Joscha Bach



 (born 1973 in Weimar, Germany) is cognitive scientist focusing on cognitive architectures, models of mental representation, emotion, motivation and sociality. Achievements include research in novel data compression algorithm using concurrent entropy models; development of microPsi cognitive architecture for modeling emotion, motivation, mental representation.

In 2000, Bach graduated with a diploma in Computer Science from Berlin, followed by a Doctor of Philosophy at Osnabrück University, Germany, in 2006. Before joining, he worked as a visiting researcher at the and the Harvard Program for Evolutionary Dynamics.

Fact finding reports by the and  found that Bach’s research was supported with more than $150,000 by the  Foundation.

Seven Principles of Synthetic Intelligence (2008)

 * When the Artificial Intelligence (AI) movement set off fifty years ago, it bristled with ideas and optimism, which have arguably both waned since. The field has regressed into a multitude of relatively well insulated domains like logics, neural learning, case based reasoning, artificial life, robotics, agent technologies, semantic web... each with their own goals and methodologies. The decline of the idea of studying intelligence per se, as opposed to designing systems that perform tasks that would require some measure of intelligence in humans, has progressed to such a degree that we must now rename the original AI idea into Artificial General Intelligence.


 * Attempts in psychology at overarching theories of the mind have been all but shattered by the influence of behaviorism, and where cognitive psychology has sprung up in its tracks, it rarely acknowledges that there is something as "intelligence per se", as opposed to the individual performance of a group of subjects in an isolated set of experiments.


 * AI’s gradual demotion from a science of the mind to the nerdy playpen of engineering was accompanied not by utterances of disappointment, but by a chorus of glee, uniting those wary of human technological hubris with the same factions of society that used to oppose evolutionary theory or materialistic monism...


 * Long ago, physics and other natural sciences... had become computational.


 * MicroPsi is a cognitive model that represents the author’s attempt to contribute to the discussion of Artificial General Intelligence (AGI)...


 * [W]e are in need of functionalist architectures. That is, we need to make explicit what entities we are going to research, what constitutes these entities conceptually, and how we may capture these concepts.


 * Early AI systems tended to constrain themselves to micro-domains that could be sufficiently described using simple ontologies and binary predicate logics, or restricted themselves to hand-coded ontologies altogether. ...AI systems will probably have to be perceptual symbol systems, as opposed to amodal symbol systems...


 * For all practical purposes, the universe is a pattern generator, and the mind "makes sense" of these patterns by encoding them according to the regularities it can find. Thus, the representation of a concept in an intelligent system is not a pointer to a "thing in reality", but a set of hierarchical constraints over (for instance perceptual) data.


 * [T]he quality of a world model eventually does not amount to how "truly" it depicts "reality", but how adequately it encodes the (sensory) patterns.


 * Robots are... not going to be the singular route to achieving AGI, and successfully building robots that are performing well in a physical environment does not necessarily engender the solution of the problems of AGI. Whether robotics or virtual agents will be first to succeed in the quest of achieving AGI remains  an open question.


 * General intelligence is not only the ability to reach a given goal (and usually, there is some very specialized, but non-intelligent way to reach a singular fixed goal, such as winning a game of chess), but includes the setting of novel goals, and most important of all, about exploration. ...[A]n  environment with fixed tasks, scaled by an agent with pre-defined goals is not going to make a good benchmark problem for AGI.


 * [M]otivation... does not arise from intelligence itself, but from a motivational system underlying all directed behavior. ...[T]here is no reason that could let us take behavioral tendencies such as self-preservation, energy conservation, altruistic behavior for granted—they... have... to be designed [including by evolutionary methods] into the system...


 * Because a naked ontological dualism between mind and body/world is notoriously hard to defend, it is sometimes covered up by wedging the popular notion of emergence into the "explanatory gap"... "strong emergence" is basically an anti-AI proposal.


