HOW DO WE EXIST IN THE UNIVERSE?
Вставка
- Опубліковано 16 тра 2024
- Atoms, things, agents, ... observers. What even defines an "observer" and which properties must all observers share? How do objects persist in our universe given that their material composition changes over time? What does it mean for a thing to be a thing? And do things supervene on our lower-level physical reality? What does it mean for a thing to have agency? What's the difference between a complex dynamical system with and without agency? Could a rock or an AI catflap have agency? Can the universe be factorised into distinct agents, or is agency diffused? Have you ever pondered about these deep questions about reality?
Please support us / mlst (and watch bonus post-interview with Friston)
/ discord
/ mlstreettalk
THE HARD PROBLEM OF OBSERVERS
This is the epic special edition show you have been waiting for! With two of the most brilliant scientists alive today.
Prof. Friston and Dr. Wolfram have spent their entire careers, some 40+ years each thinking long and hard about these very questions and have developed significant frameworks of reference on their respective journeys (the Wolfram Physics project and the Free Energy principle).
Panel: MIT Ph.D Keith Duggar
Production: Dr. Tim Scarfe
Pod version: podcasters.spotify.com/pod/sh...
Refs:
TED Talk with Stephen:
www.ted.com/talks/stephen_wol...
writings.stephenwolfram.com/2...
TOC
00:00:00 - Intro
00:10:45 - Show kickoff
00:13:23 - Wolfram gets to grips with FEP
00:37:53 - How much control does an agent/observer have
00:45:37 - Observer persistence, what universe seems like to us
00:51:16 - Black holes
00:55:52 - Inside vs outside
01:03:05 - Moving away from the predictable path
01:06:11 - What can observers do
01:17:35 - Self modelling gives agency
01:22:11 - How do you know a thing has agency?
01:33:33 - Deep link between dynamics, ruliad and AI
01:36:37 - Does agency entail free will? Defining Agency
01:43:42 - Where do I probe for agency?
01:49:58 - Why is the universe the way we see it?
01:53:35 - Alien intelligence
01:54:25 - The hard problem of Observers
01:57:05 - Summary thoughts from Wolfram
02:00:20 - Factorisability of FEP
02:07:50 - Patreon interview teaser - Наука та технологія
This was a mind-blowing conversation between two giants of science. Thank you MLST for this amazing episode!
Most of this discussion consisted of Wolfram asking what FEP is. Much of that can be found in published works by Friston. Had Wolfram done that background reading, this discussion could have gotten to more of the issues between FEP and Wolfram's "Rule-iverse." As it was, this discussion ran out of time before it could get to explanation of that Rule-iverse. However, what was discussed was interesting, well worth the viewers' time. I hope there is a follow up video to complete this discussion.
All true. Nonetheless, I'm happy it went the way it did. I imagine most people are less familiar with Friston's work than with Wolfram's. For me at least, I appreciated the fairly comprehensive introduction to FEP.
It was also a pleasure to see someone as eminent as Wolfram display such genuine interest and curiosity for someone else's theory.
@@dbz5808 Wolfram is well known for his product Mathematica. I question whether his scientific work is comparatively well known. Googling Ruliad yields about 10,000 results, the string (with quotes) "Free Energy Principle" generates about 220,000. The physics community has been fairly sanguine about the Ruliad in their review of it. Apparently, the AI community (as represented here) is more appreciative.. FEP has received support from a broad and expanding area of academic disciplines. This was my introduction to the Ruliad. The explanation Wolfram gave here, even if short, gave me no enlightenment about it, whatsoever.
I went to sleep while the video was playing. then later i woke up while the video was still playing, and Wolfram was still asking questions
In the first 30 minutes, the physicist's approach reminded me of my initial therapy sessions, where existential queries were posed with a scientific lens, akin to seeking wisdom from a clergyman.
Overall, their meeting is a testament to the beauty of interdisciplinary discourse, where the pursuit of knowledge transcends the boundaries of individual fields, encouraging us to think more holistically about the universe and our place within it.
Part two?
Great collaboration between Wolfram and Friston!
Stephen is the consummate depth interviewer. Having Stephen as a guide through Karl's mind is brilliant! Thanks for another perfect episode.
IDK how you guys pulled it off, but the last few interviews/videos have been stellar! great guests and conversations! it's like yoga for my brain...stretching and bending it in ways that kind of hurt at first, but eventually starts to feel better, stronger, and agile.
Keith- I really appreciate your demeanor and the contributions you brought to the table on this one. It seems as if things were just starting to get unpacked and hopefully next time Wolfram will have more time. Nice cameo by Tim at the end. MLST has become an integral part of my life. Keep on keeping! Many heartfelt thanks
Thank you, sincerely!
👏
Would love a part 2 where Friston did most of the talking/questioning of Wolfram. He might be too polite for that, however. ?
Fantastic discussion, well filmed, so much to think about. 👏Thank you.
