Introduction: Turing, Morphogenesis, and Collective Intelligence

• Alan Turing, known for AI, also studied morphogenesis (pattern formation), seeing a deep connection between intelligence and development. All intelligence is collective intelligence, as every cognitive system is made of parts.
• Developmental biology is crucial: We all start as a single cell and gradually become complex cognitive systems. Embryogenesis is the process of transitioning from “physics” to “mind.” There’s no sharp dividing line.

Key Concepts: Multi-Scale Competency, Goal-Directedness

• Multi-scale competency architecture: Biology uses a nested set of problem-solvers at different levels (cells, tissues, organs, organisms), each with its own goals and competencies.
• Goal-directedness is key: A powerful way to recognize, communicate with, and build unconventional agents is by understanding their goals, even if they are very different from our own. Navigation policies in diverse problem-spaces (not just 3D physical space).
• Cognitive boundary model: A framework for understanding the scaling of cognition based on the size and complexity of goals a system can maintain.
• Morphogenesis as collective intelligence: Cells collectively navigate “morphospace” (the space of possible anatomical forms) using bioelectrical networks as a communication mechanism.
• Practical, empirical predictions: These conceptual ideas lead to testable predictions and have led to discoveries in regenerative medicine and bioengineering.
• Synthetic bioengineering: Life’s competency allows creation of novel bodies/minds, raising ethical considerations.

Single Cells and Basal Cognition

• Single cells (e.g., Lacrymaria, Acetabularia) show impressive spatial and behavioral competencies, challenging the notion of cells as simple building blocks.
• Slime molds (Physarum) demonstrate decision-making, e.g., choosing the larger mass, integrating sensory information through biomechanical sensing (like sonar).
• Morphology and behavior are intertwined, highlighting a transition from solving problems in morphological space to behavioral space during evolution.

Expanding the Concept of “Agent”

• The traditional view of distinct “natural kinds” (human, rat, etc.) needs revision. Evolutionary and developmental processes show continuous transitions.
• Caterpillar metamorphosis: The brain is largely rebuilt, yet memories can persist. This raises questions about memory storage and what it’s like to *become* a different kind of agent.
• Planarian regeneration: Cut worms regrow missing parts, including the brain. Memory (e.g., learned patterns) can be stored *outside* the brain and reimprinted on the new brain.
• Tadpole plasticity: An eye induced on the tail can provide visual input to the brain, even though the eye doesn’t connect directly to the brain. This shows incredible plasticity and adaptability, even in adults.

Multi-Scale Competency Architecture: Nested Problem Solvers

• Nested hierarchy of problem-solvers: Not just whole organisms, but *all* components (cells, tissues, etc.) solve problems in different spaces (transcriptional, anatomical, physiological).
• Competency in unfamiliar spaces: We easily recognize intelligence in 3D behavioral space, but we struggle to recognize it in other spaces (e.g., physiological) due to lack of training data.
• IQ test for recognizing intelligence: Assessing another system’s IQ tests *your* ability to recognize the relevant space and goals.
• TAME Framework, Technological Apprach to Mind Everywhere: a way to evaluate other intelligence whether it be evolution-based or synthesized.

Morphogenesis as Collective Intelligence: A Case Study

• Anatomical order, how collection of cell forms anatomical complexity with fidelity, and the genome not explicitly showing this high level morphology.
• Anatomical compiler (long-term goal): Translating a desired anatomical form into stimuli that guide cells to build it, enabling regenerative medicine and bioengineering.
• Software vs. hardware: We’re good at manipulating molecular hardware (genes, pathways), but far from controlling large-scale structure. We need to understand the “software.”
• Homeostatic process around target morphlogy. Biolelectrity is one mechanism which cells use. Changing the bioelectric setpoint rather than micromanaging the system.

Bioelectricity as an Informational Medium

• The memory usecase of Biolelectricity, for example to remember a set point of anatomical morphology to reach when regrowing in planarians, even remembering the *number* of heads!
• Bioelectrical networks: Cells communicate via ion channels and gap junctions, creating electrical networks that store and process information, similar to (but not identical to) neural networks.
• Neural decoding outside the brain: The same concepts and tools used in neuroscience can be applied to non-neural electrical tissues, offering a “neuroscience outside the brain.”
• “Electric face”: Early frog embryos show a bioelectrical pre-pattern of the future face, long before genes for specific features turn on.
• Optogenetics and gap junction control: Tools to manipulate electrical states and cell communication, allowing functional experiments to test the role of bioelectricity.
• Engineering anatomical structures is more than simply changing genes. By targeting bioelectricity (using sub-routine), organs are regrown (tadpole leg regrowth)

Scaling Cognition and Goal-Directedness

• Scaling of goals: Individual cells have simple goals; cell collectives (through bioelectric networks) pursue larger, more complex goals. Goal size defines level.
• Glioblastoma: Cancer cells disconnect electrically, reverting to a unicellular lifestyle with smaller “self.” Maintaining electrical connections can normalize tumor growth.
• Cognitive “glue”: Bioelectrical networks bind cells together, enabling them to navigate morphospace as a collective.
• Measuring Goal Size to measure intelligence of all kinds. Goal persuit persistance when interrupted defines agenthood. Bending spaces by creating simple constraints from high to low levels.

