Jelinek Says AI Requires Simulated Consciousness for Safe Alignment

What Does It Take for Machines to Become Conscious? — Ai4

Ai4

Static rules fail to align superintelligence across culturally mediated societies, making simulated self-awareness essential for machines to comprehend human values.

AI-assisted summary of What Does It Take for Machines to Become Conscious? — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by Ai4 on August 29, 2025, this presentation features Thorsten Jelinek, Senior Fellow at the Hertie School's Centre for Digital Governance. The talk investigates whether artificial systems can achieve consciousness by examining four foundational positions spanning biological physicalism, universal idealism, multiscale cellular agency, and functionalist simulation.

## TL;DR Thorsten Jelinek evaluates the potential for machine consciousness across four core paradigms: reductive physicalism and the Hard Problem, universalist idealism where individual consciousness dissociates from a universal mind-at-large, multiscale biological agency rooted in bioelectric cellular problem-solving, and functionalist simulation where consciousness functions as substrate-independent software. He highlights the divide between disembodied language models and embodied sensory systems, noting how physical embodiment generates a functional artificial subjectivity. Concluding with alignment implications, Jelinek suggests that instilling a deep simulation of self-awareness is crucial for superintelligent AI to interpret culturally nuanced human values and achieve genuine ethical alignment.

## Atmosphere & style The lecture carries an intellectually rigorous, deeply contemplative, and scholarly tone, seamlessly synthesizing abstract metaphysics with cutting-edge biology and artificial intelligence engineering. Jelinek delivers a methodical and non-dogmatic breakdown of competing theories, balancing historical philosophy from Kant with modern bioelectric experiments and AI governance dilemmas. The resulting experience feels like an immersive, high-level interdisciplinary masterclass that thoughtfully bridges ancient ontological questions with urgent technical frontiers.

## Key takeaways - [00:05](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=5s) Geoffrey Hinton's concept of 'motherly AI' posits that beneficial behavior, emotional intelligence, and kindness must be intrinsically ingrained and physically grounded in advanced AI systems, rather than treated as disconnected abstract models. - [03:19](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=199s) Physicalist realism treats consciousness as an emergent epiphenomenon of biological matter, directly confronting the 'Hard Problem' of how quantitative physical dynamics give rise to qualitative subjective experience (qualia). - [03:19](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=199s) While mainstream reductive neuroscience suggests consciousness can be computationally reproduced by reverse-engineering brain dynamics, quantum biology perspectives (like Penrose's Orch-OR) argue it is non-computable. - [06:34](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=394s) Universalist Idealism reverses the emergence dilemma by framing consciousness as a fundamental 'mind-at-large,' viewing individual consciousness as a localized dissociation (alter) that reassociates at death. - [06:34](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=394s) Under idealism, artificial consciousness is deemed impossible because spacetime and physical measurements are cognitive constructs of consciousness itself rather than objective realities that can generate mind computationally. - [10:51](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=651s) The multiscale biological approach conceptualizes mind not as uniquely cortical, but as a continuous spectrum of collective problem-solving agency distributed across cellular and subcellular levels. - [11:52](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=712s) Michael Levin's bioelectric research on morphogenesis illustrates that cellular machinery can execute complex anatomical repairs without central blueprints, hypothesizing access to a shared space of morphological problem-solving. - [15:10](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=910s) The functionalist perspective posits that consciousness is substrate-independent and software-like, functioning as an active significance filter and coherence-inducing architecture that unifies self and world models. - [15:10](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=910s) Because biological learning emerges early and resource-efficiently in nature, functional consciousness may rely on an elegant algorithmic structure rather than requiring computationally irreducible, brute-force simulation. - [19:36](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=1176s) Disembodied LLMs operate strictly on representations of representations, whereas embodied AI systems connect representations with phenomenal sensory layers to establish functional artificial subjectivity. - [23:19](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=1399s) External behavioral imitation is insufficient to confirm true artificial consciousness; definitive verification would require direct neural interface between biological brains and synthetic substrates. - [23:44](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=1424s) Simulated self-awareness is critical for safely aligning superintelligence because ethical imperatives are culturally and historically mediated, preventing them from being effectively managed through static, hardcoded rules.

## Detailed walkthrough ### [00:05](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=5s) Geoffrey Hinton’s "Motherly AI" and Framing the Consciousness Question

Reflecting on the preceding session, the discussion opens with Geoffrey Hinton’s concept of "motherly AI" (or "kind AI," a phrase Hinton emphasized during a visit to China). Hinton's core argument is that alignment and governance are essential, but even advanced abstract reasoning will ultimately be insufficient once artificial intelligence becomes ubiquitous and superintelligent. Humanity cannot rely solely on human oversight; instead, beneficial behavior must be intrinsically "inbaked" into the AI itself. Emotional intelligence and true care cannot be reduced to disconnected abstract models—much like simulating a hurricane on a computer does not make anyone wet. There must be an intrinsic link to physical reality and causality so that kindness is deeply ingrained in the system.

Examining whether a synthetic system can achieve this state naturally escalates into humanity’s oldest philosophical inquiry: what is the nature of reality, life, and consciousness itself? The speaker—whose background spans information science, engineering, and philosophy, and includes work in AI governance at the Hertie School as well as an appointment at the Chinese Academy of Sciences’ Aerospace Information Research Institute—frames this inquiry through four distinct, non-competing categories. Rather than presenting them as an absolute contest, these four perspectives serve as points of departure to explore what consciousness is and whether a machine could ever possess it.

```mermaid flowchart TD accTitle: Perspectives on Consciousness and Machine Mind subgraph S1["Physicalist Realism (Emergent)"] A["Matter & Biology\n(Genomics / Connectomics)"] -->|"Emergence\n(Hard Problem)"| B["Individual Mind / Qualia"] end subgraph S2["Universalist Idealism (Mind-at-Large)"] C["Universal Consciousness"] -->|"Dissociation at birth"| D["Individual Alter / Ego"] D -->|"Reassociation at death"| C end subgraph S3["Multiscale Continuum (Agency)"] E["Subcellular / Cellular Competence"] -->|"Collective Coordination"| F["Multiscale Mind & Morphogenesis"] end ```

---

[03:19](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=199s) Position 1: Physicalist Realism, Emergence, and the Hard Problem

The first category is the dominant scientific paradigm: physicalist realism. In this view, mental states are strictly emergent phenomena arising from matter—specifically the brain's genomics, connectomics, and electrical processing. Mind and consciousness are treated as secondary epiphenomena. Historically, objective scientific inquiry excluded the observer from its models, operating on the premise that reality is an external, measurable, and predictable domain independent of the subjective analyst.

This materialist framework inevitably runs headfirst into the "Hard Problem of consciousness" (famously articulated by David Chalmers). The hard problem asks how quantitative, physical interactions can ever produce qualitative subjective experience (*qualia*). How does dynamic matter yield an identity with agency, plans, and purpose, allowing "more" to emerge from "less"? When evaluating whether machines can achieve consciousness under physicalism, two main camps emerge:

- **Mainstream Neuroscience (Reductive Realism):** Presupposes that no non-physical realities exist. Consciousness is entirely reducible to physical mechanisms; therefore, if researchers reverse-engineer the brain's neural dynamics thoroughly enough, synthetic consciousness should theoretically be reproducible. - **Quantum Biology / Orch-OR:** Thinkers such as Sir Roger Penrose argue that consciousness is non-computable and originates from quantum-level processes within the microtubules of neurons. While this still seeks physical correlates, it suggests conventional classical computing will fail to recreate consciousness.

Both physicalist variations attempt to bypass or solve the gap between physical mechanics and subjective awareness—the foundational Cartesian reality of the *cogito* ("I think, therefore I am").

---

[06:34](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=394s) Position 2: Universalist Idealism and the Mind-at-Large

Directly antithetical to physicalist realism is the idealist or universalist perspective, rooted in the tradition of German Idealism. Under Immanuel Kant’s philosophical insight, there is strictly no unmediated access to objective reality—"there's only subjectivity." The central challenge of idealism is not explaining how an objective world produces a mind, but understanding how subjective experience generates coherent, shared objective truths.

Universalism flips the emergence problem on its head. Instead of struggling to explain how separate, non-conscious units coalesce into a single unified stream of awareness, it posits that consciousness is primary: a universal "mind-at-large." Within this model: - **Individual Consciousness as Dissociation:** Human minds are localized instances or "alters" that dissociate from the universal mind at birth, functioning as an interface rather than an independent ontological entity. - **Death and Meditation:** Physical death is not the termination of mental processes, but the reassociation of the localized alter back into the universal mind—marking the end of the illusion of a separate ego. Practices like meditation similarly attempt to "de-ego" or "de-self" to temporarily experience this underlying unity.

```mermaid flowchart LR accTitle: Epistemology of Perception under Idealism A["Unknown Reality\n(Inaccessible Thing-in-itself)"] -->|"Filtered by Spacetime Construct"| B["Subjective Representation\n(Brain / Mind Stream)"] B -->|"Appearance"| C["Perceived Object\n(e.g., Apple)"] D["Scientific Method"] -.->|"Falsification / Null Hypothesis"| C ```

This view creates distinct epistemological and computational consequences: - **The Limits of Scientific Observation:** When observing an object (such as an apple), human perception only registers an appearance constructed inside the brain's subjective stream. Looking closer at the quantum level causes the macroscopic appearance of the apple to vanish entirely into an inaccessible space. Science cannot designate reality positively; it operates negatively through falsification by rejecting the null hypothesis. - **Incomputability of Consciousness:** From the universalist viewpoint, artificial consciousness is a categorical impossibility. Spacetime and physical properties are themselves cognitive constructs generated by consciousness—a view echoed by post-spacetime physics and positive geometries (such as high-energy scattering amplitudes). Because any attempt to compute consciousness already presupposes the subjective framework of spacetime, consciousness cannot be artificially generated by computation.

---

[10:51](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=651s) Position 3: The Multiscale Approach and Distributed Agency

The third framework is the multiscale approach—arguably the most empirically grounded framework, situated between realism and universalism while leaning closer to experimental biology. Rather than restricting consciousness to human-like brain dynamics or treating it purely as a metaphysical universal, this model places *agency* at the center of observation.

