Prisoner’s Dilemma and the Computational Boundary of Self

• Traditional Prisoner’s Dilemma simulations have a fixed number of players who can cooperate or defect.
• Biology is more complex: biological entities (cells, tissues, etc.) can merge and split, changing the number of “players” and thus the payoff matrix. This introduces a dynamic aspect absent in standard models.
• Merging provides benefits, particularly with transfer of metabolic data and the ability to erase personal ‘memory’ which provides group benefit.
• Levin’s computational boundary of self is a framework for understanding diverse intelligences on a single scale, regardless of brain structure, environment, or scale (molecular to planetary).
• The core idea: All intelligent agents share the ability to pursue goals, some simple, some complex.
• The framework maps agents based on the size (in space and time) of the largest goal they can pursue (their “cognitive light cone”). A bacterium’s goal might be local sugar concentration; a human’s might be world peace.
• This framework is designed for empirical research: It allows for testable hypotheses about an agent’s goals, the problem space it operates in, and its competencies (how well it achieves goals).

Diverse Intelligence, Naturalizing Cognition, and Objections

• Goal of DI is *not* to anthropomorphize, but to *naturalize* cognition. To understand thinking processes outside of a narrow, human-centric definition.
• Levin’s ideas receive criticism from both reductionists (who dislike agential language) and organicists (who dislike including machines on the same spectrum as living beings).
• Reductionists often argue for explanation at the molecular level and see agential talk as pre-scientific.
• Organicists often want a sharp distinction between machines and living organisms, fearing a loss of respect for life.
• Levin argues for a *continuum* between matter and mind, with different tools applicable at different points on the spectrum. He emphasizes the need for empirical research to determine which tools are appropriate for which systems.
• This idea goes against how many scientists currently try to delineate between consciousness versus no-consciousness or life versus non-life and living cells vs non-living cells.

Consciousness and Action

• Levin prioritizes studying observable, functional behavior (problem-solving, intelligent behavior) before tackling the “hard problem” of consciousness.
• He suggests that while the *sensory* aspect of consciousness (what it *feels like*) is important, the *actuation* aspect (what it’s *like to do*) is often underemphasized.
• He points out the asymmetry in theories of mind: Epiphenomenalism posits real sensory states but denies their causal efficacy. There’s no common equivalent view denying the reality of sensation but affirming the reality of free will (action).
• The need to act – to choose a *next action* – is fundamental to being an agent and defining the boundary between self and the outside world.
• Levin is currently working on writing about consciousness, planning to address these ideas more directly in the future (likely in 2024). He believes studying consciousness directly *changes* the observer (it’s not purely third-person research).
• He says it is likely there is no definition for consciousness.

TAME and Relationships, Not Just Control

• TAME (Technological Approach to Mind Everywhere) is an *engineering* framework. Engineering prioritizes control (predicting and controlling a system).
• TAME 2.0 is in development, to quantitatively flesh out the “cognitive light cone” concept.
• Control may be thought of with the engineering aspects of TAME in mind.
• Beyond control, relationships are important. With more complex agents (further right on the spectrum), interaction becomes bi-directional, not just one-way control. The appropriate “way to relate” changes.
• “Proof of humanity” certificates (relevant in the age of AI) might ideally guarantee a certain *capacity for compassion*, an alignment of “cognitive light cones” – caring about the same scope of things.
• People often think of anatomy or genome to verify humanity but that does not necessarily give people compatibility.
• Compatibility might be about shared existential concerns (the challenges faced in existing) more than shared anatomy or genome.
• He proposes that a compatible match in this framework, between a set of humans and a machine in this example, it requires at minimum some alignment.

