From: jimruttshow8596
The emergence of mind from the brain is a central question in its field, prompting inquiry into whether it can also arise from an artificial brain [01:23:06]. Joshua Bach defines mind as the software running on the brain [01:49:50]. He views software not as a physical entity with identity, but as a specific physical law [01:58:00]. This law describes what happens when universal components are arranged in a particular way, leading to macro states with causal structures [02:09:00].
The mind, therefore, is a principle or software running on the brain [02:48:00]. Its existence is not due to a one-to-one correspondence with individual atoms in neurons, but rather from a coherent causal structure that emerges from the neurons’ activity [03:02:00]. This structure serves as a lens to describe the collective neuronal activity [03:07:00]. Bach posits that there’s no reason why mind couldn’t arise in other substrates, provided they meet the same functional constraints and can implement the same principles, similar to how software runs on various computer types [03:15:00]. He concludes that if brains are Turing complete, minds can be built on different substrates by implementing the same principles [03:25:00].
Embodiment of Mind
The concept of the human mind being strongly embedded in its bodily substrate is a common argument [03:45:00]. Antonio Damasio, for example, suggests the root of our conscious being lies deep in the brain stem, more connected to the body than the higher brain [03:50:00]. This perspective implies that artificial minds might not be truly analogous to human minds if they aren’t embodied in similar structures [04:02:00].
However, Bach counters this by stating that one’s mind can persist even with changes to its interface with the universe [04:16:00]. For instance, being connected to sensors across a city or immersed in a virtual reality like Minecraft doesn’t mean losing consciousness [04:25:00]. Consciousness can exist even if one’s affordances are limited to internal mental processes, as occurs during dreams [04:45:00]. Therefore, embodiment can be entirely virtual [05:27:00]. What’s essential is the implementation on a physical substrate, not a human-like physical body, unless a human-like mind is desired [05:31:00].
The Role of Emotions and Feelings
Bach’s work emphasizes the importance of emotional states and valences [05:57:00]. He notes that physical reactions of emotions (e.g., heartbeat, respiration) often precede cognitive feelings [06:01:00]. He uses the metaphor of consciousness as a “little monkey” atop an “elephant” [06:40:00]. The “elephant” represents older, emotional, and motivational impulses, while the “monkey” represents younger, analytical systems [07:05:05]. Feelings serve as a means for the elephant’s emotions and impulses to become accessible to the analytic monkey [07:26:00].
Feelings are projected into the body map [08:16:00] because, he suspects, feelings were implemented as an afterthought in evolution and had to be mapped onto existing brain regions [08:40:00]. The perception of love in the heart or anxiety in the solar plexus is a projection, not actual gut involvement in computation [08:45:00]. Notably, paraplegics who lack bodily sensations still report feeling emotions in their bodies [09:02:00].
Emotions also influence risk perception. Successfully navigating a life-threatening situation can temporarily increase one’s sense of competence and lead to greater risk-taking [10:04:00]. This can also be a rescaling of perception, where major existential threats make everyday fears seem trivial [11:17:00]. These emotional adaptations to situations explain why people tend to maintain roughly the same emotional proportions as long as they are existentially safe [12:21:00]. Introducing novelty and frequent environmental changes can help maintain freshness and value new experiences [13:22:00].
Theories of Mind
Brain Frequencies and Oscillations
One prominent theory of mind from twenty years ago posited that mind was essentially a brain-wide set of interlocking frequencies [13:58:00]. Bach suggests that oscillations in the brain are necessary because neurons cannot maintain a constant state and communicate effectively without firing in synchrony [14:32:00]. This synchronization leads to periodic waves of activation that can be detected on an EEG [14:48:00]. However, this signal is a result of synchronization that leads to consciousness, not the cause of consciousness itself [15:11:00].
From a neuroscientific perspective, the binding problem (how neurons communicate) might be solved by individual neurons “tuning in” to different “programs” using frequency encoding, akin to RM or FM encoding [15:25:00]. While this is a viable model for signal transmission and synchrony in the neural cortex [16:14:00], it’s not necessarily the best engineering principle for digital computers, which can use random access memory and high-throughput buses [16:30:00]. Nature’s evolutionary solutions are often vastly different from engineered ones [17:01:00].