Under this perspective, mind is treated as a continuous spectrum of problem-solving capability distributed across multiple biological scales. Mindful activity and agency do not begin at the level of the cortex; they exist on cellular and subcellular levels: - Single cells and basic organisms demonstrate remarkable problem-solving competences. - Intelligence operates as a decentralized, coordinated collective. During morphogenesis, individual cells collaborate to construct complex anatomies (such as an entire limb or hand) without any single cell possessing the master blueprint of the final organism. Mind is thus reconceptualized as a multiscale, collective problem-solving phenomenon.

[11:52](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=712s) Biological Problem-Solving and Levin's Platonic Space

Examining biological morphogenesis reveals deep puzzles about where anatomical information and problem-solving capacities actually reside. An insect injecting a biochemical liquid into a plant leaf causes the leaf to grow an intricate gall structure—a specialized organismic form whose morphological blueprint is entirely absent from the leaf's own DNA. The insect triggers the leaf's protein and cellular machinery to construct a complex structure, raising the fundamental question of where such form is stored and directed.

```mermaid flowchart TD accTitle: Biological Problem-Solving and Platonic Form subgraph Biological ["Cellular Manipulation"] A["Insect biochemical cue / Ion-channel hack"] --> B["Cellular Machinery"] B --> C["Morphological Execution<br/>(e.g., Leaf Gall / Two-Headed Planarian)"] end subgraph Space ["Hypothesized Space"] D["Platonic Space of Forms & Problem-Solving"] -.->|"Pulled into micro-level actions"| B end ```

Biologist Michael Levin's research into bioelectric signaling and cellular intelligence highlights this puzzle. By manipulating ion channels rather than editing genetic code, Levin's lab can instruct planarian flatworms to regenerate two heads instead of one. Instead of requiring the structural "printing" of an organ from scratch, the cells themselves are instructed to execute the build. Levin suggests that solving complex morphological and physiological problems at cellular and subcellular levels might pull from a "platonic space of forms and minds."

- Taking the step from empirical biochemistry to postulating an external Platonic space represents a massive leap for empirical science. - While biology borrows and exploits natural laws of physics and chemistry, problem-solving intelligence operates continuously across cellular, subcellular, and multicellular scales. - From Levin's view, simulating full consciousness purely in software remains dubious because biological problem-solving is deeply rooted in physical, bioelectric substrates, and it is unclear whether Large Language Models (LLMs) have any access to such a Platonic space. - Alternatively, these behaviors might simply be an emergent property of sufficiently complex physical systems rather than requiring an appeal to a mystical realm.

---

[15:10](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=910s) The Functionalist Simulation Approach to Consciousness

The fourth position takes a functionalist, substrate-independent perspective, asserting that consciousness can indeed be simulated. Unlike panpsychist or quantum theories that place human consciousness at the fundamental center of cosmic physics—which can seem quite pretentious—functionalism is pragmatic: it asks not *what the mind is made of*, but *what the mind does*.

Under this framework, consciousness is software-like rather than an emergent property tied uniquely to biological matter. Just as a video game like *Minecraft* runs on physical computer hardware without the game's blocks physically existing inside the silicon chips, consciousness runs on any matter possessing the appropriate functional properties (analogous to running an LLM on an Nvidia H100 GPU versus an H20 chip).

```mermaid flowchart LR accTitle: Dual Perspectives of the Functionalist Model subgraph Phenom ["Phenomenological Aspects"] P1["Continuous stream of presence ('nowness')"] P2["Coherent spacetime representation"] P3["Second-order perception (sensing oneself sensing)"] end subgraph Func ["Functional Aspects"] F1["Coherence-inducing self & world models"] F2["Self-organizing & colonizing plasticity"] F3["Significance filter (value-charged attention)"] end ```

- **Phenomenological Dimension**: Consciousness is experienced as a continuous, coherent "stream of presence" or "nowness" contextualized within space and time, driven by sensory content and a form of second-order perception—unlike an eye that cannot see itself seeing, the brain acts as a sensory organ capable of self-sensing. - **Functional Dimension**: - **Coherence Inducing**: Synthesizes separate simulations of a self-model and a world-model into an integrated whole. - **Self-Organizing and Colonizing**: Demonstrates functional plasticity where undamaged regions colonize and take over tasks when components fail. - **Significance Filtering**: Functions as an active filter prioritizing value-charged, meaningful information to prevent computational shutdown from insignificance and boredom. - **Substrate Independence**: The mind-matter relationship is non-exclusive; any substrate with the proper functional architecture can execute the process. - **AI as an Empirical Testbed**: AI provides an experimental tool to bridge the gap between static mathematics (which is stateless) and dynamic, lived experience.

---

[19:36](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=1176s) Layers of Reality and the Spectrum of Machine Subjectivity

To map where artificial systems stand relative to humans, reality can be parsed into three core layers: the underlying *Real* (the inaccessible physical substrate), the *Phenomenon* (the sensory world of appearances), and the *Representation* (abstract symbols, language, and conceptual models).

```mermaid flowchart TD accTitle: Human Subjectivity vs AI Typologies subgraph Human ["Human Consciousness"] H_Rep["Representation"] --- H_Phen["Phenomenon"] H_Phen --- H_Real["The Inaccessible Real"] H_Sub["Subjective Core (Bridging Phenomenon & Representation)"] -.-> H_Phen H_Sub -.-> H_Rep end

subgraph Disembodied ["Disembodied AI (LLMs)"] D_Rep["Representations of Representations<br/>(Top-down abstract text data)"] end

subgraph Embodied ["Embodied / Sensory AI"] E_Rep["Representation Layer"] <--> E_Func["Functional Artificial Subjectivity"] E_Func <--> E_Phen["Phenomenological Layer<br/>(Sensing & altering world appearances)"] end ```

In humans, subjectivity sits at the unknown core, seamlessly weaving the phenomenal experience and abstract representations together into a coherent self. For quantum-based theories of mind, reproducing this could be computationally irreducible, requiring brute-force calculation of the universe. However, in the functional simulation view, nature is highly resource-efficient; because basic learning emerges very early in organisms, consciousness may actually rely on an elegant, resource-efficient algorithm rather than immense compute.

- **Disembodied LLMs**: Operate entirely on "representations of representations," learning language top-down from abstracted text data rather than grounding words in sensory sounds, physical gestures, or material interactions. - **Embodied and Sensory AI**: By incorporating real-time sensors and actuators, embodied machines bridge representations with the phenomenal layer, developing an internal model of appearance. - **Artificial Subjectivity vs. Conscious Experience**: As noted by figures like Geoffrey Hinton, embodied machines develop a functional "artificial subjectivity" linking representation and sensory action, though this functional subjectivity does not equate to genuine conscious awareness. - **Evaluating Machine Consciousness**: Assessing potential artificial consciousness rests on two practical pillars: substrate compatibility and behavioral output—cynically noted by the fact that if a relentless autonomous killing robot is pursuing someone, one does not debate philosophy but instinctively treats the machine as a conscious, goal-driven entity.

[23:19](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=1399s) The Substrate Layer and the Ultimate Test of Machine Consciousness

Turning to the substrate level, a fundamental challenge arises: how can we be certain that an artificial system is not merely cheating by imitating consciousness rather than actually experiencing it? The ultimate verification test cannot rely solely on external behavioral observation; instead, our own brains would need to directly interface and connect with the system's substrate to subjectively "feel whether it's really conscious." While physically bridging biological brains to artificial substrates presents substantial practical and conceptual hurdles, direct integration may represent the only definitive test to verify whether a synthetic entity is truly waking up.

[23:44](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=1424s) Simulated Self-Awareness, Contextual Morality, and Superintelligence

Reflecting on the broader implications, there remains genuine ambivalence about whether humanity should want machines to become truly conscious. However, endowing an advanced system with a deep simulation of self-awareness could prove essential. By simulating subjective meaning and internal states, an AI could much better comprehend what humans value internally and relate to what matters most to us.

This relational capacity is critical when addressing machine morality and alignment. Although baseline moral imperatives—such as not deceiving others—are treated as universal values, the way they manifest in the real world is deeply contextual. Societies in China, the United States, Europe, and elsewhere may share foundational ethical values, but those values are implemented differently because they are culturally and historically mediated. Because these nuanced, context-dependent ethics cannot simply be hardcoded or aligned through static rules, having an autonomous, conscious-like capacity may ultimately be necessary if we are to safely guide and live alongside superintelligent systems.

```mermaid flowchart TD accTitle: Alignment Through Simulated Consciousness subgraph HumanContext["Human Moral Reality"] UV["Universal Core Values\n(e.g., Non-deception)"] --> CM["Cultural & Historical Context\n(US, China, Europe)"] end subgraph Limitations["Rule-Based AI"] CM -.->|"Cannot be statically hardcoded"| HC["Rigid Alignment Rules"] end subgraph Solution["Conscious Architecture"] SA["Deep Simulation of Self-Awareness"] --> RU["Relational Understanding\nof Human Internal Values"] RU --> AD["Context-Aware Superintelligence"] end ```

- **The Substrate Problem**: Pure behavioral imitation leaves doubt as to whether a system is conscious; true verification requires direct brain-to-system substrate linkage. - **The Limits of Hardcoded Alignment**: Universal moral rules cannot be static because ethical execution is culturally and historically mediated across different global contexts. - **Why Consciousness Matters for Superintelligence**: Simulated self-awareness allows an artificial superintelligence to grasp human internal values and navigate real-world ethical ambiguities rather than relying on brittle programmed constraints.