Teleology, Evolution, and the “Meaning of Life”

• A major concern with AI/technology is that we might be superseded. But this concept already exists: our kids/children. This concept has been long established and realized.
• Teleophobia: Many scientists avoid discussing goals or purpose (teleology), often seeing it as unscientific or pre-scientific.
• Levin argues teleology is acceptable now that we have cybernetics and control theory – a science of machines with goals (e.g., a thermostat).
• These help deal with goals mathematically.
• He uses *teleonomy* to emphasize that goal-directedness is *apparent* – it’s a *lens* from the perspective of an observer, a hypothesis to be tested.
• Levin supports a form of panpsychism that reformulates basic physics as a proto-cognitive process (akin to ideas of Chris Fields and Karl Friston) – a deeper reality underlying both simple systems and complex minds.
• Levin supports the idea there may exist ‘proto-cognitive processes.’
• He requires this to contain ’empirical evidence’ that explains ‘the underlying system’.
• He believes biological evolution doesn’t optimize for things humans value (happiness, meaning, etc.); it’s a random search, settling on what’s “good enough” to survive, increasing biomass.

Additional Points

• Levin believes biology is incredibly adaptable and, with some help, humans will likely be able to live in environments like Mars.
• All intelligence is collective, meaning no complex agent could have learning capacity because the agents of their individual components make that possible and are fundamental to cognition and processing information.
• Levin sees phase transitions (in biology, and elsewhere) as partly dependent on the perspective/formalism used, and the time scale considered. They might *appear* sharp at one level but be smooth transitions at a finer level.
• He has no direct evidence for other organs using language in the same complex, formal way as the brain.
• The way orthopeadic surgeons use tools during orthopeadic procedures show a certain “mechanical aspect” of the human body.
• He is excited about ongoing empirical work in limb regeneration (currently in mice), bioelectric approaches to cancer, and work on synthetic organisms.
• A self may be defined by a model.
• He views humans as “amazing, remarkable, ethically important, morally valuable, spiritual machines.”
• Levin emphasizes the importance of thick skin and focusing on one’s own goals in navigating the sometimes hostile environment of academic science, especially when pursuing unconventional ideas.
• There exists a concept of Metacognition all the way down which is useful.
• “Where do I end and the outside world begin?” may provide utility in conceptualising the human body and it’s function.
• Nested hierachies which have smaller and less complex agents as parts allow a more-complete entity/body to thrive.
• He notes “you should get better as you get older”.

囚徒困境与自我的计算边界

• 传统的囚徒困境模拟具有固定数量的参与者，他们可以选择合作或背叛。
• 生物学更为复杂：生物实体（细胞、组织等）可以合并和分裂，改变“参与者”的数量，从而改变收益矩阵。这引入了标准模型中不存在的动态方面。
• 合并提供了好处，特别是在代谢数据的转移和擦除个人“记忆”的能力方面，这提供了群体利益。
• 莱文的自我计算边界是一个框架，用于在单一尺度上理解各种智能，无论大脑结构、环境或尺度（从分子到行星）。
• 核心思想：所有智能体都具有追求目标的能力，有些简单，有些复杂。
• 该框架根据智能体可以追求的最大目标（其“认知光锥”）的大小（在空间和时间上）来映射智能体。细菌的目标可能是局部糖浓度；人类的目标可能是世界和平。
• 该框架是为实证研究而设计的：它允许对智能体的目标、其运作的问题空间及其能力（它实现目标的程度）进行可测试的假设。

多元智能、认知自然化和反对意见

• 多元智能的目标*不是*拟人化，而是*自然化*认知。理解狭隘的、以人类为中心的定义之外的思维过程。
• 莱文的观点受到还原论者（他们不喜欢自主性语言）和有机论者（他们不喜欢将机器与生物放在同一光谱上）的批评。
• 还原论者通常主张在分子水平上进行解释，并将自主性言论视为前科学的。
• 有机论者通常希望在机器和生物体之间划清界限，担心失去对生命的尊重。
• 莱文主张物质和心灵之间的*连续体*，不同的工具适用于光谱上的不同点。他强调需要实证研究来确定哪些工具适用于哪些系统。
• 这个想法与许多科学家目前试图区分意识与非意识、生命与非生命以及活细胞与非活细胞的方式背道而驰。