 * In cognitive science, we currently have two major families of architectures... One, the classical school... characterized as Fodorian Architectures, as... the manipulation of a language of thought, usually expressed as a set of rules and capable of . ...The other family favors distributed approaches and constrains a dynamic system with potentially astronomically many until... behaviors [of] general intelligence are left. This may seem more "natural" and well-tuned... Yet many functional aspects of intelligence... as planning and language, are... much  harder to depict using  the dynamical systems approach.


 * MicroPsi is an implementation of ’s Psi theory... MicroPsi is an attempt to embody the principles discussed above...

Principles of Synthetic Intelligence PSI (2009)

 * : An Architecture of Motivated Cognition


 * Symbolic reasoning falls short not only in modeling low level behaviors but is also difficult to ground into real world interactions and to scale upon dynamic environments... This has lead many... to abandon symbolic systems... and... focus on parallel distributed, entirely sub-symbolic approaches... well suited  for many learning and control tasks, but difficult to apply [in] areas such as reasoning and language.


 * [D]espite inevitable difficulties and methodological problems, the design of unified architectures modeling the breadth of mental capabilities in a single system is a crucial stage in understanding the human mind...


 * [T]he designer of a unified architecture is in a similar situation as the cartographers that set out to draw the first maps of the world, based on the reports of traders and explorers returning from expeditions  into uncharted waters and journeys to unknown  coasts.


 * [T]aking [the design of a cognitive architecture] to the AI laboratory... requires the theory not merely to be plausible, but... requires it to be fit for implementation, and delivers it to the... merciless battle of testing.


 * This book is an attempt to explain cognition—thought, perception, emotion, experience—in terms of a machine, that is, using a cognitive architecture.


 * Cognitive architectures are... Leibnizian machines... designed to bring forth the feats of cognition, and built to allow us to enter... examine them, and to watch their individual parts... pushing and pulling... thereby explaining how a mind works.


 * Our... cognitive architecture is based on a formal [psychological] theory... the PSI theory... [which] has been turned into a computational model... MicroPSI... partially implemented as a computer program.


 * [T]he relationship between cognition and neurobiological processes might be similar to the one between a car engine and locomotion. ...[A] car's locomotion is facilitated mainly by its engine, but the understanding of the engine does not aid much in finding out where the car goes. ...[T]he integration of... parts, the intentions of the driver and even the terrain might be more crucial ...


 * Because there is no narrow, concise understanding of what constitutes mental activity and what is part of mental processes... cognition, the cognitive sciences and the related notions span a wide and convoluted terrain... most of [which] lies outside psychology... This methodological discrepancy can only be understood in the context of the recent.


 * Behaviorism... in the form of ... not only neglected the nature of mental entities as an object of inquiry, but denied their existence...


 * [N]egligence of internal states of the mind makes it difficult to form conclusive theories of cognition, especially with regard to language... and consciousness, so radical behaviorism... lost its foothold. Yet, methodological behaviorism is still prevalent...


 * Unlike physics, where previously unknown entities and mechanisms... are routinely postulated... and... evidence is sought in favor or against these... psychology shuns [this methodology]... Thus... cognitive psychology shows reluctance... to building unified theories of mental processes. ...Piaget's work ...might be one of the notable exceptions ...


 * [I]nfluences that lead to the study of... mental entities and structures... came from... s and cybernetics, and from formal linguistics. They fostered an understanding that mental activity amounts to ... that can be modeled as... an algorithm—working over states that encode representations.


 * Functional constructivism is based on... philosophical constructivism... that all knowledge about the world is based on... our systematic interface. ...We do not ...recognize ...objects of our environment; we construct them over the regularities ...at the system interface of our cognitive system.


 * To perceive means... to find order over patterns; these orderings are what we call objects. ...[I]t amounts to ...identification of these objects by their related patterns ...intuitively ...its features ...


 * Everything we know about ourselves is... ordering... over features available at the interface; we know of mental phenomena only insofar as they are patterns or constructed over patterns.


