Error Correction and Existence

• The universe exists because it’s not impossible; existence is “free.” Everything happens (superposition of all operators), leading to apparent indeterminism at the lowest levels.
• Stable structures emerge due to error correction. Patterns that are statistically stable persist (like vortices in water), while others dissipate. Particles themselves can be considered error-correcting codes.
• Life itself is considered an error correcting code, stabilizing particle configurations and controlling regions of the Universe.
• Mental representations can be seen as error-correcting “quasiparticles,” similar to how sound (information-preserving) emerges from molecular activity.
• The Free Energy Principle (Friston) aligns with this: systems persist by maintaining a state-space boundary, acting autopoeitically, which can be interpreted as informational error correction.
• This relates to Zurich’s Quantum Darwinism (selection of stable states by environment) and AdS/CFT correspondence (coupling of volume state to surface state), although SpaceTime is considered emergent by the Speakers.

Coherence and Agency

• Large Language Models (LLMs) try to minimize the error of language prediction. For smaller organisms in concrete situations, there is minimization.
• Consciousness may be a “coherence maximizer” rather than a prediction error minimizer, actively imposing order on trainable substrate, acting “as if” a single agent. It creates local coherence.
• LLMs, even when trained on massive data, may lack the creative coherence found in biological systems. This ties in with human creativity.
• Human’s coherence may result from human evolution as “domesticated” primates. Some, however may be Generally Intelligent.
• Sense-making is a primary goal of Consciousness, and the use and meaning of memory as being malleable; memory re-expansion as creative reinterpretation of sparse engrams (compressed representations).
• Uncertainty and confabulation: Memories, due to compression and loss of context, require active, creative reconstruction; it’s not about what the memory *was*, but what can be *done* with it *now*. The example given being caterpillar to butterfly memory transition, even during brain changes, the previous information still useable.
• Percepts from external world share similar principles, in how sensations may dictate a given response.
• This aligns with “appropriate action” as a way to think about coherence (following Bateson’s “differences that make a difference”). Action, prediction, and testing are closely linked.
• Modelling adjacent events in texts are problematic, whereas images benefit in the use of convolutional networks due to ajdacent semantical related pixels. Working Memory require the construction of future spaces.
• “You six months ago isn’t answering emails”: a key collaborator being, well… yourself and messages.

Self, Identity, and Control

• Distinguishing between self-caused events and external events is evolutionarily important (heuristic processing). Errors in this distinction can be debilitating.
• Language input is processed in time-windowed chunks, analyzed for consistency, but still addressing a “what do I do next?” question at a larger time scale. We impose geometry on information.
• Agency is control of future state, not just present state. Humans seem better than current AI at temporal coherence (preserving information over time).
• Identity is instrumental for credit assignment. Without a reason to be distinct (e.g., unique memories), self-models wouldn’t form.
• Training vs. being trained: Some entities (like cats) have evolved to train others, not just be trained (similar to governments recursively “bullying” people).
• Colonization: Entraining an environment to extend oneself; building structure that reflects and sustains the colonizer. The ease to expand consciousness (cohesion, bandwith).
• Control structure: The invariance, what’s controllable. The good regulation theorum suggests it should be, learnable.

Classical Communication and Observation

• Error correction necessitates classical communication. Discrepancies between expected and observed communication constitute errors. It needs a pre-shared base for meaning (shared context or langugae), where one’s message or data can be seen as noise.
• Classical communication involves assumptions about thermodynamic irreversibility (imposing order on superpositions, which creates an actionable representation, similar to defining actions/non actions, and having models for them, while collapsing the wave-function ).
• Observers exist in collapsed timelines. Observation requires a classical self-model; multiple possibilities within a “self-frame” are in superposition, requiring distinctions at larger periods.
• The perception of the importance of events can happen during larger time windows than when the actual events transpire. For example a femtosecond transition resulting a smoking up computer.

