset of the world to give answers instantaneously. Bridging this to increasingly dynamic physical environments and arbitrary tasks is a fascinating topic for research.

> Marvin Minsky's "Society of Mind" is a theoretical framework for understanding the nature of intelligence and how it arises from the interaction of simpler processes. The concept suggests that the mind is not a single entity, but rather a society of many different agents or processes that work together to produce intelligent behavior.

> The concept of a "Society of Mind" is still relevant today and has influenced a number of fields, including artificial intelligence, cognitive psychology, and philosophy. It has also influenced the development of artificial intelligence systems that are designed to mimic the way the human mind works, using techniques such as artificial neural networks and machine learning.

> Overall, the concept of a "Society of Mind" continues to be an important and influential idea in the study of intelligence and the nature of the mind.

Or so at least, says ChatGPT when I asked it about this just now.

How long before "ChatGPT" is just "Chat", as in "I just asked Chat and it said ..."?

I hope never. But we really like to make things harder by giving the same name to many things.

Won't they just give it a name? i.e Siri, Alexa, Jeeves

Just noticed that GPT sounds a little like Jeeves.

Or gipped (polite spelling), as in you've been had.

There is still plenty of research going on on agents, symbolic AI, and other approaches that sometimes and somewhat reflect (or have informed) ideas from Society of Mind. Some of the ideas are relevant from an application perspective (for sure, we have complex socio-technical systems where different 'agents' interact), others make it into learning, for example into RL, which was hyped some years go. Other ideas feel old-fashioned and stuck in the past; this is, in my opinion, not necessarily because the ideas are generally bad, but often because some of the sub-communities move very slowly and struggle to embrace a pragmatic approach to modern applied research.

Generally, I think it's good to maintain 'old' knowledge, and the only way to do so in a sustainable manner is to maintain a diversity of research directions, where plenty of researchers are committed to keep the lights on by slowly advancing directions that are not on top of the hype cycle at the moment.

Also look at his later book, "The Emotion Machine".

When I took Minsky's "Society of Mind" class, he was working on the later book, and many lectures were him talking about what he had been working on earlier that day.

I've read this book a couple of times, with my most recent re-read being within the last year or two. So I guess that means that I, for one at least, find something of value in SofM even now.

So the question then might be "what do you find valuable in it?"

That would take a lot of words to answer fully, but let me start by saying that I agree with a lot of the other comments in on this post. The theory, inasmuch as you can call it that, isn't super concrete, isn't necessarily something you can implement directly as such, does mostly lack any kind of experimental evidence, and is kind of hand wavy. Sooooo... what value does it have?

Well for me it's mostly something I look at as inspirational from a very high-level, abstract point of view. It strikes me as more of a framework that could support very many theories, rather than a specific realizable theory. But I believe that there's something fundamentally correct (or at least useful) about the idea of a collections of semi-autonomous agents collaborating in a style akin to SofM. And on top of that, I think there are at least a handful of specific notions contained in the book that might be realizable and might prove useful. If you want a specific example, I'd say that I think something like K-lines may prove useful.

Of course I have no experimental evidence, or much of anything else beyond intuition, to support my beliefs in this. And I'm just one random guy who's pretty much a nobody in the AI field. I just sit quietly at my computer and work, not really trying to attract a lot of attention. And in the process of doing so, I do occasionally consult Society of Mind. YMMV.

And just to be clear in case anybody wants to misinterpret what I'm saying. It's not my "bible", and I'm not a Minsky acolyte, and I don't consider SofM to be the "be all end all" any more than I consider A New Kind of Science, Godel Escher, Bach, Hands on Machine Learning with Scikit-Learn, Keras & Tensorflow, Computational Approaches to Analogical Reasoning: Current Trends, or Parallel Distributed Processing, Vol. 1: Foundations to be the "be all, end all". I'm all about applying Bruce Lee's mantra:

"Use only that which works, and take it from any place you can find it."

I haven't heard anyone talk about this book in a long time. I read it while a grad student at the MIT Media Lab in the 90s (albeit not his course). I struggled to understand its relevance even then. I think for many people that book was an introduction to ideas about multi-agent and distributed systems. I'd already had that and didn't feel the book added much to the discussion other than introducing the idea.

Academia has fashions. I feel like agent based systems will have their day again, perhaps with each agent backed by a deep learning system. It'll be easier to reason about than "convolve all these deep learning systems in a giant impenentrable network". That may be good or it may be bad.

Minsky is of course infamous for having killed neural networks for a whole generation of researchers. He lived long enough to see their resurgence, I wonder if he commented on that?

Logical Versus Analogical or Symbolic Versus Connectionist or Neat Versus Scruffy: https://ojs.aaai.org//index.php/aimagazine/article/view/894 (from 1991, and a response to the revival of connectionism that happened in the late 80s).