 * What the universe makes visible... to any observer... is... functionality.


 * [T]he notions we process... are systematically structured information, making up a dynamic system. The description of such... is the domain of cybernetics or systems science... constructive methods that allow the representation of functionality.


 * To understand... mind, we have to ask how a system capable of constructing has to be built...


 * [A] functionalist model of cognitive processes might be implemented as a computer program (computationalism)... hand in hand with... Artificial Intelligence.


 * According to [the  hypothesis]... an implemented , has the necessary and sufficient means for general intelligent action. ...[A]ny system that exhibits general intelligence will ...be a physical symbol system. ...[A]ny physical symbol system of sufficient size can be organized further to exhibit general intelligence."
 * Citing "Computer Science as Empirical Inquiry: Symbols and Search", Allen Newell & Herbert A. Simon, 1976, p. 116.]


 * The goal of building cognitive architectures is to achieve an understanding of mental processes by constructing testable information processing models.


 * Functionalist psychology is... compatible with... scientific positivism, because it makes emperically falsifiable predictions... The... model is capable of producing [or predicting] specific behavior [and] [t]he model is the sparsest, simplest one...


 * Just as the extensive theoretical bodies of physics, chemistry, [etc.]... unified theories of cognition are not isolated statements discarded when... predictions [are] refuted. Rather they are paradigms... that direct a research program...

Joscha: Computational Meta-Psychology (2015)

 * An Artificial Intelligence exploration into the creation of meaning, source.


 * What's the best algorithm that you should be using to fix your world model? ...This question ...has been answered for the first time by in the 1960s. He discovered an algorithm that you can apply when you've  discovered that you're a robot and all you've got is data. What is the world like? ...[H]is algorithm is... a combination of Bayesian Reasoning, Induction and Occam's Razor. ...[W]e can mathematically prove that we cannot do better than Solomonoff Induction. Unfortunately, Solomonoff Induction is not quite computable.


 * [E]verything that we're going to do is some approximation of Solomonoff Induction. ...[O]ur concepts cannot really refer to facts in the world out there. We do not get the truth by referring to stuff out there in the world. We get meaning by suitably encoding the patterns in our systemic interface.


 * AI has recently made huge progress in encoding data at perceptual interfaces. is about using a stacked hierarchy of feature detectors. ...[W]e use pattern detectors and we build them into networks that are arranged in hundreds of layers and then we adjust the links between these layers, usually using some kind of . ...[Y]ou can use this to classify [e.g.,] images and parts of speech. ...[W]e get to features that are more and more complex. They start with these very... simple patterns, and then get more and more complex until we get to object categories. ...[N]ow the systems are able, in image recognition tasks, to approach performance that's very similar to human performance. ...[I]t seems to be somewhat similar to what the brain seems to be doing in visual processing.


 * If you take the activation at different levels of these networks and you... enhance this activation a little... you get stuff that looks very psychadelic, which might be similar to what happens if you put certain illegal substances into people and enhance the activity on certain layers of their visual processing.


 * [O]ur best bet is not just to have a single classification with filtering. ...[I]nstead... take the low level of input and get a whole universe of features that is interrelated. ...[W]e have different levels of determinations. At the lowest level we have percepts. At a slightly higher level we have simulations, and on an even higher level we have a concept landscape.


 * How does... representation by simulation work? ...If you are a brain and you want to understand sound, you have to model it. ...Neurons do not want to do 20 Khz. That's way too fast for them. They like something like 20 Hz. So... you need to make a [which] measures the amount of energy at different frequencies. ...This  ...in our ears ...transforms energy of sound at different frequency intervals into energy measurements... This is something that the brain can model. ...[A] neurosimulator tries to recreate these patterns. If it can predict the next input from the cochlea, then it understands the sound.