FEP assumes (by observation) that things exist in the universe that are clearly bounded (the Markov blanket being that boundary), tightly connected internally, and sparsely connected but interacting with ttheir environment. The sparse connections to its environment is an important property when comparing FEP to the Ruliad. Friston's "things" by definition make them observers in the Ruliad system if such things are real / exist. Wolfram states that Ruliad theory has been unable to separate out any observer because it has not found a Ruliad entity that has sparse, interacting connections to its environment. This presents a real challenge, If FEP things do constitute Ruliad observers then being unable to recapitulate them in the Ruliad is catastrophic to the Ruliad theory. Wolfram intuited that FEP things, which are so intuitively obvious, must be a confusion caused by our inability to observe the real universe, i.e., the Ruliad. This imposes a test on Wolfram. It is incumbent on him to prove that FEP things are indeed illusions as Ruliad observers, or radically change his views of the Ruliad or dismiss the theory altogether. This is the very next task I would challenge Wolfram with.
Brilliant conversation!
Nailed it Keith. Absolutely nailed it.
Wolfram "You're getting me to think?!" That's the moment right there...
Thank you kindly! I feel I could have done better as moderator, especially carving much more time for Prof. Friston to explore Dr. Wolfram's thought space. However, if in the end we've helped these two brilliant thinkers connect, even if asymmetrically, I'll count that as a winning step forward.
I hope we can facilitate another two conversations with clearer agendas as there are at least two major areas they would like to explore together further: 1) Identifiability and observer framing in the Ruliad 2) designing cognitive computing e.g. VERSES AI.
Yeah I don't think it was doable in just one sitting, but that wasn't necessarily clear going in. We strive for balance where possible but too even is just boring. The good grace with which these two guests went back and fourth just made for such 'a (long) rush!' as @AndersHansgaard so expertly put it.
MLST just going from strength to strength right now. Great guests, great questions/opinions and slick production. This channel is a treat, just keep it coming. 🙂
@@nomenec
Thank you so much for this guys! Brilliant! ❤
Amazing, thanks for making this interview happen. I have been long trying to find common grounds between the principles these two great scientists have discovered throughout their life.
Been listening to all Friston’s and Wolfram’s episodes on MLST over the years. Many ideas harken back to cybernetics. You can hear it when they mention things like feedback, thermostats, control, etc, and in this interview Friston does mention Ross Ashby, a pioneering British cybernetician. I’ve yet to see an episode dedicated to cybernetics though on MLST. Perhaps coming up?
Also Norbert Weiner
Wiener
You want to minimize surprise now, tomorrow and days after. That's why you explore.
well said
Discrete and discreet
Lovely comment!
@@benjones1452 what a difference an ear makes
You recognize other minds by identifying attempts to maintain homeostasis. When you observe part of your external reality being sticky to being a certain way contrary to your understanding of the rules governing the movements of dumb matter/energy you've identified a program. If you can also recognize in that program an ability to recognize other minds and adapt itself accordingly you'll have identified another mind.
Regarding the difference between a rock and a brain,
I think the main differences can be accounted for with three related notions:
1) Nested Markov blankets
2) Internal feedback
3) Counterfactual depth
The one I think that was missing from the discussion is 2, the idea that a brain can have action-sensation loops that remain internal to the overall boundary of the brain by way of it having nested Markov blankets. In contrast, because a rock doesn't have enough internal Markov blanket structuring, all of it's action would be directed outward.
This might also be the distinction between free will and lack thereof. Namely, a system is acting in accordance to its free will when the sensations (inward bound causation) comes from the agent itself and remains internal to the outer most boundary of the agent. Furthermore, the further inside the Markov blanket the relevant actions are casually sourced from, the more attributable the results of that are to the "free will" of the overarching agent.
---
With regards to looking at the bits of some system (e.g. an LLM) and figuring out if it's agentic,
I'd think that you would just look for a nested Markov blanket structure with some threshold level of nesting depth, together with a small-world causal relation structure, where most bits are related to local bits, but some are related to distal bits. Where distance is measured with respect to the most internal Markov blanket that contains the bit of interest.
---
Regarding the maximum complexity of an environment that can house an observer,
my understanding is that as long as the equivalent graph is not fully connected, where each node connects to each node with the same weighting, then you necessarily have a Markov blanket.
The constraining factor is whether the environment is stable enough to have _nested_ Markov blankets.
I was thinking much the same regarding 2, that seemed obvious and crucially missing.
If the rock doesn't change internally it is not observing anything.
I would add that a rock doesn't have any differentiated parts (at a particular level, it is all very similar molecular structures), and it is rigid, it cannot change the internal complexity of itself.
It also reminds of 'how do you determine if something is living/alive or not?'
How to determine either observers or life from a bucket of bits it beyond me though, but I perhaps some machine learning model could be trained on that problem?
This is beautiful. Thanks. 😊
When I listen to Wolfram...I can't help thinking: "Someday...this guy is gonna have 'Nobel laureate' attached to his name" (...and the more I listen to Friston...and S L O W L Y grok what the hell it is he's actually saying...the more I see him in the same position). A conversation between Wolfram and Friston is like a dream match! Two truly singularly unique and exceptional minds...but also two truly distinguished and decent human beings.
Ha that's so funny and true! I've often had to watch segments of our own recordings of Friston a couple of times to grok them. Likewise I've had to read various segments of Wolfram's publications over a few times. But, it's been worth the effort!
An observation is an entanglement but not necessarily only an entanglement. The background fluctuations in the video of the two black holes is what drives time . This means that time has a random component and also that time do not have a build in direction. As the fluctuations corresponding to the pendulum in a clock the hands in a clock are the particles in space.