Bernard Baars’ Global Workspace Theory
Bernard Baars’ Global Workspace Theory suggests that the representation of consciousness, such as the sensory experience, is broadcast to wide areas of the brain, allowing various functional areas to process that information [17:30:00]. While Baars admits no neurological argument supports it, Bach believes it’s partially derived from intuitions resulting from introspection, which he believes is an underestimated tool in neuroscience [18:01:00].
Bach disagrees that consciousness is distributed across the brain. Instead, he argues that the core feature of consciousness is the ability to remember what one paid attention to [18:31:00]. This integration means information, initially distributed, becomes localized into a common protocol for later access [18:46:00]. This concept contrasts with the notion of attention in machine learning, which is gaining prominence [19:07:00].
Humans learn very rapidly compared to current neural networks, especially in complex, high-dimensional problem spaces like war games [20:15:00]. AlphaZero, for instance, requires hundreds of thousands to millions of plays to learn stereotyped games like chess or Go [21:31:00]. This suggests that human learning mechanisms are fundamentally different from gradient descent [20:21:00]. Bach attributes rapid human learning to the ability to “pay attention in the right way” [22:18:00]. If one doesn’t know what to focus on, the problem must be brute-forced, which is wasteful [22:53:00].
Bach’s early work on language processing in the 1990s, influenced by Ian Witten’s data compression perspective on cognition [23:37:00], encountered the difficulty of finding structure in language without adjacency constraints due to vast statistics [24:45:00]. This is where curriculum learning, like how children are taught simple sentences first, becomes relevant [25:01:00]. The transformer algorithm underlying GPT-3, which makes statistics over all parts in a fixed working memory window (2048 tokens), is a solution to this problem [25:56:00]. However, this fixed window limits its ability to comprehend large images, videos, or relate distant parts of a book [26:38:00].
While GPT-3 is remarkable for its stylistic emulation and plausible completions within small domains [27:10:00], it lacks genuine language understanding [27:38:00]. It’s a collection of associative statistics, but can perform semantic operations like sentiment extraction by predicting text [27:48:00]. Its “understanding” means it can consistently act as if it understands [27:57:00]. However, it lacks a global model of the universe or a global sense of meaning [28:56:00]. GPT-3’s symbols are grounded in language itself, not in the external physical world [30:54:00]. It doesn’t understand the algorithm behind arithmetic, only predicting likely text strings [32:06:00]. This suggests that simple algorithms might not be the right principle to build a mind [33:38:00], and that our attention algorithm might be a “hack” to bypass the combinatorial explosion of options [34:28:00].
Tononi’s Integrated Information Theory (IIT)
Tononi’s Integrated Information Theory (IIT) gained popularity as a theory of how mind emerges from matter [34:46:00]. Bach views IIT as largely defined by its opposition to functionalism, seeking an alternative explanation for consciousness [36:20:00]. He respects Tononi as an autonomous intellect [37:08:00] but is skeptical of the theory’s claims, especially if it offers a “new solution” to an age-old philosophical question [37:18:00]. Bach suspects the core of Tononi’s theory isn’t the quantifiable measure phi, but rather an attempt to reintroduce panpsychism [37:52:00]. It’s ironic to propose an anti-functionalist theory using information theory, as functionalism and information theory are deeply intertwined [38:49:00].
Functionalism
Functionalism, in simple terms, treats a phenomenon as a result of its implementation [39:11:00]. For example, a bank is defined by its functions: allowing accounts, storing and retrieving money, conforming to legal interfaces, and offering guarantees [39:21:00]. If an institution fulfills all these principles, it makes no sense not to call it a bank [39:39:00]. Dennett uses this to reject the concept of a “philosophical zombie”—a system identical in every feature but lacking phenomenal experience [39:51:00]. Functionalism rejects the notion of a hidden essence that lacks observable causal properties [40:37:00]; everything can be explained by its causal properties, which are ultimately implementable functions [40:47:00].
Dennett’s Pandemonium Theory and Philosophical Zombies
Dennett’s Pandemonium Theory, stemming from Selfridge and influencing Minsky’s Society of Mind, metaphorically describes the mind as a collection of agents (demons) implementing different behaviors [41:15:00]. These agents self-organize, adapt to unknown situations by composing teams of behaviors, and interact to produce combined behavior [42:23:00].