 * If we want to understand music we have to go beyond understanding sound. We have to understand the transformations that sound can have if you play a different pitch. We have to arrange the sound in a sequencer that gives you rhythms, and so on, and then we want to identify some kind of musical grammar that we can use to... control the sequencer. So we have stacked structures that simulate the world. ...If you want to model a world of music you need to have the lowest level of the precepts, then the higher levels of mental simulations, which give the sequences... and the grammars of music... [B]eyond this you have the conceptual landscape that you can use to describe the different styles of music. ...[I]f you go up in the hierarchy, you get to more and more abstract models, more and more conceptual models, and more and more analytic models. ...[T]hese are causal models...


 * [C]ausal models can be weakly deterministic, basically associative models, which tell you if this state [S1] happens, it is quite probable that this one [S2] comes afterwards. Or you can get to a strongly determined model... one which tells you, if you are in this state [S1], and this condition [c1] is met, you're going to go exactly in this state [S2]. If this state is not met, or a different condition [c2] is met, you go into this state [S3]. And this is what we call an algorithm. Now you're in the domain of computation.


 * For a long time people have thought that the universe is written in mathematics... In fact nothing is mathematical. Mathematics is just the domain of formal languages. It doesn't exist. Mathematics starts with a void. Just throw in a few axioms and if those are nice axioms, then you get infinite complexity. Most of it is not computable. In mathematics you can express arbitrary statements, because it's all about formal languages. Many of these statements will not make sense. Many of these statements will make sense in some way, but you cannot test whether they make sense because they're not computable.


 * Computation is different. Computation can exist. It starts with an initial state, and then you have a transition function. You do the work. You apply the transition function [and] you get into the next state. Computation is always finite.


 * Mathematics is the kingdom of specification and computation is the kingdom of implementation. It's very important to understand this difference.


 * All our access to mathematics... is because we do computation. We can understand mathematics because our brain can compute some part of mathematics, very very little of it and to a very constrained complexity, but enough so we can map some of the infinite complexity and noncomputability of mathematics into computational patterns which we can explore.


 * [C]omputation is about doing the work... executing a transition function.


 * [W]e saw that mental representation is about percepts, mental simulations, conceptual representations... [C]onceptual representations give us concept spaces, and... these concept spaces... give us an interface for our mental representations we can use to address and manipulate them, and we can share them in cultures. [T]hese concepts are compositional. We can put them together to create new concepts. ...[T]hey can be described using higher dimensional vector spaces. They [vectors] don't do mental simulation and prediction, and so on, but we can capture regularity in our concepts with them.


 * With this vector space you can do amazing things, [e.g.,] if you take the vector from king to queen, it's pretty much the same vector as between man and woman. ...[B]ecause [these concept spaces are] really a high dimension manifold, we can do interesting things like machine translation without understanding what it means, that is, without doing any proper mental representation that predicts the world. ...[T]his is [a] type of mental representation that is somewhat incomplete, but it captures the landscape that we share in a culture.


 * [T]here is another type of mental representation that is linguistic protocols, which is... a form of grammar and a vocabulary. ...[W]e need these ...protocols to transfer mental representations between people ...by scannning our ...representations, disassembling them ...and ...we use a discrete set of symbols to get this to somebody else... [who] trains an assembler that reverses this process and builds something that is... similar to what we intended to convey.


 * [I]f you look at the progression of AI models, it... went the opposite direction. ...AI started with linguistic protocols, which were expressed in formal grammars, and then it got to concept spaces, and now it's about to address percepts. ...At some point in the near future it's going to get better at mental simulations and at some point after that we'll get to attention directed and motivationally connected systems that make sense of the world, that are in some sense able to address meaning. This is the hardware that we have...