I asked ChatGPT to imagine this conversation. Here’s what it came up with (not as good as the real one, but interesting nonetheless.)
Sure, imagining a hypothetical conversation between Stephen Wolfram, known for his work on computational universe theory, and Karl Friston, known for his free energy principle, regarding the problem of the observer would be an exciting intellectual exercise!
---
**Wolfram:** Karl, your free energy principle posits that the brain continuously predicts sensory input and corrects its model of the world based on errors in those predictions. My work in the computational universe suggests that everything, including our thoughts and consciousness, can be understood as a computation. Could the brain be viewed as executing a sort of program to minimize surprise?
**Friston:** Indeed, Stephen. One could see the brain's efforts to minimize free energy as a computational process. But, in your framework, where does the observer - the entity that experiences the world and has a subjective perspective - fit in?
**Wolfram:** In the computational universe, any sufficiently complex computation can be associated with a kind of observer. The observer doesn't have to be conscious in the way humans are. It's just a reference frame from which computations and their consequences can be "viewed" or "experienced". Essentially, every computational process is its own observer.
**Friston:** That's intriguing. In the free energy framework, the brain constructs models of the world, acting as an observer that continually updates its beliefs. The Markov blanket defines the boundary between the observer and its environment. Your computational processes could be the mechanism that actualizes these belief updates.
**Wolfram:** Right. The challenge in my framework is to pinpoint where consciousness arises in these computational processes. Some argue it emerges from specific types of computations, while others believe it's more widespread. The idea of minimizing free energy might be one of the computations associated with conscious processes.
**Friston:** Agreed. Consciousness could be seen as the brain's most sophisticated way of predicting and interacting with its environment. The better it predicts, the more "aware" it becomes. The problem of the observer in my framework is intertwined with consciousness, as it's about how the brain models its own place within the world.
**Wolfram:** So, in essence, both our frameworks, albeit from different angles, are trying to grapple with the complex nature of observation, experience, and consciousness. Whether through computations or predictive models, we're trying to understand the nature of the "self" in the universe.
**Friston:** Precisely. The ultimate question remains: How do these processes give rise to the subjective experience of being an observer? While our approaches might differ, our goals are surprisingly aligned.
---
Of course, this is a fictional conversation and might not accurately represent the full depth of their thoughts, but it gives a glimpse into how these two intellectual giants might approach the problem of the observer.
I'm just catching up with the ideas of each of these giants, and found this conversation amazing and informative. Thank you so much for bringing this to a popular audience of the curious. 🙏👍
The question of "what is a choice" came up and I have MY definition of what it is vs a reaction (the kind of reaction a thermostat makes). A reaction is an action based on inputs. A choice requires a prediction made by an observer based on inputs, and the baysian evaluation of a course of action based on that prediction.
So I as an observer take inputs and attempt to predict my environment's future based on my mental models. I then react, not to my direct inputs, but to my model's prediction.
To me, that is the definition of choice.
you basically described model predictive control
@@user-tu1lm3kb7t I will have to look that up. Thanks!
What a (long) rush! Thanks to everyone involved for bringing us such an awe-inspiring exploration.
Pretty much no doubt that listening to this just a single time is not enough.
Thank's again. Such a great interview. There’s also room for more conversations around Wolframs implication that FEP doesn’t model for exploration. In this regard it could be fruitful to have Karl discuss with Ogi Ogas. In Journey of the Mind Ogas and Gaddam explore and clarify Stephen Grossberg’s ideas on how consciousness could arise through single celled organisms creating a balance between ‘safe’ and curiosity or hunger driven behaviours which may increase survival risk (Friston provided a blurb so is familiar with it). Also interesting about the FEP is it might offer another way of bridging the gap between the activities of non living (and presumably non thinking) things like molecules or protons, and that class of things that do create some form of model of their environment, however minimal, which humans begin to ascribe life to. In other words the ‘hard problem’ (so Chalmers, or Chris Fields). There’s a further interesting aspect to be explored to map the FEP across some concepts used by or created by Stuart Kauffman; the NK map with its strange attractors (which is useful to model objects that persist in time though not made of persistant materials, such as eddies), and the ‘Edge of Chaos’ which is a space of states which are neither too chaotic, or too ordered. It looks like most things evolve successfully if they explore the edge of chaos and not the frozen or chaotic states. So I suspect that as Wolfram is implying, the FEP is not enough to explain how certain classes of organisms persist. If all they do is minimise surprise, they will die out. They must contain a curiosity element, which in practice all living things do due to their need to acquire their own energy. It may be this hunger, together with the FEP, that is at the origin of living systems. It’s conceivable that living systems are those that are capable of running an algorithm that best balances the tensions between ‘hunger/curiosity’ and ‘surprise minimisation’ in its ecological niche. It might be that this could be a threshold for what makes a living system ie ‘ a non equilibrium steady state entity that uses exogenous energy, and persists through a period of time by balancing FEP with behaviours that will enable it to find the exogenous energy’. Kauffman also described an ‘autocatalytic set’ which could provide a model for an algorithm that is in some sense autocatalytic in that it will explore its niche while in some meaningful sense updating its priors through a form of ‘autocatalysis’. Finally Kauffman has been talking recently about how the non computability or inherent non predictability of natural evolution means that physics has limits to what it applies to. This would be an interesting question for Deutsch, Wolfram, Chris Fields and Friston.