I often wonder what Minsky would think about the current generation of AI. My guess is that he'd be critical, because while their accomplishments are pretty impressive on the surface, they do very little to explain the mechanics of how humans perform complex problem solving, or really any kind of psychological model at all, and that is what he was really interested in. This has been a methodological problem with neural net approaches for many generations now.

Minsky was as much a psychologist as a mathematician/engineer – Society of Mind owed a lot to Freud. That style of thinking seems to have dropped by the wayside, maybe for good reasons, but it's also kind of a shame. I'm not sure what insights you get into the human mind from building LLMs, powerful though they may be.

For more of Minsky's thoughts on human intelligence, here's a recent book that collected some of his writings on education: https://direct.mit.edu/books/book/4519/Inventive-MindsMarvin... (disclaimer: I wrote the introduction).

>My guess is that he'd be critical, because while their accomplishments are pretty impressive on the surface, they do very little to explain the mechanics of how humans perform complex problem solving, or really any kind of psychological model at all, and that is what he was really interested in.

The success of machine learning/neural nets--in no small part because of the amount of computation resources we can throw at them--has really led to hogging the attention compared to fields like cognitive science, neurophysiology, and so forth. Work is certainly going on in other areas but I'm still struck that some of the questions that were being asked when I took brain science as an undergrad many decades ago (e.g. how do we recognize a face as a face?) are still being asked today.

Given that ML is the thing that's getting the flashy results, it's not surprising it's the thing in a limelight--even if there's a suspicion that it maybe (probably?) only gets you so far without better understanding how learning happens in people (and other animals) and other aspects of thinking and intelligence.

It's likely also the case that brains don't work the way LLMs do. After-all, we evolved to survive and reproduce in the real world, not generate text or images. I consider AI to be a kind of alien intelligence that approximates human intelligence in some ways, and supersedes it in others (where computing power can be fully leveraged such as AlphaGo being able to play itself a million times in a few hours).

> I often wonder what Minsky would think about the current generation of AI.

I suspect he'd react similiarly to Chomsky who in, a recent interview (MLST), was highly critical of LLMs as "not even a theory" (of what, i'm not sure... language aquisition? language production? maybe both)

Minksy was more broadly critical of NNs because it wasn't clear how difficult the problems they solved actually were. Until we had a better measure of that, saying "I got a NN to do X" is kind of meaningless. He elaborates in this excellent interview from 1990, beginning at 45:00: https://youtu.be/DrmnH0xkzQ8?t=2700

It's interesting how it is almost as though if you don't understand something thoroughly that you can't pluck the fruits from it. But if that were really true we wouldn't be able to appreciate art either, nobody can point at the reasoning behind a work of art to a degree that would satisfy those criteria, you're either moved by it or you aren't regardless of how it came about.

Present day AI is much like that: we do not understand it in detail but we understand the general ideas well enough to keep improving on it. Maybe one day we'll understand it to the degree that would satisfy a Minsky or a Chomsky, but until then we'll be happy to use the results, regardless of what makes it all tick.

I suspect - but of course absolutely no way to prove this yet - that all that such understanding would do would be to result in massive optimizations, not necessarily new kinds of output. Though such an optimization may well be so strong that it will serve as a qualitative change. And if in the process we discover something about how our own minds work so much the better but that wasn't the goal to begin with, unless you want the specification for an implementation of the algorithm of consciousness. It may well simply be something messy rather than orderly.

I think it is an interesting idea and it is sort of akin to Freud's ego, superego, subconscious, etc. It is a conceptualization, an abstraction, probably a little arbitrary, that does not map well to physical constructs.

To view it from the perspective of deep learning neural networks, one would view the society of mind as a proposed super structure on top of the various deep learning neutral networks. There is this already, like the reinforcement learning structure for Chat GPT, or the multi-focus attentional systems used for code generation.

As we built out a full AGI that can interact with the whole, it is likely we will have specialized systems which mimic a society of mind, but given that Minsky's ideas are pretty rough and sort of vague, I am not sure his writings provide the best guidance, but probably can inspire a bit of work.

Think of humans as artificial neurons. Language is how we back propagate etc

Minsky's "Society of Mind" is still relevant today, IMO. It's a provocative idea that explains the complexity of the human mind as a society of simpler processes working together. In AI, it's inspired researchers to try and build systems with lots of interconnected, simple processes that work together like the human mind. And in cognitive psychology, it's a key concept that's helped researchers understand the mind as a complex network of simpler processes.

I don't think the work has become irrelevant at all. ML models are fine but they're really just big function approximators. In the context of Minsky's book they're the sort of processes or agents which when put together and interacting could maybe constitute a generally intelligent system. Which is how they actually tend to be used in the real world already, as parts of more complex systems that interact or communicate.

The technological singularity is close for purely automate-able processes.

General AI, either in the form mimicking humans or a "being" similar is a ways off, amorphous, and far off for what people might see portrayed in fiction.

