Introduction and Overlapping Ideas

• Levin appreciates Bach’s breadth in tackling computation, cognition, AI, and ethics. Bach finds overlap with Levin’s work on cellular communication and agency, particularly the idea that every cell can act like a neuron.

Cells as Agents and Information Processors

• Every cell can send and receive multiple message types, conditionally, and can learn. Each cell is a reinforcement learning agent, primarily getting rewards from its environment.
• Neurons are specialized “telegraph cells,” extending cellular communication over long distances with high-speed signals, crucial for animal movement and competition.
• The brain can be viewed as a “telegraphic extension” of the body’s cellular community, with the potential for every organism to become intelligent given enough time and shared genetic destiny.
• Cells posses turing complete abilities that allow them to behave with proactive control, making arbitrary internal represenatations indepedent from external influence.

Cross-Disciplinary Boundaries and Science

• Levin and Bach note how disciplinary boundaries in science can be protective and hinder interdisciplinary work, particularly through peer review, limiting paradigm shifts.
• An “engineering stance,” common in computer science, focuses on causal patterns and implementation, which is often missing in other fields, like neuroscience.

Critique of Current Neuroscience and Alternative Models

• Current machine learning, inspired by simplified perceptrons, doesn’t accurately reflect how biological cells organize, which is from the inside out, not through external weight updates.
• Local self-organization by reinforcement agents trading rewards offers a fascinating perspective, missing from current AI.
• The emphasis on disciplines prevents sharing insights. for example: In a neuroscience department, it is known information processing can occur through APs before gene express. yet, this may cause surprise and resistance in a molec bio group.

RNA-Based Memory Transfer and Its Implications

• Experiments (McConnell, Ungar, Glantzman) suggest memory transfer via RNA or peptides, challenging the synapse-centric view. This includes work with planaria, rats, and metamorphosing insects.
• This concept raises puzzles about decoding transferred information, especially for arbitrary, non-evolutionarily relevant memories. How does a recipient brain interpret an arbitrary RNA structure?
• An implication is that the Connectome Project may not be able to map concioussness if memory is transferred using methods and data different than physical axon connections.

Evolution and Planaria

• In planaria, an ability of an animal’s system is to recall “previous settings”, meaning biological information for an organism persists even though massive reorganization is undertaken such as loss of an entire brain.

Competency, Goals, and Constraints

• “Competency” is an engineering claim about a system’s ability to navigate a problem space toward a goal, dependent on the observer’s perspective.
• Biological systems have feedback loops to reach specific outcomes in anatomical morphospace, demonstrating competencies like recruiting cells.
• Goals can be emergent or explicitly represented, as seen in planaria’s bioelectric pattern memory, which can be rewritten (two heads instead of one).
• Constraints satisfaction: Organisms strive to move the universe’s state space towards acceptable regions (e.g., having one head), navigating substrate and functional constraints.

Multi-Scale Competency Architecture Again

• Evolution may struggle to make a genome since evolution has issues judging the “fitness” if competenct organism manage errors in-vivo using “built in software/algorithms” instead of genes, in particular with an orgamisms such as Planaria that asexually reprocduce, circumventing the normal filters, the result being a fit organism despite its genome looking horrible on paper.
• Analogy: computers which has a drive prone to errors where the software corrects for mistakes in-vivo. In computer-speak: a ‘RAID setup’.

Embryonic Development and Self-Organization

• An amniote embryo starts as a disk of cells, not inherently one individual, but potentially multiple. Symmetry breaking determines the organizer, leading to one embryo, or conjoined twins if disrupted.
• Biological systems self-construct, determining their boundaries, structure, and relevant problem spaces, unlike pre-defined robots. They are energy-limited and must choose a “lens” to coarse-grain the world.
• Planeria cells decide to follow, spatially, by gradient and other biological information cues for correct development by looking at what neighbours cells are doing: deciding “local spatial difference” cues in a cell rather than taking explicit external instrcutions, unlike an AI agent trained and designed from top down, controlled and influenced.

Implications for AGI and Collective Intelligence

• Sufficient condintions include cells connected and signalling rewards with reliability over enough units.
• Conditions for the emergence of general intelligence include: 1. Units as small agents with expectation of minimizing future target deviations. 2. Units connected and exchanging rewards or proxy rewards.
• The question of whether these biological insights can be to be scaled up is the current task. For example: Twitter and Global-Scale social Media interaction, and testing incentive structures (for exaple with Elons’ Twiter experiement, or with societies on a social/governmental scale) can have their group agency steered, where cells/units within can become grouped into emergent control behaviours through self organization.
• Twitter is explored as a potential global brain, highlighting emergent agency and the challenge of designing incentive structures for collective intelligence. This relates to governance in brains, societies, and social media.