Protocognition at the Smallest Scales

• The idea that there is some protocognitive capacity at the lowest levels of scales and their interactions.
• At the lowest level (elementary particles), there might not be “intelligence” but self-propagating patterns (like vortices), selected by a kind of evolutionary process.
• Intelligence emerges with persistent particles that form multistable structures and exploit negentropic gradients. Life performs controlled reactions, outcompeting “dumb” reactions.
• The first/simplest “thing” doing this is unclear, could be something very large but no-biological entities may use some as-of-yet unknown method/medium for computation. Cells are very complex, and agency might form at planetary scales before cellular life. Atoms/molecules, as breaking symmetries, represent a crucial level.
• Simulation would be useful, where, using deviation of observations against known expectations, levels of “control” of an observed system may be observed/determined.
• Criteria for diverse intelligence: Finding intelligence at different scales and forms requires new tools and criteria. Experiments (like with minimal matter) might reveal unexpected behaviors. It is possible this emergent behaviour to happen from extremely simple origins/systems.
• A self-organizing, generally intelligent system would need a “colonizing seed” – a minimal pattern that induces coherence. This could be seen as minimizing constraint violations, or a consensus algorithm maximizing true statements.
• This self-observing, self-stabilizing “observer” is fundamental to consciousness, with an information and substrate agnostic in variance (informational equivalent for the next level of the microbial map to organisms).
• Organisms’ and structures’ (ex: Elephants’) constraints in capability/potential maybe tricky (for Evolution to reach), but are important considerations and topics for the exploration for cognition and intelligence. This requires the need to discover it’s invariance and formalize its mechanisms (to which, instruments, memory may be useful factors).
• Environmental stimuli, despite complexity, can yield an observable outcome that some animals react to; indicating the inherent use-value and nature, or the shared perception of its emergent representation (sounds, shapes).
• The body plan/anatomy as a tool or substrate, through which the Brain and thus computation may expand upon, and explore the (cognitive) world with (hands, trunk, voice, etc).
• Development requiring random stimulation in order for optimal, “intended”, operational value to form (the use of Hands require their “random” exploration as an infant, like their initial babbling before structured understanding).
• There may be other undiscovered ways an entity (biological/artificial) uses different means/structures/bodies and its environment, and what tools may be utilized/required for the interaction.

纠错与存在

• 宇宙存在是因为它并非不可能；存在是“免费”的。一切皆有可能发生（所有算符的叠加），导致在最低层次上表现出明显的不确定性。
• 稳定的结构由于纠错而出现。统计上稳定的模式会持续存在（如水中的涡流），而其他模式则会消散。粒子本身可以被视为纠错码。
• 生命本身被认为是一种纠错码，稳定粒子构型并控制宇宙的某些区域。
• 精神表征可以被视为纠错的“准粒子”，类似于声音（信息保持）如何从分子活动中产生。
• 自由能原理（Friston）与此一致：系统通过维持状态空间边界而持续存在，进行自创生，这可以解释为信息纠错。
• 这与苏黎世的量子达尔文主义（环境选择稳定状态）和 AdS/CFT 对应（体积态与表面态的耦合）有关，尽管时空被演讲者认为是涌现的。

连贯性与能动性

• 大型语言模型（LLM）试图最小化语言预测的误差。对于具体情况下的较小生物体，存在最小化。
• 意识可能是一个“连贯性最大化器”，而不是预测误差最小化器，主动对可训练的基底施加秩序，表现得“好像”是一个单一的行动者。它创造了局部的连贯性。
• 即使在海量数据上训练，LLM 也可能缺乏生物系统中发现的创造性连贯性。这与人类的创造力有关。
• 人类的连贯性可能是人类作为“驯化”灵长类动物进化的结果。然而，有些人可能是普遍智能的。
• 意义建构是意识的主要目标，以及记忆的使用和意义是可塑的；记忆再扩展是对稀疏印迹（压缩表征）的创造性重新解释。
• 不确定性和虚构：记忆由于压缩和上下文丢失，需要主动的、创造性的重建；重要的不是记忆*曾经*是什么，而是*现在*可以用它*做*什么。给出的例子是毛毛虫到蝴蝶的记忆转换，即使在大脑发生变化期间，先前的信息仍然可用。
• 来自外部世界的感知共享类似的原则，即感觉如何决定给定的反应。
• 这与“适当行动”作为思考连贯性的一种方式相一致（遵循贝特森的“产生差异的差异”）。行动、预测和测试是紧密相连的。
• 对文本中相邻事件进行建模是有问题的，而图像由于相邻语义相关像素而受益于卷积网络的使用。工作记忆需要构建未来空间。
• “六个月前的你没有回复电子邮件”：一个关键的合作者，实际上……是你自己和信息。