So excited!
I’ve been very excited for this to drop! The All in Podcast with Stephan Wolfram wasn’t enough for me 😅
I always watch at 2x, but this is too interesting to miss live at 1x
Good stuff. Needs a follow up discussion. I think you should also make a video with narration breaking down some of the ideas going back and forth.
If an LLM is "hallucinating," it's producing counterfactuals. Get the top dozen logits, and some may be grounded, and the others may not be. We require reflection to make a good decision about groundedness. And yes, this in itself is agentic, but I think we need a functional definition of agency that includes action. Reflection is certainly an action compared to a function spitting out logits. So would opening a cat flap be an action. Therefore we need a set of criteria to differentiate between an agent's behaviors and the output of a generative function. In the end there is no real distinction, but as engineers, functionally or operationally, we need to believe there is one.
This is science podcasting at its finest.
I would posit that one answer to the FEP paradox of staying on the path is this; the more you know, the more predictable the universe is.
One neuron starts an activity and seeks out a set of neurons in the brain to form specialized networks to search for and solve problems, while other clusters keep life going in the background. The same can be said about people. An individual starts working at a task and then networks with people with similar observer attributes. As we become more connected the clusters become larger and more prolific. All while having many networks working on keeping this a safe and predictable as possible.
We're so lucky to be able the experience these 2 talking
I have so enjoyed this discussion between three professionals. I have to listen over and over to appreciate what they are saying and the depth I can understand of the discourse. Thank you very much I am totally entertained and increasing my awareness.
my feeling is that these two, each, have a piece of the puzzle that is existence: the observer and the observed.
part 2 is a must!
I have no idea where I’m going to find a spare 2 hours to watch that all over again. Thanks for bringing this to us gents ✌
Thank you!
Loved Keith’s final comments! Hoping future discussions will pick up and expand on this!
What we’re really looking for is the most efficient search physics. Coherence measurements, together with harmonic relationships arcoss scales and contexts, seems to point there. What you want is all the pieces to organize themselves based on their attributes expressed as forces or fields.
Fascinating - this conversation grapples with an old question of anthropology: how do humans deal with the chaos of the world? How do humans minimize unpredictability so that we can live in a knowable world? Anthropologists would bring meaning and culture into this conversation…
These guys amazing!
Nice! BUT I would have liked Stephen Wolfram to tell more about its ruliad thing...and then giving Karl the opportunity to ask questions...easy to interrogate someone but somewhat more difficult to explain and answer critical questions
out of all the podcasts i have listened to .. this one seems to be the most amazing one .. thanks for this .. i would like Lex Fridman to listen to this and comment on it
Agency is when you can act in your own interest, so you need to be able to have different internal states and act in the environment depending on your internal state. To pit your opinion against everyone else's, and not just be moved by your environment.
So, a sensor that opens a door when someone passes under it doesn't have agency, since it always just opens the door. If you'd program it so that a key card is needed in the evening though, it now has an opinion, a self-interest to defend, as in it will decline to open the door unless you appease its internal constraints. In this case, there are only a few possible internal states, but as you add more complexity you can see the mechanism/organism essentially developing preferences, likes/dislikes, trauma, etc.
Agency is the degree to which you can make the world act in your interest? So, a boulder is agentic only in the sense that it forces everyone to walk around it, but since it only has one possible internal state, it can't "act in its own interest".
The funny thing is that when you're talking about coupling, you could argue that the organism/mechanism that has agency in the case of the boulder people walk around is not the boulder itself, but consists of one part the boulder and one part the minds of the people walking around it (specifically, the desire to get to the other side, reliably being used by the boulder to coerce people to walk around it).
You could argue that the boulder is acting through the person, of course that would assume intentionality (the rock "wanting" to stay in its place and have people walk around it). Boulders are passive, so it's fair to say that it's only an agent from the perspective of a person having to walk around it.
Perhaps if you'd turn it into a statue, intended for people to walk around and admire for its beauty, then you could argue that the boulder gained agency (it now "actively" coerces people to serve a self-interest, rather than just passively maintaining itself). If you turned it into a stairway though, you could argue that the boulder devolved and lost agency (the environment defeated the boulder).
Note by the way that the statue has more internal states than the boulder, if you account for all the ways it is observed by the people who interact with it, who could in some sense be said to be part of its body/mechanism/organism.
In terms of coupling and agency, remember also the quote from Jung "do people have ideas, or do ideas have people?". Food for thought :).
LovedLovedLoved this! Looking (with wild surmise) for Part 2: "Karl Friston Interviews Stephen Wolfram on 'Observers!'"
This talk is amazing.
Super position is free will, one a steady state, the other a state of possibilities, division of a steady state into angles of view. Consciousness is the geometry of a steady state. Gravity it's mode of operation.