 * How difficult is it to define a brain? We know that the brain must be somewhere hidden in the genome [which] fits in a CD-ROM. It's not that complicated. It's easier than Microsoft Windows. ...[A]bout 2% of the genome is coding for proteins, and maybe about 10%... tells you when to express which protein, and the remainder is mostly garbage. It's old viruses that are left over and it's never been properly deleted [etc.] because there are no real code revisions in the genome. ...How much of this 10%, [i.e.,] about 75 megabytes code for the brain, we don't really know. What we do know is that we share almost all of this with mice. Genetically speaking, a human is a pretty big mouse, with a few bits changed to fix some of the genetic expressions. ...Most of the stuff there is going to code for cells and metabolism and what your body looks like, [etc.]...


 * [T]o encode a brain genetically, based on the hardware that we are using, we need something like at least 500 kilobytes of code... actually... it's going to be a little more, I guess. It sounds like surprisingly little... but in terms of scientific theories this is a lot. ...The universe, according to the core theory of quantum mechanics... it's like half a page of code... to generate the universe. ...[I]f you want to understand evolution, it's like a paragraph... a couple lines, really, to understand an evolutionary process. ...[T]here's lots ...of details that you get afterwards, because this process itself doesn't define what all the animals are going to look like. In a similar way, the code of the universe doesn't tell you what this planet is going to look like and you... are going to look like. It's just defining the rule book.


 * [T]he genome defines the rule book by which our brain is built. The brain boots itself, in a development process, and this booting takes some time... formation learning in which some connections are formed, basic models are built of the world so we can operate in it. How long does this booting take... about 80 megaseconds. That's the time a child is awake until it's 3 1/2 years old. By this age you understand Star Wars, and I think everything after Star Wars is cosmetics.

The Cortical Conductor Theory (2017)

 * : Towards Addressing Consciousness in AI Models


 * Mathematics is the domain of all formal languages, and allows the expression of arbitrary statements (most of which are uncomputable). Computation may be understood in terms of computational systems, for instance via defining states (which are sets of discernible differences, i.e. bits), and transition functions that let us derive new states.


 * Whereas mathematics is the realm of specification, computation is the realm of implementation; it captures all those systems that can actually be realized.


 * Computational systems are machines that can be described apriori and systematically, and implemented on every substrate that elicits the causal properties that are necessary to capture the respective states and transition functions.


 * The absence of an understanding of substrate independent machines lead Leibniz to the rejection of mechanist philosophy...


 * [W]hile dynamical systems often cannot be effectively computed on a finite state machine (such as a von Neumann computer), they can often be efficiently approximated.


 * Artificial Intelligence was the attempt of thinkers like Marvin Minsky, John McCarthy and others to treat the mind as a computational system, and thereby open its study to experimental exploration by building computational machines that would attempt to replicate the functionality of minds.


 * The failure to deliver on some of the early, optimistic promises of machine intelligence, as well as cultural opposition, lead to cuts in funding for cognitive AI, and eventually the start of the new discipline of Cognitive Science. ...Cognitive Science did not develop a cohesive methodology and theoretical outlook, and became an umbrella term for neuroscience, AI, cognitive psychology, linguistics and philosophy of mind.

Joscha Bach: Artificial Consciousness and the Nature of Reality (June 13, 2020)

 * AI Podcast #101 with Lex Fridman, source


 * At some point you have to understand the comedy of your own situation. If you take yourself seriously, and you are not functional, it ends in tragedy, as it did for Nietzsche. ...[Y]ou find the same thing in Hesse... The Steppenwolf syndrome is classic in all its sense, where you... feel misunderstood by the world and you don't understand that all the misunderstandings are the result of your own lack of self-awareness, because you think that you are [the] prototypical human and the others around you should behave the same way as you expect... based on your innate instincts; and it doesn't work out, and you become a transcendentalist to deal with that. ...It's very... understandable and I have great sympathies for this, to the degree that I can have sympathy for my own intellectual history. But you have to grow out of it.


 * You need to become unimportant as a subject, that is, if you are a philosopher, believe is not a verb. ...You have to submit to the things that are possibly true and... follow wherever your inquiry leads, but it's not about you, it has nothing to do with you.