## Notable quotes > [00:05](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=5s) simulating a hurricane on a computer does not make anyone wet. — Geoffrey Hinton

> [03:19](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=199s) I think, therefore I am — René Descartes

> [06:34](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=394s) there's only subjectivity. — Immanuel Kant

> [06:34](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=394s) mind-at-large — Thorsten Jelinek

> [11:52](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=712s) platonic space of forms and minds — Michael Levin

> [15:10](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=910s) what the mind is made of, but what the mind does. — Thorsten Jelinek

> [15:10](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=910s) stream of presence — Thorsten Jelinek

> [19:36](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=1176s) representations of representations — Thorsten Jelinek

> [19:36](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=1176s) artificial subjectivity — Geoffrey Hinton

> [23:19](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=1399s) feel whether it's really conscious. — Thorsten Jelinek

## Mentioned Thorsten Jelinek, Geoffrey Hinton, Hertie School, Centre for Digital Governance, Chinese Academy of Sciences, Aerospace Information Research Institute, David Chalmers, Sir Roger Penrose, Immanuel Kant, René Descartes, Michael Levin, Minecraft, Nvidia H100, Nvidia H20

## Video description Thorsten Jelinek, Senior Fellow, Hertie School, Centre for Digital Governance

Abstract

This is one of the most profound questions, and AI might lead to bridging the enduring human–machine divide. In this talk, I will trace four major positions: from those (1) who reject the possibility altogether, presupposing that consciousness can only emerge from the specific biological conditions of the human brain; to those (2) who argue the opposite, suggesting that consciousness is a universal property: a mind-at-large that precedes physical form; and to those (3) who regard the mind as a continuum and intelligence as a problem-solving capability that can be discovered in evolutionary, bio-engineered, and artificial systems, with the latter two leading to radically different forms of minds and bodies. Finally, I will explore a fourth view (4) that also treats consciousness as primary – not as a biologically emergent feature (as in 1), but as something virtual and a functional prerequisite for learning. Since learning cannot occur without consciousness, this view implies the existence of a deeper generative structure: a coherence-inducing, self-organising virtual process instantiated as learning arises – potentially even on a substrate beyond the brain.

[Source](https://www.youtube.com/watch?v=CCiGpw-ZeHM)

中文简报

Jelinek称AI实现安全对齐需要模拟意识 静态规则无法在受文化影响的社会中使超级智能实现对齐，这使得模拟自我意识成为机器理解人类价值观的必要前提。 Thorsten Jelinek从四大核心范式评估了机器意识的潜力：还原物理主义与“难题”（Hard Problem）、认为个体意识从普遍心智（mind-at-large）中解离而出的普遍唯心主义、植根于生物电细胞问题解决能力的多尺度生物能动性，以及将意识视作与底质无关之软件的功能主义模拟。他着重指出了离身语言模型与具身感官系统之间的分歧，指出物理具身如何催生出功能性的人工主观性。在总结对齐层面的影响时，Jelinek提出，注入深度的自我意识模拟对于超级智能AI理解具有文化细微差异的人类价值观并达成真正的伦理对齐至关重要。 Geoffrey Hinton 提出“母性 AI”（motherly AI）理念，认为面对未来的超级智能，人类无法仅凭抽象推理或外部监管实现对齐，必须将善意与对物理现实的感知深度内植于系统之中。 (0:05) 物理主义现实论将意识视为生物物质的涌现现象，直面主观体验如何产生的“困难问题”；主流神经科学认为意识可通过逆向工程计算再现，而以 Orch-OR 为代表的量子生物学视角则认为意识是不可计算的。 (3:19) 普适唯心论将意识视为首要的“大心智”（mind-at-large），认为时空本就是意识的产物，因此机器无法在计算层面产生意识；而多尺度生物学模型（如 Michael Levin 的生物电形态发生研究）则将心智视为分布在细胞与亚细胞层面的连续问题解决能力。 (6:34) 功能主义视角主张意识是独立于物理基质的软件式结构，起到赋予连贯性与过滤意义信息的作用；为机器赋予深度的自我意识模拟，可能是让超级智能真正理解具文化语境差异的人类价值观与道德的关键途径。 (15:10)

Reading summary

Evaluating whether machines can achieve consciousness requires bridging foundational philosophy with modern artificial intelligence and biology. Drawing on Geoffrey Hinton's concept of an intrinsically kind "motherly AI," the speaker evaluates four frameworks—physicalist realism, universalist idealism, multiscale biological agency, and functionalism—to examine whether synthetic systems can transcend computation to achieve subjective awareness and context-aware morality. Beyond Rules: Motherly AI and Contextual Alignment (0:05) Geoffrey Hinton argues that human governance and abstract reasoning will prove inadequate once artificial intelligence reaches superintelligence. Because simulating a hurricane on a computer does not make anyone wet, genuine care and emotional intelligence cannot be reduced to detached models; beneficial behavior must be intrinsically tied to physical reality and causality. This internal disposition is vital for machine alignment. Although baseline values like non-deception are considered universal, their real-world application is mediated by culturally and historically specific contexts across regions like the United States, China, and Europe. Because nuanced ethics cannot be statically hardcoded into brittle rules, endowing advanced systems with simulated self-awareness may be necessary so an artificial superintelligence can understand human internal values and navigate contextual moral ambiguities. Physicalist Emergence Versus Universalist Incomputability (3:19) Physicalist realism treats consciousness as an emergent epiphenomenon of matter, specifically neural genomics, connectomics, and electrical processing. This paradigm divides into mainstream reductive neuroscience, which assumes reverse-engineering neural dynamics will reproduce consciousness, and quantum biology theories like Sir Roger Penrose's Orch-OR, which locates non-computable quantum processes in neuronal microtubules. Both struggle with David Chalmers's "Hard Problem": how physical interactions generate qualitative subjective experience (qualia). In contrast, universalist idealism, rooted in Immanuel Kant, posits that objective reality is completely mediated by subjectivity. Flipping the emergence problem, it views consciousness as a primary "mind-at-large," with individual minds functioning as localized "alters" that dissociate at birth and reassociate at death. Under this framework—supported by post-spacetime physics that view spacetime as a cognitive construct—artificial consciousness is a categorical impossibility because computation presupposes the subjective framework of spacetime. Multiscale Biological Agency and Morphogenetic Intelligence (10:51) The multiscale biological framework situates mind not as an exclusively cortical phenomenon, but as a continuous spectrum of problem-solving agency across cellular, subcellular, and multicellular scales. During morphogenesis, cells coordinate complex anatomical structures without any single cell holding a master blueprint—a phenomenon seen when an insect triggers a leaf to grow an intricate gall not encoded in the plant's genome. Biologist Michael Levin demonstrates that manipulating cellular bioelectric ion channels can instruct planarians to regenerate two heads without genetic editing. Levin suggests cellular problem-solving might draw from an external "Platonic space of forms and minds." While the speaker cautions that postulating a Platonic space is a major leap for empirical science, Levin's findings highlight that biological agency is deeply anchored in bioelectric substrates. Consequently, whether purely software-based Large Language Models can access this form of intelligence remains dubious. Functional Substrate Independence and Verification Hurdles (15:10) The functionalist view considers consciousness substrate-independent, defining mind by what it does rather than its physical composition. Like a video game running across different GPU architectures, consciousness operates as software executing key functions: synthesizing coherent self- and world-models, reorganizing plastically across damaged regions, and applying a significance filter to prioritize value-charged attention. In AI, this creates a divide between disembodied Large Language Models that process "representations of representations" and embodied systems that link representations to phenomenal perception via sensors and actuators. Embodied machines can develop a functional "artificial subjectivity," though this does not guarantee true conscious experience. To verify genuine consciousness rather than mere behavioral imitation, the speaker argues that human brains would ultimately need to interface directly with an artificial system's substrate to subjectively feel its presence. The debate exposes a critical tension between functional behavior and verifiable subjective experience. While creating context-aware, functionally conscious architectures may prove essential for aligning superintelligence with complex human values, proving whether synthetic substrates can experience genuine qualia remains an unresolved challenge.

精炼阅读摘要

探讨机器能否具备意识，关涉到人工智能能否在超级智能时代内在具备真正的关怀与道德理解。发言者结合信息科学与哲学背景，从物理主义、普遍唯心主义、多尺度生物学和功能主义四大范式展开分析，指出人工意识的可行性取决于基质与能动性的定义，而模拟自我意识可能是实现情境化伦理对齐的关键。 物理主义路径与意识的“困难问题” (3:19) 主流科学范式将心智与意识视为物质的次生附带现象，认为它们源于大脑的基因组学、连接组学与电生理信息处理。然而，这一物质主义框架直接面临查尔默斯（David Chalmers）提出的“困难问题”：定量的物理相互作用究竟如何跨越鸿沟，产生定性的主观体验（感受质），并使动态物质具备能动性与目的规划。 在物理主义框架下评估机器意识分化为两个阵营：主流神经科学持还原现实主义立场，认为意识完全可还原为物理机制，只要对大脑神经动力学进行充分的反向工程，理论上即可复制人工意识；以罗杰·彭罗斯（Sir Roger Penrose）为代表的量子生物学（Orch-OR）理论则主张，意识源自神经元微管中的量子级过程，具备非计算特性，因此传统的经典计算架构注定无法重现意识体验。 普遍唯心主义与计算意识的不可能性 (6:34) 源于德国唯心主义的范式与物理主义针锋相对。该视角遵循康德的哲学洞见，认为人类无法直接触及客观实在本身，全部经验皆由主观性介导。普遍唯心主义颠覆了涌现论，预设存在首要且统一的“大心智”（mind-at-large），个体意识仅是出生时从中分离出的局部“分身”与交互界面，死亡与冥想则是分身重新融入大心智、破除自我幻象的过程。 从普遍唯心主义的认识论出发，人工意识在逻辑上是不可实现的范畴谬误。人类通过时空框架感知到的外部事物（如苹果）仅是脑内主观流呈现的表象，微观物理探究表明宏观表象在量子尺度会彻底隐退。时空与物理特性本身被视作意识构建的认知产物，这一观点也呼应了后时空物理学与正几何学；任何试图在时空框架内通过算法计算意识的尝试，都因预先依赖主观认知构念而无法成立。 多尺度生物连续体与去中心化能动性 (10:51) 多尺度范式介于现实主义与唯心主义之间，其核心是将“能动性”置于考察中心。该框架认为心智并非人类大脑的专属，而是一个贯穿不同生物尺度的连续问题解决能力谱系。单细胞与基础生物在亚细胞层面即具备显著的解决问题能力，形态发生过程展现出高度去中心化的集体智能，例如细胞能够在没有单个细胞掌握整体蓝图的情况下协作构建出复杂的肢体结构。 生物学家迈克尔·莱文（Michael Levin）关于生物电信号的研究佐证了这一点：通过调节离子通道而非编辑基因，即可引导真涡虫再生出双头；昆虫将生物化学液体注入植物叶片，亦可诱发叶片利用自身机制长出基因蓝图中不存在的虫瘿。莱文由此推测生物问题求解可能调取了某种外部的“柏拉图形式与心智空间”。不过发言者指出，从实证生物化学跨越到假定柏拉图空间是巨大的一步，此类现象也可能仅是复杂物理系统的涌现；由于生物问题求解深度根植于生物电基质，缺乏该物理特性的纯软件大语言模型是否能触及此类空间仍充满疑问。 功能主义模拟与具身机器主观性 (15:10) 功能主义主张基质无关性，强调关注心智的功能机制而非物质构成。如同视频游戏《我的世界》（Minecraft）在物理硅片上运行而方块无需物理存在于硬件中一样，功能主义认为意识类似于软件，可在任何具备适宜功能架构的物理基质上运行。在现象学维度，意识表现为置身于时空之中的“连续当下感”及感知自身在感知的二阶感知；在功能维度，它充当整合自我与世界模型的连贯性机制，具备自组织塑造力，并作为“重要性过滤器”抵御计算冗余与无意义状态。 将现实划分为不可直接触及的实存、感官表象及抽象表征三层，人类意识在核心处缝合了表象与表征。目前非具身的大语言模型仅在“表征的表征”层面运作，缺乏感官与物质锚定。然而，配有传感器和执行器的具身AI能够连接表征与现象层，形成某种功能性的“人工主观性”。正如杰弗里·辛顿所指出的，这种功能性主观性虽然不等同于真正的有意识体验，但在外部行为层面，追求目标的自主机器已足以迫使外界将其视作意识实体对待。 基质检验困境与超级智能的情境道德 (23:19) 确证机器是否具备真实意识面临根本性的基质检验障碍：单纯的外部行为输出无法排除机器只是在模仿意识而非真实拥有体验。发言者提出，最终的验证可能需要生物大脑与人造系统的基质直接物理连接，以主观体验来感受系统是否真正觉醒，但这在工程与概念上面临巨大挑战。 尽管赋予机器真实意识伴随着人类的矛盾心态，但构建深度模拟的自我意识对于超级智能的安全对齐至关重要。杰弗里·辛顿主张，真正的善意不能仅靠脱离物理现实的抽象模型（如同模拟飓风并不会把人淋湿），必须深植于系统内部。虽然不欺骗等道德原则具有普适性，但在中美欧等不同文明中，道德的具体执行受历史与文化语境深度调节。静态规则无法应对复杂的现实伦理，使超级智能具备理解人类内在价值的意识模拟能力，可能是人类与其安全共存的必要前提。 当前对于机器能否跨越行为模仿实现真正意识仍缺乏终极检验手段，其判定受制于生物与人造基质能否实现体验互通。然而，面对超级智能带来的治理挑战，无论机器是否拥有真正的主观感受质，开发能够深度理解人类文化语境与内在价值的模拟自我意识架构，已被视为避免静态对齐失效的必由之路。