@24:50 I couldn't help thinking that obligatory reading / watching in advance of a chat with Karl is the wonderful interview he had with Sean Carrol on Sean's Mindscape channel (Episode 87) where @ 23:38 Sean highlights a wonderful paradox from a Physicists perspective which Karl removes by saying "minimising free entropy will require a maximisation of entropy from the Physicists point of view and that's right but the key distinction is that: what I will do at this moment in time will always be to maximize my accuracy (or my energy) in terms of minimising prediction error whilst at the same time maximising my entropy, keeping my options open because the entropy that we are talking about is an attribute of a belief about the causes of my data so this is not an entropy measure of the brain physically". The idea of an entity modelled using a markov blanket seeking it way towards its attracting (using the free energy principle) set in a non equilibrium steady state is a wonderful "north star" in understanding emergent properties in complex Systems of such systems. The "dark room" paradox questions to me seem unfair or premature, after all Karl is not suggesting that the compass is sailing the ship. Stefan Thurner ( professor for Science of Complex Systems at the Medical University of Vienna ) has a cracking book "Introduction to the Theory of Complex Systems" with a brilliant treatment of: "The three faces of entropy for complex systems -- information, thermodynamics and the maximum entropy principle". It would be great to hear what Karl and Stefan would have to say in a future episode. Paper ref here [Stefan Thurner, Bernat Corominas-Murtra, and Rudolf Hanel Phys. Rev. E 96, 032124 - Published 15 September 2017]
Since the 1st time we designed purpose into the function of an object we've been searching for owe design, function, purpose and mortality. It's about that's immortal and that which dies. In a discrete system separation has to be superposition, not what's inside and what's outside. CA is the 1st real break regarding this question outside of DNA and RNA. Now all we have to do is identify the energy source. Throughout what we see magnetic fields generate electric fields and vice a versa, gravity seems to be same; compression could be the source of consciousness, a property of gravity.
This is probably the most accessible talk about free-energy principle out there! Probably it takes a WolfRam to take it out of FriSton 😊 Thanks for doing this 🙏
Inspiring! Thanks
This was fascinating and a much appreciated dive into the FEP. I wish, however, Karl had had more of an opportunity to ask Stephen questions. Can you perhaps get these two together for another round at some point?
The confusion about curiousity can be explained in this metaphorical way: cells in a human body may behave in ways that do not simply preserve their own existence, but that of the body as a whole, similarly, a human being may behave in ways that do not directly seek for their own existence, but that of the society. Curiosity and novelty seeking behaviors increase the fitness of the ambient super-organism of which each person is a part.
Observer doesn't need to have a choice or free will about what to do. The fact that all known observers can act is because observeness developed evolutionarily to help the thing to promote it's existence in the first place. That's why it was naturally selected to be an actor, not just passive thing. When you passive, you don't need observeness because it would help you survive. Some bottom ocean organisms have a brain at larva stage and then when it stick to the bottom, it consumes it's brain as the source of nutrients and not useful in any other way.
Watching this video is actually altering my reality!
This conversation is truly a fountain of inspiration; even when certain aspects might elude my grasp, I find myself irresistibly drawn into deep contemplation on the topic at hand. My heartfelt gratitude goes out to Dr. Friston and Dr. Wolfram, as well as to the remarkable team at MLST for their outstanding work.
A question, at around the 38:00 mark of your discussion. You delve into the topic of an observer navigating along the path of least uncertainty, with uncertainty stemming from the observer's current model of its environment. I am curious to know when the observer strays off this path and ventures into realms of greater uncertainty, is there a method for the observer to discern whether the perceived randomness or uncertainty is a result of inadequacies in its own model? Or is it an indication that the observer has indeed stumbled upon a domain of computational irreducibility?
Perhaps what I'm trying to ask is: Is it possible for computational irreducibility to be a universal phenomenon, applicable across all observers and thus discernible from one's own model? Or does its existence inherently depend on the limitations of the observer, implying that computational irreducibility is a relative concept, emerging due to inadequacies in the observer’s model of its surroundings?
Thank you in advance for any replies :)
Thank you for your kind words! We put so much effort into this work and it's wonderful to have it appreciated. As for computational irreducibility, below is my understanding.
Computational irreducibility, i.e. no short cuts, is a statement about a complete computation. In other words, take a system of N states (think of these are cells in a cellular automata or nodes in a hyper-graph), and apply a set of computational rules, R, for some number of time steps t. Let's denote this computational evolution as: N0 -- R.t -> Nt. This computation (N0, R, t) is computationally irreducible if there is no simpler set of rules S that can be run for steps s Ns == Nt. I.e. there is not a simpler computation which would allow us to know "everything" at time t which is Nt.
However, computational irreducibility does not imply the absence of patterns, i.e. functions of Nt, which can be computed by a simpler set of rules. In other words, even if (N0, R, t) is computationally irreducible, there may be a function (a statistic for example) F(N) such that F(Nt) = F(Ns) so that we can predict F about the irreducible (N0, R, t) by applying the simpler computation S and then computing F(Ns).
In fact, in our previous Wolfram interview, he states there are in fact a near infinite number of such predictable patterns. And, interestingly, the degree to which an observer is computationally bound, i.e. just how limited their computational capacity is determines which simpler computations and patterns, S and F, they are able to discover!
Indeed, human intelligence is now significantly externalized and distributed all around the world throughout our people, our books, our computers, our drives, etc. The ongoing technological development of human kind for tens of thousands of years has been pushing our combined computational bound further and further out. As thus we continue to discover more and different patterns in the Universe.
