From: jimruttshow8596
Defining Mind
Joshua Bach, Vice President of Research at the AI Foundation, defines the mind as the “software that runs on the brain” [01:50:00]. He clarifies that software isn’t a physical thing with an identity, but rather a “very specific physical law” [02:05:00] that describes what happens when components are arranged in a particular way, leading to observable macro states with a causal structure [02:09:00].
Can Mind Arise from an Artificial Brain?
Bach asserts that mind can indeed arise from an artificial brain [01:31:00]. If the brain is Turing-complete, then minds can be built on different substrates, provided the same class of principles are implemented [03:22:00].
Embodiment and the Mind
While some argue that the human mind, and consciousness, is deeply embedded in its bodily substrate [03:45:00], Bach suggests that embodiment can be entirely virtual [05:27:00]. He posits that changing sensory interfaces (e.g., being connected to city sensors, like a “city” [04:25:00]) or limiting affordances (e.g., during dreams or playing Minecraft in VR) [04:45:00] does not mean a loss of consciousness [04:51:00]. The implementation needs a realization on the physical substrate layer, but not necessarily a human-like body [05:31:00].
Emotions and Feelings
Feelings are a way for the analytical “monkey” part of our mind to access the emotions and motivational impulses of the older, “elephant” part [07:27:00]. These feelings are often projected into the body map for communication between brain subsystems [08:15:00]. This projection, such as feeling love in the heart or anxiety in the solar plexus, is a mapping into existing brain regions [08:45:00]. Even paraplegics still experience emotions in their body, suggesting it’s a projection rather than direct involvement of physical organs [09:02:00].
Experiences like walking across a dangerous bridge can influence risk-taking behavior, potentially by leading to a higher sense of competence or by rescaling the perception of what is important [10:04:00], [11:17:00]. This adaptation of emotional range to the agent’s situation results in a normalization of risk perception [12:24:00], [12:52:00].
Theories of Mind
Brain Oscillations
The theory that consciousness is a specific brain frequency lacks functional sense, as it doesn’t explain how such a frequency is necessary and sufficient to produce the phenomenon [14:14:00]. Brain oscillations are likely a result of synchronization, necessary for neurons to fire in sync and for waves of activation to pass through the brain [15:11:00]. This is a viable model for how the brain solves signal transmission and synchrony over distances, possibly using something akin to FM encoding for neurons to tune into different computations [16:00:00]. However, this biological mechanism might not be the most efficient engineering principle for digital computers that use random access and high-throughput buses [16:30:00].
Global Workspace Theory
Bernard Baars’s Global Workspace Theory, which suggests that the “sensorium” of consciousness is broadcast to wide areas of the brain for processing, is acknowledged for its insights derived from introspection [17:30:00], [00:18:01]. However, Bach argues that it is the localization of previously distributed information into a common, accessible protocol that is the core feature of consciousness, enabling memory of what was attended to [18:43:00]. This contrasts with attention mechanisms in machine learning that often work by pulling from different regions [19:07:00].
Integrated Information Theory (IIT)
Bach views IIT as a theory largely “opposed to functionalism” [36:20:00] and believes its popularity within the physics community might be partly political [36:32:00]. He suggests that the core of Tononi’s theory is not the quantifiable measure of phi, but rather an attempt to reintroduce panpsychism, likely because Tononi doesn’t see how functionalism can solve the problem of consciousness [38:14:00]. Bach finds it ironic that IIT uses information theory (strongly intertwined with functionalism) to propose an anti-functionalist theory [38:55:00].
Functionalism Explained
Functionalism treats a phenomenon as the result of its implementation [39:11:00]. For example, a bank is defined by its functions (accounts, storing/retrieving money, legal compliance) rather than an inherent “bankness” [39:21:00]. This perspective leads to the rejection of concepts like philosophical zombies (systems identical in all features except phenomenal experience), arguing there’s no “hidden essence” without observable causal properties [40:10:00], [40:39:00]. All causal properties, from a functionalist view, are ultimately functions that are computable and realizable in a physical system [40:50:00], [40:54:00].
Artificial Intelligence and Learning
Human vs. AI Learning
Humans are remarkably efficient learners, often using transfer learning from analogous experiences, requiring a tiny number of examples compared to current Artificial Intelligence [20:51:00], [21:11:00]. Even advanced systems like AlphaZero require hundreds of thousands to millions of plays to learn stereotyped games like chess or Go [21:31:00]. This suggests human learning mechanisms are fundamentally different from gradient descent and similar methods [20:21:00]. The ability to “pay attention in the right way” is key to rapid human learning [22:18:00].