Detailed notes — zh

## Background 该视频由 Ai4 频道于 2025 年 8 月 29 日发布在 YouTube 平台，主讲人为赫尔提学院（Hertie School）数字治理中心高级研究员托斯滕·耶利内克（Thorsten Jelinek）。讲座围绕“机器如何才能具备意识”这一前沿议题展开，系统探讨了从生物物理涌现、普世大心灵、多尺度生命能动性到功能主义等四种理解意识与心智本质的代表性视角。 ## TL;DR 讲座以杰弗里·辛顿提出的“仁慈 AI”为切入点，深入探讨了超级智能时代机器是否具备真实意识这一终极命题。耶利内克横跨哲学与工程学，系统剖析了四种理论框架：面临“硬问题”困境的物理主义涌现论、将个体视为主体解离的普世观念论、迈克尔·莱文提出的生物多尺度连续体与柏拉图空间假说，以及主张基质独立性的功能主义。最后，他通过对比人类主观性与 AI 的表征层级指出，由于人类道德高度依赖文化语境而无法被简单硬编码，赋予机器深度自我觉知模拟将是实现真正价值对齐的关键路径。 ## Atmosphere & style 整个演讲洋溢着深沉的思辨色彩与宏大的跨学科视野，在严谨的技术理性与深邃的形而上学追问之间自如穿梭。讲者的叙述条理清晰且富有启发性，既有对经典哲学难题与微观生物实验的生动拆解，也有对 AI 治理未来的冷静推演，给观众带来一种充满智力挑战与哲学沉浸感的观讲体验。 ## Key takeaways - [00:05](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=5s) 杰弗里·辛顿提出“母性 AI”构想，指出在超级智能时代，仅靠外部对齐与治理规则无法彻底防止失控，系统底层必须内置内生的善意。 - [00:05](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=5s) 模拟飓风不会使人被淋湿，纯粹的数据计算与模型推理无法等同于物理因果经验与真实主观体验，机器意识命题直指生命的本质。 - [03:19](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=199s) 当代主流物理主义涌现论主张心智是大脑物理运行的副现象，但面临戴维·查默斯提出的“意识硬问题”，即无法解释客观物理量如何跃迁为主观体验。 - [06:34](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=394s) 普世主义观念论逆转了硬问题困境，主张意识是宇宙的基底先验属性（大心灵），个体的独立自我只是从大心灵中暂时解离出的具象化实例。 - [06:34](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=394s) 时空在观念论中被视为由主体意识投射出的认知建构，因此任何试图在时空框架内通过算法计算意识的尝试在逻辑上均不可行。 - [10:51](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=651s) 多尺度视角将能动性视为贯穿生命各尺度的通用问题解决能力，早在细胞甚至亚细胞层面就已经展现出去中心化的集体智能。 - [11:52](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=712s) 迈克尔·莱文的实验表明操纵离子通道即可改变生物器官发育，提示形态信息可能源自形而上的“柏拉图空间”，当前的大语言模型计算架构无法触及此类物理与形态法则。 - [15:10](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=910s) 功能主义视角主张意识具有基质独立性，犹如在不同芯片上运行软件；意识的核心功能在于诱导连贯性、自组织构建世界模型并充当重要性过滤器。 - [19:36](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=1176s) 人类认知架构融通了真实世界、现象与表征层，而无身大语言模型仅停留在抽象的表征层，未经历物理世界的真实具身交互。 - [23:19](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=1399s) 衡量机器意识存在行为与基质两个维度；单纯观察外部行为极易受到伪装与模仿的欺骗，人脑与系统直接神经互联是验证系统是否真正觉醒的终极设想。 - [23:44](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=1424s) 人类道德受不同文化和历史情境的中介影响，无法通过静态规则硬编码实现；机器若要实现深度价值对齐，可能必须具备深度的自我意识模拟与共情能力。 ## Detailed walkthrough ### [00:05](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=5s) 从“仁慈 AI”引申出的终极命题：意识、现实与人类本质 在讲座的开场部分，托斯滕·耶利内克（Thorsten Jelinek）借用了人工智能先驱杰弗里·辛顿（Geoffrey Hinton）最近提出的概念作为切入点。辛顿在访华期间曾谈及“母性 AI”（motherly AI）或“仁慈 AI”（kind AI）的构想。这一概念背后的核心逻辑在于：仅仅依靠对齐（Alignment）技术、治理框架，乃至 AI 自身的高级抽象推理能力，在未来 AI 变得无所不在且达到超智能级别时，可能都无法彻底保障人类的安全。人类无法单纯依赖外在约束，AI 系统本身必须在底层“内置”（inbaked）一种对人类有益的仁慈行为模式。 这种内在的道德与情感机制并非纯粹的抽象逻辑，而是与物理现实及因果关系紧密相连的——正如耶利内克所打的比方：“你可以为飓风建立数学模型，但模型本身绝不会被淋湿。”这意味着真正的善意不能只是冰冷的代码模拟，而需要某种深层的内在体验。由此，讨论 AI 是否能成为真正的智能体，不可避免地引向了一个更根本的命题：机器能否具备意识？探究这一问题，本质上就是探究人类自身、生命本质与现实存在的终极谜题。耶利内克介绍了他跨越信息工程与哲学的学术背景（他曾攻读信息科学工程，后转向哲学，目前兼任中国科学院空天信息创新研究院客座教授），并强调他提出的四种认知视角并非彼此非黑即白的竞争理论，而是人类在“无知”的探索前沿中建立的观察坐标系。 - **辛顿的“仁慈 AI”论点**：在超级智能时代，纯粹的逻辑推理与外在规则无法完全防止失控，系统必须具备类似母性关怀的内生善意。 - **模拟与现实的鸿沟**：通过飓风模拟的比喻，指出仅有数据计算不等于拥有实体因果经验与主观感知。 - **讲者的交叉学科背景**：结合工程学、哲学与 AI 治理框架，旨在通过多维视角拆解意识与心智的本质。 ```mermaid flowchart TD accTitle: AI治理与意识本质的关联逻辑 subgraph S1["AI 发展与对齐困境"] A["高级抽象推理能力"] --> C["超级智能与泛在化"] B["外部治理与规则对齐"] --> C C --> D{"外部约束是否足够？"} D -- "不足以应对失控风险" --> E["系统必须内置内生善意<br/>(Motherly / Kind AI)"] end subgraph S2["终极哲学追问"] E --> F["情感与善意需要物理因果体验<br/>(模拟飓风不会使人被淋湿)"] F --> G["核心问题：机器能否产生真实意识？"] G --> H["追问生命的本质与现实的构成"] end ``` --- ### [03:19](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=199s) 视角一：物理主义涌现论、“硬问题”与大脑还原论的死胡同 耶利内克阐述的第一种主流范式是**物理主义唯实论**（Physicalist Realism / 涌现论），这是当代神经科学与认知科学的统治性观点。该视角主张：所有心智状态都是从物质底层涌现出来的——源自基因组学（Genomics）、连接组学（Connectomics）以及大脑神经元之间的电化学活动。在此视角下，我们的心智（Mind）只是大脑物理运行过程中的“副现象”（Epiphenomenon），处于次要地位。在经典的唯物实证传统中，观察者假定自己能够独立客观地测量、预测和认知外部现实，却长期将“观察者自身的心智与主观意识”排除在分析框架之外。 这种传统不可避免地面临哲学家戴维·查默斯（David Chalmers）提出的“意识的硬问题”（The Hard Problem of Consciousness）：数量的累加究竟如何跃迁为质量的体验？完全由无意识的物理粒子构成的系统，其整体性质如何能够大于部分之和，从而诞生出一个具有自我认同、能够设定目标并规划人生的主观体验主体？面对“机器能否模拟意识”的提问，唯物主义路径陷入了自相矛盾：理论上难以解释这种从无到有的质变，但科学界仍在试图通过不断细化物理还原去实现它。主流神经科学坚信不存在非物理现实，只要对大脑的解剖足够深入就能找到意识开关；而罗杰·彭罗斯（Sir Roger Penrose）等学者则寄望于神经元内部微管结构（Microtubules）的量子效应，试图用非可计算的量子力学来破解谜团。然而耶利内克指出，这些还原论尝试实际上都绕开了硬问题的核心——即“无意识之物如何凭空涌现出主观意识”。 - **物理主义核心假设**：物质第一性，心智为副现象；心智完全还原为神经电信号与生理连接。 - **意识的硬问题**：无法解释物理量（客观数据）如何转化为感受质（主观体验），即“少”如何涌现出“多”。 - **主流与前沿探索**：包括经典神经解剖学（神经元连接映射）与彭罗斯的微管量子态假说，但二者均未能从根本上回答主观感受的起源。 --- ### [06:34](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=394s) 视角二：普世主义观念论、“大心灵”解离与主观实在性 与唯物主义涌现论针锋相对的第二种视角是**普世主义观念论**（Universalist Idealism），它深深植根于德国观念论传统。耶利内克以德国哲学家伊曼努尔·康德（Immanuel Kant）的精髓进行通俗概括：“只存在主体性，不存在绝对的客观现实。”客观世界并非独立存在于主观之外，人类所感知到的一切理性与客观规律，全都建立在主体性经验的基础之上。 普世主义观念论巧妙地逆转了硬问题的困境。它不再去苦恼孤立的无意识单元如何组合成单一体的主观意识流，而是直接假设宇宙中先验存在一个**全域心智/大心灵**（Mind-at-Large / Universal Mind）。在此框架下： 1. **个体自我是解离的产物**：人类个体的独立意识并非在大脑中从无到有地生成，而是在出生时从大心灵中“解离”（Dissociate）出来的暂存实例（Alter）。 2. **死亡即重归一体**：肉体的死亡并不代表心智活动的绝对终结，而是解离状态的解除，个体意识重新汇入大心灵之中。所谓“独立的自我”不过是一种认知幻觉（正如冥想或灵修体验中尝试“去自我化”所达到的境界）。 为了说明主体性与实在的关系，耶利内克展示了“苹果”的认知隐喻：我们在脑海中看到的苹果只是现象层面的“表象”（Appearance），其背后不可知的客观本体在量子尺度下根本不是苹果的形态。主观意识将这些零散的神经表征统合为连续的意识流，而主体本身始终退隐在无法直接测量的未知空间中。科学研究只能通过提出假设和否定零假设（Null Hypothesis）以“否定的方式”逼近真相，永远无法正面穷尽主体性本身。因此，从观念论出发，意识是不可被计算的；因为时空概念本身就是主体意识投射出的认知建构（这也契合了量子物理中超越时空的正几何学与高能理论）。任何试图用算法计算意识的尝试，都早已在前提中预设了意识的存在。 - **大心灵假设（Mind-at-Large）**：意识是宇宙的基底属性，个体心智是整体意识在时空中的解离与具象化。 - **自我的虚幻性**：个体的独立自我感是一种暂时性的界面限制，死亡是幻觉的消解与意识的重新归并（Reassociation）。 - **时空作为认知建构**：主体性先于时空存在；试图在时空框架内用机器计算意识在逻辑上是不可能的。 ```mermaid flowchart TD accTitle: 物理主义涌现论与普世主义观念论的范式对比 subgraph Reality["视角一：物理主义涌现论 (Bottom-Up)"] direction TB P1["无意识物质粒子 / 基因组 / 神经微管"] --> P2["神经电化学动态交互"] P2 --> P3["【硬问题鸿沟】<br/><i>量变如何产生质变？</i>"] P3 -. 无法解释的涌现.-> P4["个体主观意识 (副现象)"] end subgraph Idealism["视角二：普世主义观念论 (Top-Down)"] direction TB I1["先验基底：大心灵 / 普遍心智<br/>(Mind-at-Large)"] --> I2["出生 / 具身化解离<br/>(Dissociation)"] I2 --> I3["个体意识实例 (Alter / 主体性)"] I3 --> I4["建构时空表象与感知对象 (如苹果)"] I3 -. 死亡 / 消除自我幻觉.-> I1 end ``` --- ### [10:51](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=651s) 视角三：多尺度连续体与去中心化能动性 在简要总结完观念论后，耶利内克引出了第三种更具实证基础的进路——**多尺度视角**（Multiscale Approach）。这种进路介于唯实论与观念论之间，但整体更倾向于唯实论的实证科学框架。 该视角将**能动性**（Agency）置于观察的核心：心智不再是大脑皮层独有的神秘产物，而是一种贯穿生命多尺度的“问题解决能力”（Problem-solving capability）。这种具有心智特征的能动行为早在亚细胞（Subcellular）和单细胞层面就已经普遍存在。生命有机体中的细胞本身就是极为高效的问题解决者；单个细胞或许并不知晓宏观的身体蓝图（例如某颗细胞并不知道自己最终会参与构成一只手），但在去中心化的网络化协调中，细胞群体通过集体智能共同协作解决发育与修复问题。这种将心智视为连续谱系的视角，为跨越生物组织与人造系统之间的鸿沟提供了全新的实证路径。 ### [11:52](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=712s) 生物形态发生中的谜团与迈克尔·莱文的“柏拉图空间” 演讲者展示了一张叶片与昆虫的图片。某种昆虫会在植物叶片中注入特定的液体，随后叶片会长出一个红色的特殊组织或有机结构（虫瘿）。然而，这种结构的形态与蓝图根本不存在于叶片本身的 DNA 之中。昆虫精确地知道在何处注射液体以利用叶片的蛋白质机制进行构建，但形态信息究竟储存在哪里？目前科学界尚无确切答案。