Yes thanks for this, it was the most interesting point for me too, Keith engages immediately, reaching for the critical insight. We've heard Keith and Karl speak together and it's alway fascinating, that's the fruit of mutual respect.
Love Karl Friston but the best FEP explanations I've come across ( I still don't fully understand this) have come from persons other than him. Wolfram has a knack for distilling things like these and so this was a great watch.
There is a free pdf book you can download by Parr et al, highly recommend. Friston admits himself that others explain FEP more succinctly. direct.mit.edu/books/oa-monograph/5299/Active-InferenceThe-Free-Energy-Principle-in-Mind this is the book, I’m on my phone now but Google around there is a free pdf download
@@MachineLearningStreetTalk thank you so much for replying :) I'll definitely have a look..
Regarding my life, I seek shelter so as to predict a safe place to be. I find it, then next thing I know, I've thrown it away. Like a gyro I find a new balance in a new adventure. Karl and Stephen find their way, but never through it away.
This chat /discussion represented a concerted effort to establish commonalities and facilitate a dialogue between the internal and external realms, each characterized by distinct attributes and experiences. Notably, these two intellectual figures have been rigorously examining some of the core assumptions that underpin our daily conceptual framework. This framework shapes our internal representation of the world, guiding our predictions and strategies to navigate and excel in the face of external challenges. The exploration centers around the realms of possibilities and potentialities, particularly in scenarios where our predictive capabilities encounter disruptions or unpredictable elements. While some parts of such discussions especially those by Wolfram and his thought experiments on black holes and singularities, might initially appear as overly abstract, the timely inputs of prof Friston have been instrumental in anchoring this otherwise abstract exploration in the tangible aspects of our experiential reality.
Excellent video Much Love for the Nudge ❤
I think it's safe to say that Wolfram is more of a philosophical realist, and Friston is more of a philosophical idealist 🙂
What makes this discussion great is that they both seem to value pragmatic philosophies. That is they want to apply their ideas to real solutions, make actual products that do something and help us in our daily lives, so their differences in philosophical approaches is beneficial for all of us.
Marvelous conversation. High Intelligence truly listening to one another. /The one and the many/
I felt that this was an opportunity missed. I was disappointed that Stephen Wolfram packed in so many questions in one rambling segment after another. I felt it was difficult for KF to be his usual crystal clear self as he was left searching for which question among questions to respond to. We didn't get into the full ramifications of FEP and active inference, let alone the ruliad. So let's have more! And more focus... Thank you for organizing this event
On my planet I find myself tossed about , but then I stop the balancer and through myself, I toss myself into the unknown. Argue about free will without ever trying free will. I've tried it and I'm still standing. Every time I do so I scare me. Free will is scarey.
There is a harmonic relationship within a recursive feedback network that gets closer to “self” than one might imagine. Picture a camera pointed at its own monitor, and there is a new thing created. This feels closer to “self” than most other things, like introspection is thinking about thinking, both are regressive reflections.
I’m slightly more than halfway through and still the “feel” that both of these guys might think there’s a single fixed scale at which thickness and observer-ness appears.
If so, i think this is wrong.
I believe things, per se, can be defined… no, *must* be defined… as constrained by three things: temporal boundaries, spatial boundaries, and interaction boundaries. Things are not things. Independent, spatial, temporal and interaction boundaries.
(A target cell’s apprehension of a single real spike protein as a “thing“ is quite a bit different than our linguistic abstract perception of the spike protein as a thing. If this difference is obscured, much is lost. “ Things are not things”)
This is also brilliant and ridiculously complex. Delightful!
Now, off to go learn (a lot) more about Friston’s work!
Thanks to video creators. This was an impossible job and the presentation was top drawer. Thank you.
Ok! At about 1:30, they start to get into it…but…Wolfram keeps trying to push it back to “bits”, which is the lowest possible scale.
Not commensurate.
(Implicitly, they are talking about detecting quality in part ( not exclusively) and Friston seems to know there is no solution to this…)
Beauty to discovery then loop.
💓 High value contect.
Makes sense to me, complexity emerges from simplicity. In the beginning was the rule, and the rule said let there be space, and there was space. Modern physics hack the universe, which is great, but the hack isn't fundamental. The rule is fundamental to maths, so logically that has to be our starting point.
Nice.
Mathematics seem to be tricks or shortcuts that we memorize from our compressed perception of the world.
I love to be sitting at home listening to this. Such a privilege.
Great conversation. If I’m tracking (and I may not be) I’m curious if understanding consciousness as a “coherence inducing operator” and an emergent VIRTUAL property (ie Joscha Bach) would bring more clarity to Dr. Wolfram’s model of how the computational observer fits into the computational universe? It seems to me that the virtual space has different rules from the the physical universe even if they are integrated. While the virtual space runs on the quantum hypergraph, We can’t predict or model what is happening to each particle in the universe (eg electrons in our brains) without incorporating psychology and social science. Here, immaterial cognitive models take on more than than physical laws that govern particles directly (if/then) but also influence other physical-law creating systems in operating in the noosphere (other conscious organisms and machines). All this to say is that the observers experience is not directly beholden to the same physical laws that we can observe in the same way. Sorry if I’m totally missing the thread😅
We need a round two.
Loved this conversation, thank you!
I don't believe electrons are black holes though.