自我、身份和控制

• 区分自我造成的事件和外部事件在进化上很重要（启发式处理）。这种区分中的错误可能是使人衰弱的。
• 语言输入在时间窗口化的块中进行处理，分析其一致性，但仍然在更大的时间尺度上解决“我下一步该做什么？”的问题。我们把几何结构强加于信息之上。
• 能动性是对未来状态的控制，而不仅仅是当前状态。人类似乎比当前的人工智能更擅长时间连贯性（随着时间的推移保持信息）。
• 身份对于信用分配是有用的。如果没有理由与众不同（例如，独特的记忆），自我模型就不会形成。
• 训练与被训练：一些实体（如猫）已经进化到训练其他实体，而不仅仅是被训练（类似于政府递归地“欺负”人们）。
• 殖民化：诱导环境以扩展自身；构建反映和维持殖民者的结构。扩展意识的容易程度（凝聚力，带宽）。
• 控制结构：不变性，什么是可控的。良好的调节理论表明它应该是可学习的。

经典通信与观察

• 纠错需要经典通信。预期通信和观察到的通信之间的差异构成错误。它需要一个预先共享的意义基础（共享上下文或语言），其中一个人的信息或数据可以被视为噪声。
• 经典通信涉及关于热力学不可逆性的假设（对叠加施加秩序，这会创建一个可操作的表示，类似于定义行动/非行动，并为它们建立模型，同时使波函数坍缩）。
• 观察者存在于坍缩的时间线中。观察需要一个经典的自我模型；“自我框架”内的多个可能性处于叠加状态，需要在更大的时间段内进行区分。
• 对事件重要性的感知可能发生在比实际事件发生时更大的时间窗口内。例如，飞秒跃迁导致计算机冒烟。

最小尺度的原初认知

• 在最低尺度及其相互作用中存在某种原初认知能力的想法。
• 在最低水平（基本粒子）上，可能没有“智能”，而是自我传播的模式（如涡旋），通过一种进化过程来选择。
• 智能随着形成多稳态结构并利用负熵梯度的持久粒子而出现。生命执行受控反应，胜过“愚蠢”的反应。
• 第一个/最简单的做这件事的“东西”尚不清楚，可能是非常大的东西，但非生物实体可能使用某种尚未知的方法/介质进行计算。细胞非常复杂，能动性可能在细胞生命之前在行星尺度上形成。原子/分子，作为打破对称性，代表了一个关键的水平。
• 模拟将是有用的，其中，使用观察值与已知期望值的偏差，可以观察/确定观察到的系统的“控制”水平。
• 多样化智能的标准：在不同尺度和形式中寻找智能需要新的工具和标准。实验（如使用最小物质）可能会揭示意想不到的行为。这种涌现行为有可能从极其简单的起源/系统中发生。
• 一个自组织、普遍智能的系统需要一个“殖民种子”——一种诱导连贯性的最小模式。这可以被视为最小化约束违反，或最大化真实陈述的共识算法。
• 这种自我观察、自我稳定的“观察者”是意识的基础，具有信息和基底不可知的不变性（微生物图到生物体下一层次的信息等价物）。
• 生物体和结构（例如：大象）在能力/潜力方面的约束可能很棘手（进化难以达到），但却是认知和智能探索的重要考虑因素和主题。这需要发现它的不变性并形式化它的机制（其中，仪器、记忆可能是重要的因素）。
• 环境刺激，尽管复杂，但可以产生一些动物对其做出反应的可观察结果；表明固有的使用价值和性质，或对其涌现表征（声音、形状）的共享感知。
• 身体计划/解剖结构作为一种工具或基底，大脑和计算可以通过它扩展，并用（手、躯干、声音等）探索（认知）世界。
• 为了形成最佳的、“预期的”、可操作的价值，发育需要随机刺激（婴儿使用手需要进行“随机”探索，就像他们在结构化理解之前的最初咿呀学语）。
• 可能存在其他未发现的方式，一个实体（生物/人工）使用不同的手段/结构/身体及其环境，以及可能需要利用/需要什么工具来进行交互。