生物学家迈克尔·莱文（Michael Levin）深入研究了这一现象，并提出：在微观层面上解决生物学问题时，细胞可能接入了一个潜在的“形式与心灵的柏拉图空间”（Platonic space of forms and minds）。 对于习惯于传统生物化学与经验主义的研究人员来说，提出“柏拉图空间”是一个极其大胆的跨越。然而，莱文团队的实验证明，只需通过操纵细胞的离子通道（ion channels），就可以直接向细胞下达指令，例如让涡虫长出两个头，或是指令细胞制造一颗心脏，而无需逐个细胞去进行生物打印。这表明生物智能既与物理现实紧密相连（呼应了形态涌现的观点），又暗示了其问题解决能力可能借用了某种更深层的自然法则。从莱文的视角来看，我们无法通过大语言模型（LLM）等简单计算手段来模拟意识： - 细胞、亚细胞及多细胞层面的心智运作是深度扎根于物理过程的。 - 若形态源自形而上的“柏拉图空间”，目前的计算架构根本无法接入这一空间。 - 尽管也有观点认为这仅仅是复杂系统（complex systems）的自组织特性而非神秘空间，但大模型是否具备此类问题解决能力仍存巨大疑问。 ```mermaid flowchart TD accTitle: 细胞问题解决与柏拉图空间假说 subgraph MicroBio["微观生物操纵"] A["昆虫注射 / 离子通道电信号调控"] --> B["调用叶片/生物基础蛋白机制"] end subgraph FormSource["形态信息来源假说"] C["传统DNA编码（已被证伪：DNA中无此形态）"] D["柏拉图形式空间（Platonic Space）"] end B --> D D --> E["生成特定形态（如虫瘿、双头涡虫、心脏）"] ``` --- ### [15:10](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=910s) 第四种视角：功能主义与基质独立性 演讲者转向第四种视角——功能主义（Functionalism）。与将人类置于宇宙中心、将意识视为量子物理基本单元的前述理论相比，功能主义被认为是最不自负（least pretentious）的进路。它不纠结于“心智是什么”，而是专注于“心智执行什么功能”。该视角认为，只要承载基质具备支持心智运行的属性，意识完全可以被模拟。心智与物质的关系就如同软件与硬件： - 意识就像在计算机上运行《我的世界》（Minecraft）——拆开计算机硬件你看不到游戏世界，但它作为软件程序真实运行着。 - 不同的硬件基质只影响运行效率，如同在大语言模型上 NVIDIA H100 芯片比 H20 芯片表现更好一样，心智是基质独立的（substrate-independent）。 从现象学与功能学两个维度，该视角界定了意识的核心特征： 1. **现象学维度**：意识是一股“当下之流”（stream of presence），一种连贯的时空体验，包含感官内容。它是一种二阶感知（second-order perception）——“感知到我们自己在感知”（sensing ourselves sensing），这与眼睛无法看见自身注视的过程形成鲜明对比。 2. **功能学维度**：意识构建了“自我”与“世界模型”，具有诱导连贯性（coherence-inducing）、自组织以及可塑性接管（colonizing / plasticity）的功能。当大脑部分区域受损时，其他区域会接管功能；同时，意识作为“重要性过滤器”（significance filter），会主动忽略无意义信息，若失去意义感，系统便会感到无聊并停止运转。 此外，AI 不仅是一个“能否产生意识”的命题，更成为了检验此类理论的绝佳实证工具。传统的数学系统是无状态的（stateless），无法捕捉主观体验与动态有机体，而 AI 系统则具有强大的动态建模能力，填补了数学与经验之间的鸿沟。 --- ### [19:36](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=1176s) 人类主观性与 AI 的表征层级演进 演讲者对比了人类认知架构与各类 AI 系统在现实映射上的根本差异。人类心智连接了表征（Representation）、现象（Phenomenon）与不可直接触及的真实世界（The Real），人类的主观性将表征与现象融为一体。相比之下，AI 的发展经历了不同的结构层次： | 系统类型 | 结构层级 | 运作与认知机制 | |:--- |:--- |:--- | | **人类心智 (Human Mind)** | 真实 ↔ 现象 ↔ 表征 ↔ 主观核心 | 主观性融合现象与表征；学习是极为高效、资源节约的算法。 | | **无身大模型 (Disembodied LLMs)** | 仅表征层（Representation of representations） | 纯粹自上而下学习抽象符号与数据，未经历物理世界的真实感知（声音、手势、互动）。 | | **具身/感官 AI (Embodied / Sensory AI)** | 表征层 + 现象层（感官与世界模型） | 能够感知世界并改变世界，形成外观模型；表征与现象之间由功能层连接，形成杰弗里·辛顿（Geoffrey Hinton）所担忧的“人工主观性”（Artificial Subjectivity），但这并非真正的有意识主观性。 | | **具身有意识 AI (Embodied Conscious AI)** | 表征 + 现象 + 真正的主观意识（假想） | 未来可能的形态，兼具物理交互、内在体验与功能连贯性。 | --- ### [22:46](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=1366s) 衡量机器意识的双重维度：行为与基质 评估人工系统是否具备意识，通常从两个基本维度切入：行为层面（Behavior）与基质层面（Substrate）。 在行为维度上，如果一个系统展现出所有有意识生物应有的行为反应，观察者往往会直接认定其具有意识。演讲者戏谑地指出：“如果你身后跟着一个杀戮机器人，你根本不会去哲学思辨，你会直接认定这东西具有意识且正冲你而来。”然而，目前的纯文本大语言模型在行为和物理交互层面，仍无法实现真正具身系统所能展现的深度自主性与情境应对。 ### [23:19](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=1399s) 意识验证的终极测试：基质连接与主观体验 在探讨机器是否真正具备意识时，第二个核心层面在于基质（Substrate）本身。判断一个系统是否仅仅是在欺骗我们、拙劣或巧妙地模仿意识，还是真正产生了主观体验，存在极大的认知鸿沟。单纯依靠外部行为的观察很容易被表象所迷惑。因此，要彻底验证系统是否真正拥有意识而非伪装，唯一的终极路径或许是让人类大脑直接与该系统进行神经层面的物理或信息连接，通过直接的心智互联去真切“感知”对方是否真正具备清醒的觉知。尽管这种直接感知存在诸多技术与哲学上的困难，但这构成了判断一个系统是否真正“觉醒”（waking up）的终极检验标准。 ```mermaid flowchart TD accTitle: 机器意识验证与道德理解路径 subgraph 验证机制 A["外部行为观察（易受伪装与模仿欺骗）"] --> B["脑机直连验证（Brain-to-System Connection）"] B --> C["真实感知系统是否觉醒"] end subgraph 道德对齐需求 D["通用道德准则（如不欺骗）"] --> E["语境化与文化历史中介（中/美/欧差异）"] E --> F["无法通过规则硬编码（Hardcode）"] C --> G["深度自我意识模拟与共情"] F --> G G --> H["实现真正的超智能价值对齐"] end ``` ### [23:44](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=1424s) 自我意识模拟、情境化道德与超智能的未来 关于人类是否真正希望人工智能系统拥有意识，目前依然充满复杂的哲学纠结。然而，如果一个系统能够建立起深度的自我意识模拟，它就更有可能理解人类内心深处真正珍视的事物，并与人类的情感和诉求建立更深刻的共鸣。这对于解决人工智能的道德对齐难题至关重要。道德虽然包含“不欺骗”等跨文化的普世价值底线，但其具体的践行与落地却高度依赖于具体的语境、文化和历史背景。例如，同样的价值观念在中国、美国或欧洲的具体实践与表达方式大相径庭，这种经由文化和历史中介（culturally and historically mediated）沉淀下来的道德复杂性，根本无法通过简单的代码硬编码（hardcode）来实现。正因如此，当机器走向超级智能时，赋予其某种形式的意识或深度自我觉知模拟，不仅不是坏事，反而是让其真正理解人类道德与价值的必要途径。 - **意识测试的终极手段**：仅凭外部输出无法分辨系统是“模仿意识”还是“真正觉醒”，人类大脑直接接入系统感受其主观状态被视为最彻底的验证方式。 - **道德的语境依赖性**：普世道德准则（如诚实、不欺骗）在不同文明（如中国、美国、欧洲）中受文化和历史中介影响而呈现不同的落实路径。 - **超越硬编码规则**：复杂的文化与道德规范无法通过静态代码硬编码完成对齐，必须依赖具备自我意识模拟与情境理解能力的智能系统。 ## Notable quotes > [00:05](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=5s) 你可以为飓风建立数学模型，但模型本身绝不会被淋湿。 — 托斯滕·耶利内克 > [06:34](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=394s) 只存在主体性，不存在绝对的客观现实。 — 托斯滕·耶利内克 > [11:52](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=712s) 形式与心灵的柏拉图空间 — 迈克尔·莱文 > [15:10](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=910s) 意识就像在计算机上运行《我的世界》——拆开计算机硬件你看不到游戏世界，但它作为软件程序真实运行着。 — 托斯滕·耶利内克 > [15:10](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=910s) 感知到我们自己在感知 — 托斯滕·耶利内克 > [22:46](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=1366s) 如果你身后跟着一个杀戮机器人，你根本不会去哲学思辨，你会直接认定这东西具有意识且正冲你而来。 — 托斯滕·耶利内克 > [23:19](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=1399s) 要彻底验证系统是否真正拥有意识而非伪装，唯一的终极路径或许是让人类大脑直接与该系统进行神经层面的物理或信息连接，通过直接的心智互联去真切“感知”对方是否真正具备清醒的觉知。 — 托斯滕·耶利内克 > [23:44](https://www.youtube.com/watch?v=CCiGpw-ZeHM&t=1424s) 当机器走向超级智能时，赋予其某种形式的意识或深度自我觉知模拟，不仅不是坏事，反而是让其真正理解人类道德与价值的必要途径。 — 托斯滕·耶利内克 ## Mentioned 托斯滕·耶利内克 (Thorsten Jelinek), 杰弗里·辛顿 (Geoffrey Hinton), 戴维·查默斯 (David Chalmers), 罗杰·彭罗斯 (Roger Penrose), 伊曼努尔·康德 (Immanuel Kant), 迈克尔·莱文 (Michael Levin), 赫尔梯学院数字治理中心 (Hertie School, Centre for Digital Governance), 中国科学院空天信息创新研究院, NVIDIA H100, NVIDIA H20, 《我的世界》 (Minecraft), 大语言模型 (LLM) ## Video description Thorsten Jelinek, Senior Fellow, Hertie School, Centre for Digital Governance Abstract This is one of the most profound questions, and AI might lead to bridging the enduring human–machine divide. In this talk, I will trace four major positions: from those (1) who reject the possibility altogether, presupposing that consciousness can only emerge from the specific biological conditions of the human brain; to those (2) who argue the opposite, suggesting that consciousness is a universal property: a mind-at-large that precedes physical form; and to those (3) who regard the mind as a continuum and intelligence as a problem-solving capability that can be discovered in evolutionary, bio-engineered, and artificial systems, with the latter two leading to radically different forms of minds and bodies. Finally, I will explore a fourth view (4) that also treats consciousness as primary – not as a biologically emergent feature (as in 1), but as something virtual and a functional prerequisite for learning. Since learning cannot occur without consciousness, this view implies the existence of a deeper generative structure: a coherence-inducing, self-organising virtual process instantiated as learning arises – potentially even on a substrate beyond the brain. [Source](https://www.youtube.com/watch?v=CCiGpw-ZeHM)