Re electrons, true, but they are tantalizingly close!
en.wikipedia.org/wiki/Black_hole_electron
Great! Stephen Cellanlaged Karl Friston
Embedding artificial intelligence (AI) within a latent space to make it intelligent presents a groundbreaking approach in AI research, merging machine learning with data representation and AI autonomy. This innovative concept enables AI to actively interpret and manipulate data within the latent space, using advanced machine learning models. It allows for dynamic data management, where the AI can adjust the latent space structure for more effective data representation. Additionally, its self-learning capabilities enable continuous adaptation without external input, enhancing predictive and prescriptive analysis.
The potential applications of this technology are vast, ranging from advanced data analytics in various sectors like finance and healthcare to personalized recommendations in e-commerce. This intelligent latent space can offer deeper insights, predict future trends, and suggest strategies, significantly advancing AI's role in decision-making. While it poses challenges like computational complexity and ethical considerations, the benefits of improved AI autonomy and sophisticated data analysis are immense. This concept marks a significant step forward in AI research, promising interdisciplinary innovations and a new era of intelligent data processing.
Could we think our way out of it? I think we possibly could, in a sort of science fiction sense,
9:03
simply because we've got the right kinds of models of our lived world, which is just the physics that you're talking about.
9:09
But lots of other observers would not be able to do that unless they had that right kind of internal model, that implicit way of modeling the
9:19
causes of all of my sensations. I mean, I think you were talking
9:25
about thinginess, so to speak, which is a lower bar probably than observer ness. I mean,
9:32
thinginess might just be the ability to maintain an independent object,
9:38
maintain something that we can consider to be a persistent object. Yeah, there's Chris Fields. Chris Fields. Yes.
9:44
And so he frames exactly what you were just saying in terms of there
9:49
being little pockets of concentrated planning, sort of sort of self modeling of the kind that invokes the consequences of one's own.
9:59
An agent, he'd articulate that in terms of a minimal Markov blanket. So something that cannot be reduced, that has, if you like,
10:08
the capacity to be an agent, and then it sees itself enacting its plans through a hierarchy of subsequent Markov blankets.
10:17
So you could say that the, you know, the inclusive Markov blanket around
10:23
this little packet or irreducible Markov blanket was an agent, but the agency sits at the heart. And of course he thinks that's
10:32
consciousness. You're both absolutely brilliant. And I think Tim and I both look up to you quite a b
so part of the simplicity of the the free energy principle lies in that almost tautological
16:18
framing of the free energy principle, whereby you're just saying that if something exists in harmony and in relation to its environment,
16:29
then it will show these kinds of properties that can on a on a certain reading, be interpreted as minimizing surprise, maximizing predictability.
...1:32:25
is a reasonable thing to say. On the other hand, if one's trying to make a distinction between observer like things, well, there's a couple
1:32:34
of levels of distinction here. There's the distinction between just everything and things and then things and observer like things.
1:32:43
And I think the and I think several of these distinctions are kind
1:32:48
of complicated. And I'm, I'm. You know, in order to have
1:32:54
something which seems like a a kind of a scientifically progressing kind of set of ideas, it seems like one really has to
1:33:03
be able to make a definite distinction between sort of the.
1:33:11
You know, how do I tell from the bag of bits that I have a an observer
1:33:17
like thing, as opposed to just a thing and to, to kind of make that a bit more technological?1:37:15
think that's agency, even if tell me what you mean by agency. What?
1:37:21
So what I mean by that is, is literally that it's performing a computation about sensory inputs. And as a result of that computation,
1:37:30
deciding some some output. But so any old classifier, any old machine learning classifier does that.
1:37:38
Does it have it's not performing. It's not performing an action though, right? It could be if you connect it to
1:37:44
a robot or something, it can perform plenty of actions. And as soon as you do then, then it becomes a then it
1:37:49
becomes an agent, you know, acting, acting in the world. So, so I mean, the, the, you know, image identifying cat flap that
1:37:57
decides, is it a cat or not? That cat flap has agency in your
1:38:02
data. I would I would say it does. It has a very simple form of agency. But I think that that's so. So your definition of agency is
1:38:10
that there is there are large number of possible inputs. You are sort of, you know, whittling that down to something
1:38:17
about which a decision can be made. And then you are actuating in
1:38:23
the world perhaps. So, for example, does it have agency in some environment. Yeah. Okay. Does it have agency if the only
1:38:29
thing it does is to change a bit somewhere inside itself?...
You know, we don't have a way to describe how brains work at an
1:45:16
intermediate level between kind of the neuron firings and the words that
1:45:22
get spoken by the brain type thing. In other words, that. And so in the case of, you know, the I, the whatever else,
1:45:30
I'm curious because that's the case where we don't have any limitation on the underlying data. We know what every bit is.
1:45:36
And so now my question is, can I run, what test would I run, what kind
1:45:41
of correlation would I measure? What kind of what would I do to say, hey, you're a counterfactual thinking I and this one over here,
Posit:
"An agent is your 'avatar' in the context 9f the environment whereto you refer."