 * You cannot define objective truth without understanding the nature of truth... So what does the brain mean by saying that it's discovered something as truth... A model can be predictive or not... [T]here can be a sense in which a mathematical statement is true because it's defined as true under certain conditions. So it's... a particular state that a variable can have in the assembled game and then you can have a correspondence between systems and talk about truth, which is again a type of model correspondence.


 * [T]here also seems to be a particular kind of ground truth, [e.g.,] you are confronted with the enormity of something existing at all. ...It's stunning when you realize something exists, rather than nothing. ...[T]his seems to be true. There is an absolute truth in the fact that something seems to be happening.


 * The easiest answer is existence is the default. ...So this is the lowest number of bits that you need to encode this. ...Nonexistence might not be a meaningful notion. ...If everything that can exist, exists... it probably needs to be implementable. The only thing that can be implemented is finite automata so maybe the whole of existence is... a superposition of all finite automata, and we are in some region of the fractal that has the properties that it can contain us. ...Imagine that every automaton is... an operator that acts on some substrate [something that can store information], and as a result you get emergent patterns.


 * It may not have a why. This might be the wrong direction to ask... [T]here could be no relation in the "why" direction... It doesn't mean that everything has to have a purpose or a cause...


 * The last big things that we discovered was the constructivist turn in mathematics... to understand that the parts of mathematics that work are computation. That was a very significant discovery in the first half of the 20th century. ...[I]t hasn't fully permeated philosophy and even physics yet. Physicists checked out the code libraries for mathematics before constructivism became universal. ...Gödel himself ...didn't get it yet. Hilbert could get it. Hilbert saw that [e.g.,] Cantor's set theoretic experiments in mathematics led him to contradictions, and he noticed that with the current semantics we cannot build a computer in mathematics that runs mathematics without crashing, and Gödel... could prove this.


 * What Gödel could show is using classical mathematical semantics you run into contradictions, and because Gödel strongly believed in these semantics... he was shocked. It... shook his world to the core, because in some sense he felt that the world has to be implemented in classical mathematics.


 * For Turing it wasn't quite so bad. ...[T]uring could see that the solution is to understand that mathematics was computational all along. ...For instance π|pi in classical mathematics is a value. It's also a function, but it's the same thing. In computation, a function is only a value when you can compute it, and if you cannot compute the last digit of pi, you only have a function. You can plug this function into your local sun, let it run until the sun burns out... This is it. This is the last digit of pi you will know. But it also means that there can be no process in the physical universe, or in any physically realized computer that depends on having known the last digit of pi. ...Which means that there are parts of physics that are defined in such a way that cannot strictly be true, because, assuming that this could be true leads into contradictions.


 * I don't actually have an identity beyond the identity that I construct. ...[T]he Dalai Lama... identifies as a form of government. [He] gets reborn, not because he is confused, but because he is not identifying as a human being. He runs on a human being. He's... a governmental software that is instantiated in every new generation anew. So his advisors will pick someone who does this in the next generation. So if you identify with this, you are no longer human and you don't die in this sense... only the body that you run on. To kill the Dalai Lama you'd have to kill his tradition.


 * ... Facebook... Twitter... are companies that... own a protocol... imposed on a community and... [and have] different components for monetization... user management... user display... rating... anonymity... for import of other content... Imagine that you take these components of the protocol apart and... communities are allowed to mix and match their protocols, and design new ones... [e.g.,] the UI and the UX can be defined by the community, the rules for sharing content across communities can be defined, the monetization... the way you reward individual users... the way users represent themselves... can become part of the protocol... [I]n some communities it will be a single person that comes up with these things; in others it's a group of friends. Some might implement a voting scheme... Who knows what might be the best self-organizing principle for this. ...It can be automated so people can write software for this. ...Let's not make an assumption about this thing if you don't know the right solution... In most areas there is no idea whether it will be people designing this ad-hoc, or machines doing this. Whether you want to enforce compliance by social norms, like, or with software solutions, or with AI that goes through the post-op people, or with a legal principle... If you let the communities evolve, and you just control it in such a way that you are incentivizing the most sentient communities. The ones that produce the most interesting behaviors, that allow you to interact in the most helpful ways to the individuals. ...So that you have a network that gives... information that is relevant to you. It helps you to maintain relationships to others in healthy ways. It allows you to build teams... to... bring the best of you into this thing and goes into the coupling, into a relationship with others in which you produce things that you would be unable to produce alone.