GPT-3: Capabilities and Limitations
GPT-3, based on the transformer algorithm, made a significant leap by processing statistics over all parts of a working memory to find relationships, a problem that was difficult to solve previously [26:01:00].
While GPT-3 is “both remarkable and frustrating” [27:10:00], capable of style emulation and plausible text completions within short frames [27:17:00], it has significant limitations:
- No Global Understanding: It lacks a “global model of the universe” or a “global sense of meaning” [28:56:00]. Its symbols are grounded in language, not reality [30:54:00].
- Limited Arithmetic: While it can do two-digit arithmetic reliably, it fails at three digits because its training data (human texts) doesn’t heavily feature complex arithmetic [31:46:00]. It doesn’t understand the underlying algorithms [32:06:00].
- Fixed Working Memory Window: GPT-3 is limited to 2048 adjacent tokens in its working memory, unable to actively change or construct working memory contexts like the human mind [18:49:00], [1:17:47:00].
- Offline Learning: It is an “industrial production system” that does not continuously learn, having stopped learning in October 2019, meaning it has no knowledge of events like COVID-19 or George Floyd [1:19:14:00].
- Lack of Relevance: GPT-3 doesn’t intrinsically care about relevance. Its apparent understanding of relevance stems from being trained on texts that humans found relevant enough to write [1:20:05:00]. For systems interacting with the world, a motivational system is needed to focus on promising parts of the model [1:20:50:00].
Future Directions for AI
For the next step in explaining mind emerging from matter and advancing AGI, attention-based models are crucial [1:13:30:00]. Future improvements include:
- Larger Attentional Window: Extending attention to allow active construction and changes in working memory context [1:17:47:00].
- Online Learning: Building agents that can continuously learn and track reality in real-time [1:19:40:00].
- Relevance/Motivational System: Implementing a motivational system that assigns relevance to learning and meta-learning, which current AI systems lack [1:20:50:00].
- Multi-Modal Representation: Moving from language-only representation to a multi-modal representation that is agnostic to what it represents [1:18:00:00].
- Agent Integration: Integrating components like vision-to-text and text-to-motor action modules to create a fixed context for the AI, allowing it to interpret camera images and tell stories about its interactions with the world [1:15:53:00].
The Symbol Grounding Problem
GPT-3’s understanding is primarily “grounded in language” [30:54:00], creating a self-referential loop within language rather than grounding in physical reality. This points to the famous symbol grounding problem [30:49:00].
Affective Part in AI
The “Psi theory” (Bach’s own work) posits that an agent can be described using homeostasis as a guiding principle [1:21:42:00]. Needs, when frustrated, produce pain signals and when satisfied, pleasure signals [1:22:07:00]. Acting on these needs creates a hierarchy of purposes that strives for coherence, forming the “structure of our soul” [1:22:26:00]. Adding this “affective part” involves incorporating these motivational systems into AI development.
Debate on AI understanding and consciousness
Panpsychism and Roger Penrose
The discussion touches on panpsychism and Roger Penrose’s theory that consciousness arises from quantum gravity effects in microtubules [57:07:00], [1:01:29:00]. This perspective suggests that computation alone cannot explain consciousness, linking it to the uncomputable parts of mathematics as interpreted from Gödel’s proof [59:22:00]. Bach counters that Gödel’s proof implies that only “constructive mathematics” (computable) is real, while classical mathematics with infinities cannot be implemented in physical causal structures [1:00:11:00].
The “Dream” Hypothesis and Psi Phenomena
The conversation explores the idea that if phenomena like telepathy or clairvoyance (Psi phenomena) were real, it would imply that the universe operates not purely mechanically, but like a “dream” with symbolic, non-physical interactions [1:02:42:00], [1:03:30:00]. Such phenomena would be incompatible with known physics and statistics of the universe [1:02:22:00]. Bach proposes that the ability to experience Psi phenomena within our internal model is because we are already living in a “dream” generated by our minds, which are physically grounded elsewhere [1:04:08:00], [1:04:48:00].
The Nature of Reality and Perception
Our perception of reality is a representation, a “simulacrum,” not an isomorphic experience of the physical world [1:05:06:00]. “Realness” itself is a model property that the mind attaches to its perceptual dimensions, indicating predictability of future sensory patterns [1:09:12:00]. While our everyday human perceptions are “gross” compared to what scientific instruments can probe [1:10:13:00], the observed lawfulness across many orders of magnitude strengthens the case for an objective universe [1:10:48:00]. To have an enlightened relationship with reality, one must realize that perception and even self-perception are representations, and cultivate awareness of how the attentional system constructs this reality [1:11:42:00].