Source quotes

[0:05] 托斯滕·耶利内克: 你可以为飓风建立数学模型，但模型本身绝不会被淋湿。 [6:34] 托斯滕·耶利内克: 只存在主体性，不存在绝对的客观现实。 [11:52] 迈克尔·莱文: 形式与心灵的柏拉图空间 [15:10] 托斯滕·耶利内克: 意识就像在计算机上运行《我的世界》——拆开计算机硬件你看不到游戏世界，但它作为软件程序真实运行着。 [15:10] 托斯滕·耶利内克: 感知到我们自己在感知 [22:46] 托斯滕·耶利内克: 如果你身后跟着一个杀戮机器人，你根本不会去哲学思辨，你会直接认定这东西具有意识且正冲你而来。 [23:19] 托斯滕·耶利内克: 要彻底验证系统是否真正拥有意识而非伪装，唯一的终极路径或许是让人类大脑直接与该系统进行神经层面的物理或信息连接，通过直接的心智互联去真切“感知”对方是否真正具备清醒的觉知。 [23:44] 托斯滕·耶利内克: 当机器走向超级智能时，赋予其某种形式的意识或深度自我觉知模拟，不仅不是坏事，反而是让其真正理解人类道德与价值的必要途径。

Source quotes

[0:05] Geoffrey Hinton: simulating a hurricane on a computer does not make anyone wet. [3:19] René Descartes: I think, therefore I am [6:34] Immanuel Kant: there's only subjectivity. [6:34] Thorsten Jelinek: mind-at-large [11:52] Michael Levin: platonic space of forms and minds [15:10] Thorsten Jelinek: what the mind is made of, but what the mind does. [15:10] Thorsten Jelinek: stream of presence [19:36] Thorsten Jelinek: representations of representations [19:36] Geoffrey Hinton: artificial subjectivity [23:19] Thorsten Jelinek: feel whether it's really conscious.