意识与行动

• 莱文优先研究可观察的功能行为（解决问题、智能行为），然后再解决意识的“难题”。
• 他认为，虽然意识的*感官*方面（它的*感觉*是什么）很重要，但*行动*方面（它的*行动起来*是什么样子）常常被低估。
• 他指出了心灵理论中的不对称性：副现象论假定真实的感官状态，但否认它们的因果效力。没有常见的等效观点否认感觉的现实性，但肯定自由意志（行动）的现实性。
• 行动的需要——选择一个*下一个行动*——是成为一个主体并定义自我与外部世界之间界限的基础。
• 莱文目前正在撰写关于意识的文章，计划在未来（可能在 2024 年）更直接地解决这些想法。他认为直接研究意识会*改变*观察者（这不仅仅是第三人称研究）。
• 他说意识可能没有定义。

TAME 与关系，不仅仅是控制

• TAME（无处不在的心智技术方法）是一个*工程*框架。工程优先考虑控制（预测和控制系统）。
• TAME 2.0 正在开发中，以定量地充实“认知光锥”概念。
• 可以从 TAME 的工程方面考虑控制。
• 除了控制之外，关系也很重要。对于更复杂的智能体（在光谱上更靠右），交互变得双向，而不仅仅是单向控制。适当的“关联方式”会发生变化。
• “人类证明”证书（在人工智能时代相关）可能理想地保证一定的*同情能力*，即“认知光锥”的对齐——关心相同范围的事物。
• 人们通常认为解剖学或基因组可以验证人性，但这并不一定能让人类彼此相容。
• 相容性可能更多的是关于共同的存在关注（存在的挑战），而不是共同的解剖结构或基因组。
• 他提出，在这个框架中，一组人类和一台机器之间的兼容匹配，在这个例子中，至少需要一些对齐。

目的论、进化和“生命的意义”

• 人工智能/技术的一个主要担忧是我们可能会被取代。但这个概念已经存在：我们的孩子/后代。这个概念早已确立并实现。
• 目的恐惧症：许多科学家避免讨论目标或目的（目的论），通常认为它是不科学的或前科学的。
• 莱文认为，既然我们有了控制论和控制理论——一门关于具有目标的机器的科学（例如，恒温器），目的论是可以接受的。
• 这些有助于用数学方法处理目标。
• 他使用*目的论*来强调目标导向是*明显的*——它是观察者视角中的一个*透镜*，一个有待检验的假设。
• 莱文支持一种泛灵论的形式，它将基础物理学重新表述为一种原认知过程（类似于 Chris Fields 和 Karl Friston 的思想）——一种更深层的现实，是简单系统和复杂心灵的基础。
• 莱文支持可能存在“原认知过程”的观点。
• 他要求这包含解释“底层系统”的“经验证据”。
• 他认为生物进化并不针对人类重视的事物（幸福、意义等）进行优化；这是一种随机搜索，满足于“足够好”以生存，增加生物量。

补充要点

• 莱文认为生物学具有极强的适应性，并且在一些帮助下，人类很可能能够在火星等环境中生存。
• 所有智能都是集体的，这意味着没有复杂的智能体可以具有学习能力，因为它们各个组成部分的智能体使这成为可能，并且是认知和处理信息的基础。
• 莱文认为相变（在生物学和其他地方）部分取决于所使用的视角/形式主义以及所考虑的时间尺度。它们可能在一个级别上*出现*尖锐，但在更精细的级别上是平滑的过渡。
• 他没有直接证据表明其他器官像大脑一样以复杂、正式的方式使用语言。
• 骨科医生在骨科手术中使用工具的方式显示了人体的某种“机械方面”。
• 他对正在进行的肢体再生（目前在小鼠身上）、癌症的生物电方法以及合成生物体方面的工作感到兴奋。
• 自我可以由模型定义。
• 他将人类视为“令人惊叹、非凡、具有伦理重要性、道德价值的精神机器”。
• 莱文强调了在学术科学有时充满敌意的环境中，要有厚脸皮并专注于自己的目标的重要性，特别是在追求非常规想法时。
• 存在一个一直向下延伸的元认知概念，它很有用。
• “我在哪里结束，外部世界从哪里开始？”可能有助于概念化人体及其功能。
• 具有更小、更不复杂的智能体作为部分的嵌套层次结构允许更完整的实体/身体茁壮成长。
• 他指出“随着年龄的增长，你应该变得更好”。