导言与思想交融

• 莱文赞赏巴赫在计算、认知、人工智能和伦理学方面的广度。巴赫发现莱文关于细胞通讯和自主性的工作存在思想交融，特别是每个细胞都可以像神经元一样运作的观点。

细胞作为自主体和信息处理者

• 每个细胞都可以有条件地发送和接收多种类型的消息，并且可以学习。每个细胞都是一个强化学习自主体，主要从其环境中获得奖励。
• 神经元是专门的“电报细胞”，将细胞通讯以高速信号扩展到长距离，这对动物运动和竞争至关重要。
• 大脑可以被视为身体细胞群体的“电报延伸”，只要有足够的时间和共享的遗传命运，每个生物体都有可能变得智能。
• 细胞拥有图灵完备的能力，使它们能够以前摄性控制的方式行动，进行独立于外部影响的任意内部表征。

跨学科界限与科学

• 莱文和巴赫注意到，科学中的学科界限可能是保护性的，并且通过同行评审阻碍跨学科工作，限制了范式转变。
• 计算机科学中常见的“工程立场”侧重于因果模式和实现，这在神经科学等其他领域往往缺失。

对当前神经科学的批判和替代模型

• 当前受简化感知器启发的机器学习，并没有准确反映生物细胞如何从内而外组织起来，而不是通过外部权重更新。
• 通过交换奖励的强化自主体的局部自组织提供了一个引人入胜的视角，这在当前的人工智能中是缺失的。
• 对学科的强调阻碍了洞见的分享。例如：在神经科学系，众所周知，信息处理可以在基因表达之前通过动作电位发生。然而，这可能会在分子生物学组中引起惊讶和抵制。

基于RNA的记忆转移及其影响

• 实验（McConnell、Ungar、Glantzman）表明记忆可以通过RNA或肽转移，挑战了以突触为中心的观点。这包括对涡虫、大鼠和变态昆虫的研究。
• 这个概念提出了关于解码转移信息的难题，特别是对于任意的、非进化相关的记忆。接收者大脑如何解释任意的RNA结构？
• 一个启示是，如果记忆使用不同于物理轴突连接的方法和数据进行转移，连接组计划可能无法映射意识。

进化与涡虫

• 在涡虫中，动物系统的一个能力是回忆“先前的设置”，这意味着生物体的生物信息即使在进行大规模重组（例如失去整个大脑）后仍然存在。

能力、目标和约束

• “能力”是关于系统在问题空间中朝着目标导航的能力的工程断言，取决于观察者的视角。
• 生物系统具有反馈回路，以在解剖形态空间中达到特定结果，表现出诸如募集细胞的能力。
• 目标可以是涌现的或显式表示的，如涡虫的生物电模式记忆所示，它可以被重写（两个头而不是一个）。
• 约束满足：生物体努力将宇宙的状态空间移动到可接受的区域（例如，有一个头），导航基质和功能约束。

再谈多尺度能力架构

• 进化可能难以产生基因组，因为如果能力强的生物体使用“内置软件/算法”而不是基因在体内纠正错误，特别是对于无性繁殖的涡虫等生物体，绕过正常过滤器，进化很难判断“适应性”，结果是适应性强的生物体，尽管其基因组在纸面上看起来很糟糕。
• 类比：计算机的驱动器容易出错，其中软件在体内纠正错误。用计算机术语来说：一个“RAID 设置”。

胚胎发育与自组织

• 羊膜胚胎开始时是一个细胞盘，本质上不是一个个体，而是潜在的多个个体。对称性破缺决定了组织者，导致一个胚胎，或如果中断则导致连体双胞胎。
• 生物系统自我构建，确定它们的边界、结构和相关问题空间，不像预先定义的机器人。它们是能量受限的，必须选择一个“透镜”来粗粒化世界。
• 涡虫细胞通过查看邻近细胞正在做什么来决定在空间上跟随梯度和其他生物信息线索进行正确发育：决定细胞中的“局部空间差异”线索，而不是像自上而下训练和设计的AI自主体那样接受明确的外部指令，受到控制和影响。

对AGI和集体智能的影响

• 充分条件包括连接的细胞和通过足够数量的单元可靠地发送奖励信号。
• 通用智能出现的条件包括：1. 作为小自主体的单元，期望最小化未来的目标偏差。2. 单元连接并交换奖励或代理奖励。
• 这些生物学见解是否可以扩大规模的问题是当前的任务。例如：推特和全球范围内的社交媒体互动，以及测试激励结构（例如，埃隆的推特实验，或社会/政府规模的社会）可以引导它们的群体自主性，其中细胞/单元可以通过自组织组合成涌现的控制行为。
• 推特被探索为潜在的全球大脑，突出了涌现的自主性以及为集体智能设计激励结构的挑战。这涉及到大脑、社会和社交媒体中的治理。