Xenobots: Engineered by Subtraction

• Evolution pivots problems accross dimensions like behavorial vs morphological.
• Xenobots: Skin cells, isolated from a frog embryo, spontaneously self-organize into novel organisms with emergent behaviors (movement, pile-making). No brain, no neurons.
• Kinematic self-replication: Xenobots build copies of themselves from loose cells – a behavior discovered by AI (Josh Bongard’s lab), and *not* found in frogs.
• Engineered by Subtraction” what cell behaviour appears if freed, so their full behavioural capability comes from both constraints and lack thereof, the “default”.
• Evolution creates machines with broader capabilities for generalized problem solving, than simply for just fulfilling a very narrow set of use-cases.
• The ability to self-assemble shows potential to be agents or problem-solvers in other non-standard evolutionary usecases, posing many societal and safety implications for ethics and decision making.

Other Notes

• Survivorship Bias can easily occur due to our ability to engineer novel constructs which are not constrained by evolutionary history, survival, or adaptability.
• The electrical pattern or set point in an organism may adaptively, be shaped over a wide range of acceptable changes to the ‘settings’, so changes to the organism’s electrical signaling or ion channels can create different results (making for great interface), as cells respond according to the new voltage state.
• Agency Claims are claims on how an organism or collection could act, not the specific low-level mechanism.
• A multi-scale agent (one agent inside a bigger one inside etc…) have a spectrum of interfaces to manipulate to persaude behaviour, with gap junctions helping to meld boundaries to coordinate (the higher scales imposing simple contraints such as by guiding and making easy following down energy grandients) or be interfered with for persuasion by setting constraints.
• The evolution to new intelligence or agent starts as simple single celled goals and then to networking for multiple ones. The networks enable a scaling of not only intelligence but more shared memory between units and prediction capability, along with its homeostatic ability that keeps on growing, and error tolerance through adaptability.

导言：图灵、形态发生与集体智能

• 艾伦·图灵，以人工智能闻名，也研究过形态发生（模式形成），认为智能与发育之间存在深刻的联系。所有智能都是集体智能，因为每个认知系统都是由多个部分组成的。
• 发育生物学至关重要：我们都始于一个单细胞，并逐渐成为复杂的认知系统。胚胎发生是从“物理”到“心智”的转变过程。两者之间没有明确的分界线。

关键概念：多尺度能力、目标导向性

• 多尺度能力架构：生物学在不同层次（细胞、组织、器官、生物体）使用一组嵌套的问题解决者，每个层次都有其自身的目标和能力。
• 目标导向性是关键：识别、交流和构建非常规智能体的一种强有力方法是理解它们的目标，即使它们与我们自己的目标非常不同。在不同的问题空间（不仅仅是三维物理空间）中的导航策略。
• 认知边界模型：一个理解认知规模的框架，基于系统可以维持的目标的大小和复杂性。
• 形态发生作为集体智能：细胞利用生物电网络作为沟通机制，集体地在“形态空间”（可能的解剖形式空间）中导航。
• 实用的、经验性的预测：这些概念性想法带来了可测试的预测，并导致了再生医学和生物工程领域的发现。
• 合成生物工程：生命的潜能允许创造新颖的身体/心智，这引发了伦理方面的考量。

单细胞和基础认知

• 单细胞（例如，泪腺虫、伞藻）显示出令人印象深刻的空间和行为能力，挑战了细胞作为简单构建块的概念。
• 黏菌 (Physarum) 展示了决策能力，例如，选择较大的质量，通过生物力学传感（如声纳）整合感官信息。
• 形态和行为相互交织，突出了进化过程中从解决形态空间问题到解决行为空间问题的转变。

扩展“智能体”的概念

• 传统的“自然种类”（人类、老鼠等）的明确区分需要修正。进化和发育过程显示出连续的过渡。
• 毛毛虫变态：大脑大部分被重建，但记忆可以持续存在。这引发了关于记忆存储以及*成为*另一种智能体的感受的问题。
• 涡虫再生：被切割的蠕虫会再生缺失的部分，包括大脑。记忆（例如，学习的模式）可以存储在大脑*之外*，并重新印在新的大脑上。
• 蝌蚪可塑性：在尾巴上诱导出的眼睛可以向大脑提供视觉输入，即使眼睛不直接连接到大脑。这表明即使在成年人中，也存在令人难以置信的可塑性和适应性。