One also has to ponder the functional definitions of self-awareness, intelligence(s), and consciousness as not magical but as emergent properties of the "individual" inferred by others through behaviors, especially communication. It is anthropocentric and arrogant to assume other agents are lesser or incapable simply by lacking a common language or mutual behavioral understanding. Learning and optimization for improving one's power and resources (fitness function, gradient vectors, bank account balance;), etc.), especially through play and speed of adaptation through feedback would be strong signals of this.

I think this is a great framework to explain dyspraxia. It's not a matter of getting distracted, or slow thinking, it's random failures in each component part.

You don't get distracted and forget what number you were counting, you jump from 5 to 7 because the increment process is unreliable.

It's not that you "Aren't paying attention", you let the pot boil over because your agent that watches the environment in the background randomly left the office because you thought about something else for a brief moment and the agent that manages the stack and returns to the previous task went on vacation.

Your conscious experience is the same, but you can take wildly incorrect actions because whatever agent gave you the data is wrong.

Im not an expert in this field, but I think things have actually gone in the other direction. Look at IBM’s Watson (which won on jeopardy), and it was a system which consisted of diverse agents which would all evaluate the question in their own way and report back a result and a confidence score.

Now look at GPT, it is transformers all the way down and it is doing much more diverse things than Watson could ever do.

So I think the key is not in the diversity of agents, but in the diversity of data representations. GPT is limited by the text representing language, but what if you could train on data even more fundamental and expressive

One of the first AI proposals was from Oliver Selfridge. He called it Pandemonium because it was a set of demons (into processes) and the loudest demon was successful.

In response, Paul Smolensky made the “Harmonium”—which was the first restricted Boltzmann machine. There whichever process produces the most harmony among the elements was successful. It’s still a really great paper.

Harmony maximization was the same as free energy minimization. When Smolensky and Hinton collaborated (they were both Postdocs under David Rummelhart and Don Norman at UCSD), they called it “goodness of fit.” Still used today!

Interesting! I started reading through this pdf after reading your comment here and it has a lot of cool ideas:

https://stanford.edu/~jlmcc/papers/PDP/Volume%201/Chap6_PDP8...

Did you get a takeaway on society of mind?

I’ve been curious about the topic of distributed cognition and non-individualism in philosophy. When we don’t think that we have a single unified soul/psyche/self, then a lot of really interesting and fairly practical things happen. It’s easier to change one’s mind, for instance—and even death is less of a concern (since the essential parts of one’s self are still rather living).

I believe "the society of mind" contains a bunch of really good but unorganized ideas for building intelligent models, but was written in such a way that it remained virtually impossible to implement them into a working program. Minsky's last book called "The Emotion Machine" tries to reorganize these ideas into one giant architecture composed of at least five interconnected macrolevels of cognitive processes built from specialized agents. Having said that, "The Society of Mind" is one of the most difficult books I've read.

In a clinical setting, the idea is alive and well in the form of Internal Family Systems and other "parts work."

I wouldn't be surprised if microservices come from the same root of inspiration as well, via object oriented programming (message passing), etc.

The very idea of intelligence arising from communicating parts I think originated from that time-period and has influenced many fields, though there could be earlier references.

Unfortunately, Minsky's work is tainted by his association with Jeffrey Epstein: https://www.theverge.com/2019/8/9/20798900/marvin-minsky-jef...

I'm not sure if that sticks though.

From Minsky's Wikipedia page: "There has been no allegation that sex between them took place nor a lawsuit against Minsky's estate. Minsky's widow, Gloria Rudisch, says that he could not have had sex with any of the women at Epstein's residences, as they were always together during all of the visits to Epstein's residences."

I've been wondering about this too. The book gave me a way to think about consciousness, but I do wonder whether we'll ever see machines that use concepts at the described level. Because humans don't seem to be built that way, and the models we've built so far don't either.

You can argue that there might be some similarities or analogies to be made. But that's it. The book is actually very irrelevant and it had literally no impact on how these new systems were created and conceptualized.

https://socraticmodels.github.io/ seems somewhat related, using a LLM as the top-level model.

it reminds me a little bit of the thousand brains theory from Numenta. we'll see what they turn out in the future. i think philosophically they're a closer match to minsky.

The society of agents idea helped me understand politics more.

Of course he applies the idea to a single mind where each agent is a neuron / set of neurons, and in politics each agent is a mind.

You can follow a lecture series at MIT presented by Marvin Minsky on Ai. I think it was recorded before Nvidia fitted GPUs in a shoe box and changed the game and the price.

Society of Mind is just a bunch of unfalsifiable speculations... more New Age mysticism than science or engineering. Not sure how it would have any impact

I just might have to dust off the copy on my shelf and give it a re-read.

Minsky is hand waiver to say the least.

Reply on news.ycombinator.com