 * Minds are software states... Software doesn't have identity. Software in some sense is a physical law. ...The maintenance of the identity is not terminal. It's instrumental to something else. You maintain your identity so you can serve your meaning. So you can do the things that you are supposed to do before you die. ...For most people the fear of death is the fear of dying before they are done with the things that they feel they have to do even though they cannot quite put their finger on... what that is.


 * The fuzzy idea is the one of continuous existence. We don't have continuous existence... because it's not computable. There is no continuous process. The only thing that binds you together with the last week and from yesterday is the illusion that you have memories about them. So if you want to upload, it is very easy. You make a machine that thinks it's you. ...It's the same thing that you are. You are a machine that thinks it's you.


 * [Y]ou don't know this state in which you don't have a self. You can turn off yourself... You can... meditate yourself [into] a state where you are still conscious, where still things are happening, where you know everything that you knew before, but you're no longer identified with changing anything. ...[T]his means that your self ...dissolves. There is no longer this person... you know that this person construct exists in other states and it runs on this brain... but it's not a real thing. It's a construct. It's an idea... and you can change that idea, and if you let go of this idea... If you don't think you are special, you realize it's just one of many people, and it's not your favorite person even... It's just one of many, and it's the one that you are doomed to control... and that is... informing the actions of this organism as a control model. This is all there is, and you are somehow afraid that this control model gets interrupted, or loses the identity of continuity.


 * [M]editation is... just a bunch of techniques that let you control attention. ...[W]hen you can control attention you can get access to your own source code, hopefully not before you understand what you are doing, and then you can change the way it works, temporarily or permanently. ...Everything else is downstream from controlling attention.


 * Normally we only get attention in the parts of our mind that create heat, where you have a mismatch between [the] model and the results that are happening. So most people are not self-aware, because their control is too good. If everything works out roughly the way you want, and the only things that don't work out are whether your football team wins, then you will mostly have models about these domains. ...It's only when... your fundamental relationships through the world don't work [that, attention or self-awareness arises].


 * [T]he types of models that we form right now are not sparse enough... which means that, ideally, every potential model state should correspond to a potential world state. So... if you vary states in your model, you always end up with valid world states. ...[O]ur mind is not quite there... an indication is especially what we see in dreams. The older we get, the more boring our dreams become, because we incorporate more and more constraints that we learned about how the world works. So many of the things that we imagine to be possible as children turn out to be constrained by physical and social dynamics, and as a result fewer and fewer things remain possible. It's not because our imagination scales back, but the constraints under which it operates become tighter and tighter. ...So the constraints under which our neural networks operate are almost limitless, which means it's very difficult to get a neural network to imagine things that look real.


 * We probably need to build dreaming systems... [P]art of the purpose of dreams is... similar to a... generative adversarial network. We learn certain constraints, and then it produces alternative perspectives on the same set of constraints, so you can recognize it under different circumstances. Maybe we have flying dreams as children because we recreate the objects that we know, the maps that we know, from different perspectives, which also means from the bird's-eye perspective.


 * It's relatively easy to build a neural network that... learns the dynamics. The fact that we haven't done it right so far doesn't mean it's hard... [A] biological organism does it with relatively few neurons. ...[Y]ou build a bunch of neural oscillators that entrain themselves with the dynamics of your body is such a way that the regulator becomes isomorphic and it's modeled with the dynamics that it regulates, and then it's automatic. ...[I]t's only interesting in the sense that it captures attention when the system is off [kelter].