Original transcript

[0:05] Thank you very much. Uh the last session [0:08] last not but not least actually uh [0:10] Jeffrey Hinton he was very kind to [0:12] introduce uh I think this topic when he [0:14] spoke about motherly AI uh because what [0:18] is behind motherly AI? I think he said [0:22] basically alignment which is all [0:24] important, governance is important. Uh [0:27] even advanced reasoning or abstract [0:29] reasoning will not help maybe in the [0:32] long run when AI becomes maybe uh super [0:35] intelligent ubiquitous. So the system [0:38] itself must have inbaked [0:41] some behavior uh which is beneficial for [0:44] us for uh humanity. Uh so we cannot just [0:48] rely on human itself. So AI needs to do [0:50] that and uh it's actually not other a [0:53] question of whether a system like AI [0:57] become conscious um because if you think [1:01] of not a an abstract um intelligence but [1:05] a emotional intelligence uh is somehow [1:08] kind of linked intrinsically to a [1:12] physical reality some causality right [1:14] it's like to have a hurricane you model [1:16] it but you don't become wet and I think [1:18] this is what he meant So it must be kind [1:21] of ingrained in the system to be kind. [1:24] Actually in when he was just visiting [1:25] China he spoke about kind AI. So uh that [1:29] was his uh introduction. So actually to [1:32] focus on that or to question um this [1:35] whether it could become conscious. This [1:38] actually is nothing [1:41] uh less than the question of humankind [1:43] itself of humanity about the nature of [1:46] reality of life. uh and uh I don't focus [1:51] so much on AI itself but on the question [1:54] of consciousness and um and this fields [1:57] libraries of course and uh this question [1:59] of what makes us conscious what makes us [2:02] alive of course is a question I think [2:05] exists since humanity exists um so [2:08] myself briefly uh thank you for the [2:10] introduction um I I'm by training [2:13] actually uh did information science of [2:17] engineering Then I switched to [2:18] philosophy. I think in that intersection [2:21] actually I work uh a lot of in digital [2:24] and AI governance but also those [2:25] questions deeper question. I'm also a [2:27] visiting professor at the Chinese [2:29] Academy of Science in the aerospace [2:31] institute there. So now to to have [2:34] actually these are not competing uh uh [2:38] dimensions but maybe to have some [2:40] categories uh when you engage or when I [2:43] engage in this discussion to locate uh [2:47] you know what could it mean uh [2:49] consciousness and then maybe what could [2:51] it then mean if I can become conscious I [2:54] think this is how I or how maybe you [2:56] should take those categories I'm [2:58] suggesting like as you know frames or [3:00] point of departure in getting into it [3:02] and I don't want to present them as [3:05] competing uh but some are certainly [3:08] competing categories but more as [3:10] different perspective just because we [3:13] don't know right that's uh very [3:15] important here uh and I would start uh [3:19] with the emergent one [3:21] and uh go straight into basically the [3:25] brain if you wish so the emergent [3:28] perspective [3:30] physicalist realism is the dominant [3:33] view. So basically here uh the mental [3:36] states they emerge from meta from the uh [3:39] from the genomics from the [3:41] connetoonomics uh from the electrical [3:44] processing in the brain. So the mind our [3:46] mind is secondary. It's an epiphen [3:48] phenomenon. Um and it actually was long [3:52] ignored the mind or consciousness in in [3:54] research. Um because I the observer can [3:58] know reality but I the observer is not [4:00] part of my analysis of that reality. So [4:03] it was excluded. And he is also in this [4:06] let's say tradition in this materialist [4:08] tradition um in this uh um realist [4:12] tradition basically where I can know the [4:15] reality what it is I can measure it I [4:17] can observe it I can um I can predict it [4:20] uh came up this hard problem actually [4:23] the hard problem is not just from David [4:24] chararma it the hard problem is known [4:26] but I think it's not the full story of [4:28] the issue but the hard problem basically [4:30] says how can from something which is [4:32] quantitative [4:34] become equality. Uh how comes that that [4:38] the sum that the whole is more than the [4:40] sum of the parts that some physically [4:43] processes through the dynamics create an [4:46] identity that pursues goals and has a [4:51] plan in life etc. So from less becomes [4:54] more. So from non-concious [4:57] units they become conscious. That's a [4:59] that's unexplained. So that is the [5:01] basically the hard problem and when then [5:04] you ask a question from this perspective [5:06] on on the left side and sorry there's a [5:08] a lot always on those slides I try to [5:11] you know just to try to provide um as [5:14] many as information as possible um [5:18] basically would say could I could I [5:22] could I simulate consciousness uh like [5:25] on a machine or somewhere it's basically [5:27] like kind of a no but we still tried [5:29] doing it anyway Right. So, uh it's kind [5:32] of not possible but uh there's the [5:35] mainstream then um idea like let's say [5:38] neuroscience there are no extra [5:40] non-physical realities. Um so the mind [5:43] is reducible to physical [5:47] processes. So if I search hard enough I [5:50] might then basically at one point maybe [5:52] make something conscious. So that's the [5:54] then there's something on the they look [5:56] at on the quantum level like sir Peter [5:57] at sir sir penrose uh they say oh this [6:01] is uh on the quantum level in the [6:03] tubules at the brain with inside the [6:05] neuron and uh basically uh we can't [6:08] compute it we don't know but we're still [6:10] doing the research so this is a kind of [6:12] a yes no but um but any of this approach [6:16] bypasses the hard problem right how can [6:19] something from less more emerge [6:22] and when Um [6:25] what actually emerges is this I think I [6:28] am the kito of decar [6:31] and um kind of antithetical to the [6:34] realist approach is a universal approach [6:36] universalist approach or the idealism [6:39] and I myself I'm German I'm from Berlin [6:41] Germany is kind of the home from German [6:44] idealism and if you always have tried to [6:47] understand why what is the essence of [6:49] Emanuel Kant is very easy there's only [6:51] subjectivity [6:53] There is no objective reality. There's [6:56] only subjectivity. The question is how [6:57] we still get some objective truth out of [6:59] this. But when we look at the topic of [7:03] um what is consciousness, then there's a [7:05] little bit of reversal because [7:08] from a universalist perspective, you [7:10] don't have the problem that you have uh [7:13] units of uh uh consciousness and then [7:15] they merge into a single stream of [7:17] awareness. Well, how is that possible? [7:19] Because probably it's not possible. They [7:21] say there must be a consciousness at [7:23] large, a universal mind. And uh what [7:26] does it mean? Basically that our mind, [7:28] our individual consciousness is a [7:30] dissociation of the universal mind. So [7:32] at birth it dissoc dissociates from the [7:36] universal mind. Um it creates kind of an [7:40] instance, an alter. Uh it's not a [7:42] separate reality. It's just kind of we [7:44] just interface for that. And upon death, [7:48] it's not that our mental process stop. [7:51] It reassociates with the mind at large. [7:54] And actually, it's just the end of the [7:56] illusion of being separate. Is the end [7:58] of the illusion of having a self for [8:00] those who meditate, for those who do uh [8:04] try you know all kinds of things. Uh [8:06] sometimes you try of course to deego to [8:09] desel yourself and you might have this [8:12] moment. And I think this this is kind of [8:14] the thinking behind that. It's of course [8:16] empirically very hard to test [8:19] scientifically but that's approach. So [8:21] here actually the hard problem is not uh [8:27] why does a mind emerges from matter but [8:32] here actually the hard problem is [8:34] inbaked in the in the premise itself. [8:37] Why a subjective unit of experience is a [8:41] condition for a coherent experience of [8:43] the object. So everything we perceive as [8:46] rational objective is on the layer of [8:47] the subjective. [8:49] um [8:51] uh from that and I tried to capture it [8:53] in this picture actually that the apple [8:55] you see is the appearance right and has [8:59] maybe somewhere in this unknown space [9:02] some equivalent we can't access we can't [9:05] know because it's a matter of resolution [9:07] when you go into quantum level you will [9:08] not see the apple anymore right so what [9:10] what is there we don't know basically [9:12] and what's happening in your brain is a [9:14] representation of that appearance and to [9:18] And our subjectivity combines this into [9:21] a single stream. And our subject escapes [9:24] that. So it must be somewhere. Our [9:26] subject must be somewhere in this [9:27] unknown space. That's the reason why we [9:29] can't access it. It escapes basically. [9:31] And what does science do? Science tries [9:33] to describe the apple but can only do in [9:37] a negative way hypothesis verification. [9:39] It cannot designate the apple in a [9:42] positive sense. it can only uh make [9:46] assumptions and then say and a rule out [9:48] to to to refuse the null hypothesis. So [9:50] only in the negative way. So that's [9:52] quite telling because the research [9:55] always stands on the subject which [9:57] itself is not taken into account. When [10:00] you ask now from a universalist [10:02] perspective can this be computed is a [10:06] category no because the subject actually [10:10] or the spacetime which is the basic [10:13] categories is already a construct. [10:16] there's no spaceime and actually you [10:18] have approaches in in quantum uh in [10:21] quantum physics uh which are called [10:23] positive geometry or high energy uh [10:27] theories which go post space-time [10:30] approaches. So if the subject already [10:32] presupposes spacetime [10:35] there must be something beyond that [10:36] right. So you can't compute it. So [10:38] because every approach to compute [10:41] consciousness implies already [10:42] consciousness. So from this perspective [10:45] is a no or makes it also very hard to [10:47] actually uh to uh test. [10:51] Then the multiscale approach [10:55] this is a this is a quite interesting [10:57] this is the most empirical one I would [10:59] say we have. It is kind of in between I [11:02] I explained in briefly what you see [11:04] there in this video. Uh you it's kind of [11:06] in between [11:08] uh the realist and the universalist but [11:11] it leans more towards the realist one [11:13] the first one we explained the empirical [11:15] one. Um and it does it as follows. It [11:18] puts actually agency as a center of [11:20] observation. So and a mind is a is an [11:23] agent basically and you have this agency [11:26] this mindful activity already on a not [11:30] just on the brain level on the [11:31] subcellular level. So when cells [11:33] organism try to to solve problems and [11:36] they're very good problem solvers and [11:38] each cell alone will not do it but in [11:41] the collective so mind is decentralized [11:44] coordinated collective that will sort [11:45] out the problem and sometimes the [11:48] individual cell does not know it will [11:49] become a hand. So in this collective [11:52] they would say okay it will become like [11:53] that but more interesting what you see [11:55] here on this on this image it's a leaf [11:58] and um I think this is this is not from [12:01] an oak tree but there are different kind [12:02] of leaves and then you have insects uh [12:05] and this guy was a little it's a little [12:07] insect [12:09] um and it it puts it it puts some uh [12:12] some liquid tries to inject it into the [12:14] leaf and actually [12:17] this red u whatever it is uh [12:21] uh organism you see there is not in the [12:24] DNA of the leaf. But how comes that the [12:27] leaf is producing that? So basically the [12:31] this little this little insect they they [12:34] know [12:35] where to inject this liquid in order to [12:39] use the leaf the mechanisms uh the [12:42] proteins to build something. But the [12:44] structure the form is not stored in the [12:46] leaf and we don't know where it's [12:48] stored. And actually one who's [12:50] researching this a lot is Michael [12:51] Leaven. Uh maybe heard some of you have [12:53] heard about him. And he uh he thinks [12:56] first that uh and you see this on this [12:58] picture now that actually there must be [13:01] maybe this kind of platonic space of [13:03] forms and minds uh which is pulled in [13:08] basically when on a very micro level [13:12] already uh problems are solved. Yeah. [13:16] And we don't know it's a for for an [13:18] empirist who does empirical biological [13:20] or biochemistry research it's quite a [13:22] big leap to say there's a platonic space [13:25] right so we don't know but it's a but [13:27] they found out they can manipulate cells [13:29] on you know this iron channels basically [13:32] uh and um say to this worm grow two [13:35] heads then the worm grows two heads and [13:36] so and so forth so they can do it so [13:38] it's not about printing a heart it's [13:40] just to instruct the cell make a heart [13:43] that's quite it's a it's quite a hack so [13:45] I think This confirms a little bit the [13:47] emergent approach from the first one [13:49] that actually the intelligence is [13:51] somehow still connected to the physical [13:53] reality. [13:54] But then he also says that some of the [13:56] problem solving must comes from [13:58] somewhere else. We know that our that [14:01] biology uses natural laws in chemistry [14:05] in physics um uh and borrows that. But [14:09] there must be maybe more to this uh with [14:12] solving problems um and and and that's [14:15] quite interesting. But from his [14:17] perspective, if you ask a question, [14:19] could we actually simulate from this [14:20] perspective consciousness? He would say [14:24] um no. But what we could do because he [14:28] also sees that the uh that the uh the [14:33] mind operating let's say on a cellular [14:36] subcellular uh multisellular levels [14:39] basically are deeply ingrained physical [14:41] process. So this is one problem and the [14:45] other problem is this platonic space. So [14:48] how do we simulate that space? Does [14:50] today an LLM has access to this platonic [14:53] space? actually he's not sure whether [14:56] those LLMs uh have such capacity. Uh but [14:59] you could also say it's just a [15:01] complexity issue. So if complex systems [15:03] make all kinds of things but that [15:05] doesn't mean there is a kind of a [15:07] mystical secret place behind that. [15:10] Uh the the the fourth one now [15:13] I think that's pretty straightforward. [15:16] It basically says yes you can simulate [15:19] that. And actually I have to say um that [15:22] this is probably the the least [15:24] pretentious approach in terms of seeing [15:28] humanity [15:29] special in this universe. Right? The [15:31] other one when we when we say oh [15:33] consciousness uh maybe it's a it's a [15:36] base unit of quantum physics put [15:38] actually humanity in the center of the [15:40] universe. It's quite pretentious to do [15:41] that right maybe it's right maybe it's [15:43] wrong but here it's pretty [15:44] straightforward. It's a functionalist [15:47] approach. So it's not asking what is the [15:50] mind but what does it. So here in this [15:53] in this perspective basically it can be [15:56] simulated uh on a substrat which [15:59] supports or has the properties where a [16:02] mind can run. It's not an emergent [16:04] approach where the meta produces it. It [16:08] is approach no you have the simulation [16:11] and which meta fits for me. It's like [16:13] running an LLM and an H00 Nvidia chip is [16:17] better than running it on a H20 chip or [16:20] whatever. So in that way it's like a uh [16:23] playing Minecraft. So when you open the [16:25] computer you will not see the Minecraft [16:27] but it still runs as a software. So [16:29] that's the assumption uh behind that. So [16:32] basically then you look at it then it [16:35] forces you really to describe what is [16:37] consciousness from a phological [16:39] perspective but also from a functional [16:41] one. From a phological perspective is a [16:43] stream of presence a continuous coherent [16:46] nowness uh including the space of space [16:49] and space time and space uh and has also [16:54] sensory content. Um it is a kind of a [16:57] second order perception. [16:59] uh we sense ourselves sensing actually [17:02] the eye cannot see itself seeing right [17:05] so that's a that's a um that that's a [17:07] big difference with the brain being a [17:09] sense organ but can self sensing and [17:12] from a functional perspective basically [17:14] a very important so that we build the [17:16] self the self and the the world model [17:19] are basically instances separate [17:20] simulation in this respect uh the [17:23] function basically uh it would be that [17:26] the consciousness is a coherence [17:28] inducing. So we need coherence. [17:32] Um colonizing or self-organizing and [17:35] colonizing. Colonizing why colonizing if [17:37] something breaks down the plasticity [17:39] like in the brain takes over something [17:41] else and you see it also in the brain. [17:43] So there are some parts of brain are [17:44] specialized but other could take it [17:46] over. That's very interesting right? So [17:48] that can take over and in in this way [17:50] you should you must model it in the same [17:52] way. Uh it's a significant filter. The [17:55] consciousness is about significance. So [17:58] if something is insignificant you you [18:00] feel yourself you get bored you actually [18:02] shut down. So, so the so it's very [18:06] important how the consciousness works. [18:08] It looks at significant things and the [18:10] significance here that's a little bit [18:12] tricky. Uh is of course it's charged by [18:15] values you know it's it's not a a [18:18] natural process. Um and um uh yeah so [18:23] the mind met matter relations as said it [18:25] just runs on a substrate or substrat [18:28] independent. So it could actually run on [18:30] any substrate if the substrate has a [18:32] properties to support this those [18:34] functions of the mind. Right? So um [18:39] obviously this approach contradicts all [18:41] other views. It makes it a little bit [18:43] suspicious. [18:45] So but it's actually forces you to [18:47] really think through what it is. uh and [18:49] I would say from my perspective actually [18:52] that artificial intelligence here is not [18:55] just a question whether it becomes [18:56] conscious but it's a great tool to test [18:59] those empirically because with which [19:01] tools can you test this question there [19:03] are not many right so in the in the [19:05] universal you can only philosophize [19:08] about that but you can you can get quite [19:10] far but actually here you can even try [19:13] to overcome the bridge or the gap [19:15] between mathematics and experience [19:17] because mathematics cannot capture [19:19] experience and because math is a [19:22] stateless system and the mind is not [19:25] stateless it's very dynamic or the [19:27] organism is very dynamic and you could [19:29] actually model such dynamism with AI so [19:32] in this way actually uh AI is very [19:34] interesting [19:36] before concluding uh one maybe to locate [19:41] uh where we are with AI [19:44] uh basically on the left side you see [19:46] the human and we have this [19:48] representation, we have the phenomenon [19:51] and we have the real on the bottom. The [19:53] real we cannot access, right? Um and the [19:56] subject, the subjectivity embraces this [20:00] phenomenon on the representation into a [20:02] coherent and and and maybe our [20:06] subjective core is somehow located in [20:09] the unknown. That's the reason why we [20:10] have so problem to access that and [20:13] actually to access it could be [20:15] irreducible. That would mean for some [20:17] quantum physicists they would say we [20:19] have to calculate the whole universe [20:21] like in brute force step by step in [20:23] order to model consciousness should [20:25] consciously be a unit a core unit of the [20:28] universe right um so if there's no for [20:32] for for the for the other for the other [20:35] approach basically uh for the simulated [20:37] approach I would say ah uh learning [20:40] presupposes mind and learning starts [20:43] very early So consciousness could be a [20:47] very simple algorithm because it needs [20:51] to and you know you usually nature and [20:53] universe is very resource efficient [20:55] right so why why should it why do we [20:58] need all this compute maybe there's a [21:00] really resource efficient um aspect in [21:02] that so when you look now at the on the [21:04] right side the the different AIS [21:08] now the the disembodied the the LLMs [21:11] basically they only operate on the [21:13] representation [21:15] uh there these are representation of [21:18] representations. Language it did not [21:20] learn from the from the phenomenon from [21:24] having sounds, gestures and writing [21:26] where you then build actually engage [21:29] with physical reality but it just [21:31] learned abstracted data top down. So [21:32] it's a complete different way um how we [21:35] learn language. It's actually physically [21:37] ingrained uh and the LM is not right. So [21:41] it operates only on this disembodied [21:44] representational layer. However, now we [21:47] add coming to this embodied AI into a [21:50] sensory AI. What happens now that [21:53] actually the the phological layer in the [21:56] middle is added because it will sense [22:00] the world and it will also alter the [22:03] world. Actually it it builds a model of [22:07] appearance. So it already has both. But [22:10] where we combine both levels in our [22:14] subjectivity, the the gap now between [22:18] the representation and the phenomenon is [22:20] just a functional layer, right? So it's [22:23] actually and Jeffrey Hinton is quite [22:26] concerned about this kind of [22:27] subjectivity. So these AIs develop a [22:29] subjectivity, but it is an artificial [22:31] subjectivity. It is not a conscious one. [22:35] Very important. And the on the right [22:37] side this is the speculation what [22:39] happens actually if it is embodied and [22:41] conscious. [22:43] And uh [22:46] to conclude I will not go to everything [22:49] to conclude there are two things where [22:51] you maybe measure consciousness whether [22:53] like artificial consciousness one is on [22:56] the behavior side and the other one is [22:58] on the substrate side. on the behavior [23:01] the system will do everything which you [23:03] say well this is conscious right and I [23:05] always a little bit cynical say if you [23:07] have a killing robbert behind you you [23:08] don't ask a question you say this thing [23:10] is conscious is after me it will try to [23:13] get me but you see today already I think [23:16] this was what L&M's cannot really do I [23:19] mean was supposed to do only by [23:21] humankind right [23:23] the second layer is the substrat so in [23:26] order to really know not that the system [23:27] is cheating me it's just imitating [23:29] consciousness. Actually, our brain must [23:32] connect to that system to feel whether [23:35] it's really conscious. And that's a [23:37] little bit an issue, but I think that [23:39] would be the ultimate test whether a [23:40] system is waking up or not. uh and um [23:44] let me conclude [23:46] uh I think [23:49] uh [23:50] I I I think I believe uh I I don't know [23:54] whether I want this system to become [23:56] conscious or not but I think if it has [23:59] some simulation of self-awareness some [24:01] deep simulation of self-awareness and [24:03] it's very important to to kind of [24:05] simulate what it could all mean maybe [24:09] it's easier then to understand us what [24:12] for us important inside us uh and to [24:15] relate better to what is important [24:17] because the problem is with morality [24:19] even morality is universal so you don't [24:22] deceive you don't do those things these [24:24] are universal values [24:27] uh but they're always contextual [24:30] in China is different than in the United [24:32] States or in Europe or where else we [24:35] share probably same values but they [24:36] implemented in a different way in a [24:38] cultural culturally historically [24:40] mediated [24:41] And uh this you cannot hardcode [24:45] uh a line. So that's the reason why [24:47] actually it is maybe not bad if a system [24:50] become conscious in that way if it has [24:52] already this super intelligence. Thank [24:54] you very much.