多尺度能力架构：嵌套的问题解决者

• 嵌套的问题解决者层次结构：不仅是整个生物体，而且*所有*组成部分（细胞、组织等）都在不同的空间（转录、解剖、生理）中解决问题。
• 在不熟悉的空间中的能力：我们很容易识别三维行为空间中的智能，但我们很难识别其他空间（例如，生理）中的智能，因为缺乏训练数据。
• 识别智能的智商测试：评估另一个系统的智商会测试*您*识别相关空间和目标的能力。
• TAME 框架（Technological Approach to Mind Everywhere, 处处皆心智的技术方法）：一种评估其他智能的方法，无论它是基于进化的还是合成的。

形态发生作为集体智能：一个案例研究

• 解剖秩序，即细胞集合如何忠实地形成解剖复杂性，以及基因组并未明确显示这种高水平形态。
• 解剖编译器（长期目标）：将所需的解剖形式转换为指导细胞构建它的刺激，从而实现再生医学和生物工程。
• 软件与硬件：我们擅长操纵分子硬件（基因、通路），但远未控制大规模结构。我们需要理解“软件”。
• 围绕目标形态的稳态过程。生物电是细胞使用的一种机制。改变生物电设定点，而不是对系统进行微观管理。

生物电作为信息媒介

• 生物电的记忆用例，例如记住涡虫再生时要达到的解剖形态设定点，甚至记住头的*数量*！
• 生物电网络：细胞通过离子通道和间隙连接进行通信，创建存储和处理信息的电网络，类似于（但不完全相同于）神经网络。
• 大脑外的神经解码：神经科学中使用的相同概念和工具可以应用于非神经电组织，提供“大脑外的神经科学”。
• “电面孔”：早期青蛙胚胎显示出未来面部的生物电预模式，远早于特定特征的基因开启。
• 光遗传学和间隙连接控制：操纵电状态和细胞通信的工具，允许进行功能实验来测试生物电的作用。
• 工程解剖结构不仅仅是改变基因。 通过靶向生物电（使用子程序），器官得以再生（蝌蚪腿再生）。

扩展认知和目标导向性

• 目标的扩展：单个细胞有简单的目标；细胞集体（通过生物电网络）追求更大、更复杂的目标。目标大小定义了级别。
• 胶质母细胞瘤：癌细胞电连接断开，恢复到具有较小“自我”的单细胞生活方式。保持电连接可以使肿瘤生长正常化。
• 认知“胶水”：生物电网络将细胞结合在一起，使它们能够作为一个集体在形态空间中导航。
• 测量目标大小以衡量各种智能。被打断时目标的持续存在定义了自主性。通过从高到低创建简单的约束来弯曲空间。

异种机器人：通过减法进行工程设计

• 进化在行为与形态等维度上枢转问题。
• 异种机器人：从青蛙胚胎中分离出来的皮肤细胞，自发地自组织成具有涌现行为（运动、堆积）的新型生物体。没有大脑，没有神经元。
• 运动学自我复制：异种机器人从松散的细胞中构建自己的副本——这是人工智能（Josh Bongard 的实验室）发现的一种行为，在青蛙中*找不到*。
• “通过减法进行工程设计”：如果解放，细胞会出现什么行为，所以它们的全部行为能力来自约束和缺乏约束，“默认”。
• 进化创造了具有更广泛的通用问题解决能力的机器，而不仅仅是为了满足一组非常狭窄的用例。
• 自组装的能力显示出在其他非标准进化用例中成为智能体或问题解决者的潜力，对伦理和决策提出了许多社会和安全影响。

其他说明

• 生存者偏差很容易发生，因为我们能够设计出不受进化历史、生存或适应性限制的新型结构。
• 生物体中的电模式或设定点可以适应性地、在“设置”的广泛可接受变化范围内形成，因此对生物体的电信号或离子通道的改变会产生不同的结果（使其成为很好的界面），因为细胞会根据新的电压状态做出反应。
• 自主性声明是关于生物体或集合如何行动的声明，而不是特定的低层机制。
• 多尺度智能体（一个智能体在更大的智能体内部，等等…）具有一系列可操作的界面来诱导行为，间隙连接有助于融合边界以协调（更高尺度通过引导和使遵循能量梯度变得容易来施加简单的约束）或通过设置约束来干扰以进行说服。
• 新的智能或智能体的进化始于简单的单细胞目标，然后是多个目标的网络化。网络不仅可以扩展智能，还可以在单元之间共享更多的记忆和预测能力，以及其不断增长的稳态能力，以及通过适应性实现的容错能力。