 * How much common sense knowledge do we actually have. Imagine that you are a really hard working all your life and you form two new concepts every half-hour... You end up with... a million concepts, because you don't get that old. ...That's not a lot. ...[H]ow many cycles do your neurons have in your life, it's quite limited.


 * [I] think of the concepts as the address space for our behavior programs. The behavior programs allow us to recognize objects [also mental objects] and react... [A] large part of that is the physical world that we interact with, which is this ' thing... basically the navigation of information in space... [I]t's similar to a ... a physics engine that you can use to describe/predict how things that look in a particular way, that feel... a particular way, enough, enough auditory perception... the geometry of all these things... [T]his is probably 80% of what our brain is doing... dealing with that... real time simulation... [I]t's not that hard to understand... [O]ur game engines are already approximating the fidelity of what we can perceive... in the same ball park... just a couple of orders of magnitude away from saturating our perception, from the complexity that [the brain] can produce. ...[T]he computer that you can buy... is able to give a perceptual reality that has the detail that is already in the same ball park as what your brain can process.'''


 * [E]verything else are ideas about the world, and I suspect that they are relatively sparse, and also the intuitive models that we form about social interaction. ...[T]he priors are present in most social animals ... Many domestic social animals ...have better social cognition than children.


 * There are some animals like elephants that have larger brains than us and they don't seem to be smarter. ...Elephants seem to be autistic. They have very, very good motor control and they're really good with details, but really struggle to see the big picture. ...[Y]ou can make them recreate drawings stroke by stroke... but they cannot reproduce a still life... of a scene... Why is that? Maybe smarter elephants would meditate themselves out of existence because their brains are too large. So... that elephants that were not autistic, they didn't reproduce.


 * Arguably most people are not generally intelligent, because they don't have to solve problems that make them generally intelligent. ...[I]t's not yet clear if we are smart enough to build AI and understand our own nature to this degree... [I]t could be a matter of capacity, and for most people it's... a matter of interest. They don't see the point because the benefits of attempting this project are marginal, because you're probably not going to succeed... and the cost... requires complete dedication of your entire life.

Quotes about Bach

 * Joscha Bach is one of those rare people whose primary motivation is unbound curiosity and inspiration. He clearly loves what he does and you can’t help but notice his radiating passion and youthful exuberance. Joscha has an impressively wide and deep knowledge in a variety of scientific, philosophical and artistic disciplines and I had to do my best just to keep up with Bach’s brilliant fast-paced mind and stream of consciousness.
 * Nikola Danaylov, "Joscha Bach: We need to understand the nature of AI to understand who we are" (Nov 14, 2018) Introduction to Podcast interview.


 * I have a lot of good things to say about Joscha Bach. He has a deep conceptual understanding of AGI and its grounding in cognitive science, philosophy of mind and neuroscience; a software system embodying a subset of his ideas and doing interesting and useful things; and a concrete plan for what needs to be learned to expand his system and his associated theory gradually toward human-level intelligence. What's more... his  approach  might  work, if extended and grown intelligently from its current state. There aren’t many AGI researchers I can say all that about.
 * , "Joscha Bach: Understanding the Mind" (May, 2011) as quoted in Between Ape and Artilect (2013)


 * Joscha Bach has written a blog post criticizing my suggestion that the universe as a whole is, in a sense, akin to a cosmic nervous system. ...Bach and I have engaged... in an extensive email exchange discussing precisely the points he brought up in his blog post. ...[H]e chose to make his criticisms public now, whilst ignoring the many clarifications I sent him by email back then. ...Bach's criticisms are straw-men; every single one of them.
 * Bernardo Kastrup, "The 'cosmic nervous system': response to Joscha Bach" (Jan 26, 2013)