Bach notes that many philosophers and students dislike Dennett because he seems to “miss the problem” they try to explain, not giving enough importance to phenomenal experience [42:51:00]. He speculates this might be because Dennett, being a “nerd,” is highly constituted on the conceptual side, potentially lacking intuitive empathy [43:24:00]. Scientists and true philosophers, he suggests, might be “aberrations” who permanently trust ideas more than feelings, often due to social failures in childhood [00:44:00]. This leads them to build rational models for interaction, becoming good at analytical models but poor at intuition [45:13:00]. Science helps with difficult edge cases where intuition fails, but not with most everyday problems [46:21:00].
Chalmers’ Hard Problem of Consciousness
David Chalmers’ “hard problem” of consciousness, which Dennett rejects, concerns how physical systems give rise to subjective conscious experience [50:21:00]. Bach suggests Chalmers sometimes explores explaining why people think there’s a hard problem, rather than explaining the phenomenon itself [50:50:00]. This psychological certainty requires explanation [51:02:00].
Parapsychology and Non-Lawful Universe
Bach explores parapsychological phenomena, such as clairvoyance and out-of-body experiences, noting that many reputable people and even Alan Turing [55:54:00] considered the possibility of telepathy [51:27:00]. However, such phenomena appear incompatible with known physics and the standard model, as they would require forces beyond the four known ones [52:22:00]. Quantum non-locality does not explain the transmission of information needed for psi phenomena [52:36:00]. If psi phenomena were real, it raises questions about whether our consciousness is computed in our own brain or if the brain is merely an antenna for consciousness computed elsewhere [56:47:00].
Bach contrasts the “mechanical universe” hypothesis of physics (everything is causally closed and expressible as a computer program) with the idea that we live in a “dream” [01:02:05]. A dream involves symbolic or “magical” interactions that transcend known physical laws, suggesting a higher level of causation outside basic game dynamics [01:02:45]. He posits that we do live in a dream, one generated by a mind on a different level of existence, which is physics [01:04:06]. What we perceive (colors, sounds, the squishy brain) is the dream, not physics itself [01:04:32].
Penrose and Quantum Consciousness
Roger Penrose, affiliated with theories that go beyond computation for consciousness, interprets Gödel’s proof as showing that human mathematicians can do parts of mathematics that computation cannot [59:22:00]. Bach argues that this implies those “uncomputable” parts of mathematics were never “real” and were ill-defined [59:55:00]. For Bach, constructive mathematics (which is computable) is real, while classical mathematics containing infinities is not implementable in physical causal structures [01:00:11]. Penrose connects the uncomputable parts of mathematics with consciousness, suggesting they go beyond known physics, perhaps linked to quantum gravity, which is the only part not explained in the computational paradigm [01:00:51].
Bach mentions Stuart Hameroff, who builds a “psychedelic sculpture garden” of theory, using “pi resonant quantum underground” mechanisms to explain consciousness, anesthesia, psychedelics, evolution, and emotion [01:01:50]. Bach views this as more art than science due to its lack of resonance in the scientific community [01:00:13].
The Next Step in the Science of Mind
The next decade in the science of mind should focus on attention-based models, with the transformer being just the beginning [01:32:58]. Bach identifies three main issues with current systems like GPT-3:
- Limited Working Memory Window: GPT-3 has a fixed window of 2048 adjacent tokens, unlike the human mind which can construct working memory contents with more degrees of freedom, even if it’s smaller in capacity [01:48:51]. Humans can actively change and construct working memory contexts, recalling and revising past insights [01:17:10].
- Offline Learning: GPT-3 learns offline and stops updating after its training cutoff (e.g., October 2019), leading to a lack of real-time knowledge about current events [01:19:14]. Future agents need to continuously learn and track reality in real time [01:19:40].
- Lack of Relevance: GPT-3 doesn’t inherently care about relevance; its “relevance sensation” comes from being trained on human-written texts that are already deemed relevant by people [01:20:05]. For systems interacting with the world and processing rich sensory data, a motivational system is needed to focus on the most promising parts of the model [01:20:46]. This affective (emotional) component is crucial, guiding a system through homeostasis, where needs produce pain signals when frustrated and pleasure when satisfied [01:21:34]. This hierarchy of purposes, striving for coherence, forms the “soul” of the agent [01:22:26].
The property of “realness” is not a feature of physical reality itself but a model property [01:09:12]. It’s a label the mind attaches to parameters, distinguishing imagination from the experienced world based on predictability of future sensory patterns [01:09:30]. An enlightened relationship to reality requires recognizing that perception, including self-perception, is a representation, not an immediate reality [01:11:42].