Keith, I think we need to be careful about extending agency outwards. Would we need to include the ocean as a tool and the wind, etc. as all part of the agent; for in order for the plan to be successful, all these things are utilized. I really don't know where or how to set the boundary. I suspect the problem is multidimensional and difficult to describe. Perhaps this kind of understanding is intractable for us. Are some "suitcase-words" highly dimensional ideas? We model semantic space, but how complex is actual human semantic space? If it models the Universe, however imperfectly, it's necessarily intractible.
I would have loved to ask.. Are we as observers individual or a collective? given that as a species we build upon knowledge over many contributions... Therefore what is the boundary of the Markov blanket of human knowledge.. And does it encompass consciousness?
I've been running a ruliad since 2003.
I think Wolfram's question about how we can explain the existence of scientists using FEP can be re-framed as "How can we explain risk-taking and surprise-minimization simultaneously". I think the answer must involve competition and reward.
It's worth noting that Wolfram seems to believe that the components that constitutes agency can be formalized.
So I'm curious...yes but Karl is curious also! What Stephen (as an observer!!!) doesn't pick up on is that a good and productive interaction is NOT a one way street of interrogation...round two with a more switching of roles....please
Agree, Stephen has a clear lack of intelectual empathy, asking question just like a policeman. Perhaps the moderator should be more active in balancing the conversation.
The best contemporary thing about Galileo is that his middle finger is on display at the museum dedicated to him, in Florence.
It cannot happen but a conversation with Woolfram and Douglas Adams would have been an amazing thing. ( maybe it did happen ) their time lines were at some point available.
At 59:00 compute resources are insufficient.. in this universe.
If someone could communicate with other universes in a multiverse, could they then accumulate enough compute power to simulate this universe?
I know we can't even tell if there is a multiverse or not, I am just wondering if it in principle would be possible to make an algorithm that computes simulations in parallel universes and returns a prediction in one?
thanks for making this happen!
a bit of feedback: the slow zoom in and out of friston's face was very distracting, especially juxtaposed with wolfram's still face, and I *really* wish you hadn't edited it that way
Recognizing what is expected or not, what is signal or noise, is dependent on the framing applied by the observer. How would an observer understand whether their particular framing is accurately capturing the expected or unexpected signals and noises?
It's easy to say this works for scientists, because scientists seek the unknown in order to make the world more predictable, and they'll happily spend most of their career just narrowing down some variables to a quantifiable replicable value, and the nice neat ordered house where everything is predictable works great for neurotypicals,
But there are minds like mine that just genuinely seek diversity, uniqueness and chaos because they hate the repetitive and predictable with every fibre of their being.
I seek novel stimuli not because it makes the world more predictable, but precisely the opposite. Repetition is boring.
I like new unexpected art, drug-fuelled hallucinations, concepts and structures for their own sake. Understanding will strip the joy out of anything.
I say that as a mechanical, electrical and machine learning engineer that regularly uses reductionism to strip the joy out of things I love, because that has it's utility too,
but there are some things that are necessary to understand for a basis to explore the unexpected safely.
I think it's worth separating the concept of the unexpected from random chaos and potential danger. I don't want to experience pure entropy, I appreciate structured chaos.
Pure entropy is pretty predictable on the macroscale too. You can kind of expect it to all kind of just swirl around in a hot mess, probably not spontaneously forming a mazda miata or a boltzmann brain that lasts until the end of the universe.
What you describe is curiosity, which is very much a part of FEP and was given attention in this video.
Rare occasion when I can better understand Wolfram as other participants :D
So good, thank you. Since these are heavyweight prize fights, what about putting Karl in the ring with David Deutsch, who afaict seems less enamoured with Bayesianism....
Not a bad idea! I loved his recent interview on Sean Caroll
@@MachineLearningStreetTalk Add in Robert Sapolsky who appears to have a negative, at least critical view, of Bayesian brain theory.
A foton, i belive to be made of 4 microspaces conected by a mathematical low, 2 vortex enthengeled. Electron îs some more complex mathematical structure.
I loved this but felt that Wolfram tried to control the narrative throughout, I am not sure that he should have got away with been interrogator and not being interrogated more, (he appeared tp. Be reluctant to a round 2.0)he also seemed hostile at times in a dismissive manner and I felt that was unfair.I acknowledge my own bias towards Karl and the FEP. I did appreciate that it extracted information from Karl I have not heard previously.
Wolfram seemed to have a real thirst for perspective🙂
Discreet and discrete
so, how often did you replay the conversation?
Funny. Keith “ the observer “ facial expressions provided delightful insights.
Cool
I wonder if lineier energy migration is a black holes feeding system 🤔 in a way refactoring the waves of vibrational energy and the spitting it out is the result of its own over charges and it has to release the same to keep going. ❤
1:08:00 All things are patterns. The human body replaces every atom every 7 years. What persists is the pattern. Or, spirit.
We need a definition of agency that makes changes in the environment because there are consequences of this kind of action. If a model produces text, and that text doesn't affect changes in the environment, it's not dangerous. It's an arbitrary distinction, but it's also important. It's the difference between a thermometer and a thermostat. A rise in mercury can be considered an action, but it's somehow different from turning on a furnace. I would call the latter an agentic behavior but not the former. Yes, it's both an arbitrary distinction and an important distinction. At the heart of the matter, it's probably a failure of our language.
The more I see of Dr Wolfram, the more I want to buy him a new shirt lol
Benford’s Law may validate the depth and scope of the Free Energy principle.