Caterpillar-Butterfly Metamorphosis and Memory

• Caterpillars and butterflies have drastically different bodies and behaviors, requiring massive brain remodeling during metamorphosis. Remarkably, memories learned by the caterpillar can be retained by the butterfly, demonstrating information survival through significant physical restructuring. This challenges computational architectures where data storage is fragile.
• The crucial point is not just *persistence* of memory, but *remapping* of information onto a new organism with different needs. A butterfly doesn’t need specific caterpillar memories (e.g., finding a particular leaf color) but needs to *reinterpret* the underlying *principles* for its new life (e.g., how it interacts/moves/ navigates to it’s benefit/goal).

Bateson’s Paradox and the Necessity of Change

• Species must change to survive environmental shifts, but changing fundamentally means the original entity “disappears”. This poses a paradox: how to persist while constantly changing? This applies to individual learning, transformative experiences, and even puberty.

Self-Improving Memory Architecture (Bow Tie/Autoencoder)

• Levin proposes a “bow tie” or “autoencoder” architecture. A wide funnel of diverse inputs (experiences, stimuli) is compressed into a “generative kernel” (a simplified representation, removing unnecessary details) which can persist/exist in that compressed “smaller” structure (less information). This kernel is the memory engram. This stored memory is then expanded/decompressed by the output. This process repeats endlessly, like “Now Moments” continuously, expanding out from a center point to make the cognitive cone larger.
• This compression is necessary due to energy/time constraints. Organisms can’t track micro-details (like a Laplacian demon); they must generalize. The central part are your memories of previous learned things, and your constant adaptation, where these generalized learned experiences get re-applied to the situation you are experiencing.
• Memories are reinterpreted in the *present*. They are “messages” from the past self, constantly requiring re-evaluation of *meaning*. The present self isn’t bound by the past self’s interpretation. This interpretation step can introduce adaptations.
• The left side (compression) is algorithmic. The right side (reinterpretation) is creative and underdetermined, involving active “sensemaking.” All the organism gets are some memory engrams and have to reinterpret them, but also the ability/incentive exists, to completely change that interpretation, too.

Evolution and the Unreliable Substrate

• Biology operates on an “unreliable substrate.” Organisms can’t rely on a stable environment *or* their own components (due to mutations, etc.). The DNA you inherited can create something unexpected. The frog genome has everything to become either a tadpole/frog/xenobot.
• Organisms *must* interpret information from their ancestors (the genome), but they are not obligated to interpret it the same way. This gives rise to plasticity and problem-solving capacity (e.g., planaria making heads of the wrong species, xenobots/anthrobots).
• This “creative reinterpretation” of information drives intelligence. Agents become better at manipulating information, leading to “confabulation” as a feature, not a bug (adapting information to the present).

Memories as Agents

• Agency is a term to apply when you conduct some experiment and find an entity displays that capacity (it helps your understanding/modelling), and doesn’t require that the object to move through the space (example it could move in idea-space like genetic regulatory networks (GRNs), cells, tissues, organs, even memories themslves..
• Levin proposes exploring *memories themselves* as having agency. There’s a spectrum, from fleeting thoughts to persistent/recurrent thoughts (which can alter brain structure) to personality fragments (dissociative identity disorder). The thinker/pattern itself can be viewed from the perspective of the memory/pattern on it.
• Memories are viewed as temporary patterns (like hurricanes or metabolic processes), blurring the line between “thinkers” and “thoughts”.
• Patterns may strive to *persist* and *expand* their influence. The memory patterns could help incentivize the organism that the data gets reinterpreted. The information itself gets reinterpreted but has features which could help make it easier for the organism/thinker to encode and propogate these learnings.
• New Perspective: Viewing the *physical body* as the memory medium (“tape”) and bioelectric patterns as the driving agents. This is highly speculative, and work on testing that is now only just beginning.

Confabulation and Storytelling

• Confabulation is defined as generating explanations or narratives that aren’t necessarily true to the original event or memory. These “narratives” may help adaptation to future scenarios/outcomes more.
• Split-brain patients exemplify confabulation: the speaking hemisphere invents justifications for actions of the non-speaking hemisphere. This “story telling” tendency goes beyond just “lying”, as that organism is merely driven to form coherence of experience with their past.
• This is fundamental to intelligence: continuous model-building of self and the world. Going too far, becoming only useful for the very very short-term leads to bad longer-term outcomes.
• This is related to how AI can output hallucination: outputs adaptive for *present* context but untrue to *past* context (prioritizing saliency over veracity).
• Humans have a basic inherent drive to see patterns and come up with explainations of their surroundings and events, so for our brains to come up with good patterns on why we have the thoughts/patterns that persist, can help.
• Organisms that become really good at “course graining”, in being good at reducing a vast multitude of experiences into single category to apply the “rules”/learnings is a good survival adaptation.
• Storytelling (at all cognitive-levels): creating narratives about oneself is essential even down to the single cells and pathways levels: Story telling involves not just a total/summary of microstates but, by applying some of the inherent/built in capcity for creative intepretation can allow it to re-intepret those memory/learnings such that the new learning, better reflect new experiences/information, in effect to use old tools to achieve different purposes.

Implications and Future Research

• Polycomputing Framework is a model of using an “evolutionary” computing paradigm for new types of biological computational platforms. The “computation” has a vast variety of potential, like having a computer’s components using biological ones, because organisms are just fundamentally very plastic.
• Research will focus on *mechanisms* of creative reinterpretation. How do engrams get mapped to new situations? Synthetic models (xenobots, anthrobots) are crucial because they lack specific evolutionary history, forcing novel interpretations of their genetic material.
• Scientific papers themselves, act as bow-tie architectures, from the author’s understanding that has to be compressed into text/equations in publication and that text/equations will once again need to be expanded out for use (hopefully it helps!) for the paper’s consumers.
• Future Directions: computational models, applying “polycomputing” to this, biological mechanisms of interpretation. This is being tested using a range of biological models/tools like xeno/anthro-bots.

毛毛虫-蝴蝶变态与记忆

• 毛毛虫和蝴蝶拥有截然不同的身体和行为，在变态过程中需要进行大规模的脑部重塑。 值得注意的是，毛毛虫学到的记忆可以被蝴蝶保留，这表明信息能够在重大的物理结构重组中存活。 这对数据存储脆弱的计算架构提出了挑战。
• 关键点不仅仅是记忆的*持久性*，而是将信息*重新映射*到一个具有不同需求的新生物体上。 蝴蝶不需要特定的毛毛虫记忆（例如，找到特定的叶子颜色），而是需要为了它的新生活*重新解释*潜在的*原则*（例如，它如何交互/移动/导航以实现其利益/目标）。

贝特森悖论与改变的必要性

• 物种必须改变才能在环境变化中生存，但根本性的改变意味着原始实体“消失”。 这提出了一个悖论：如何在不断变化的同时持续存在？ 这适用于个人学习、变革性经历，甚至是青春期。

自我改进的记忆架构（领结/自编码器）

• 莱文提出了一种“领结”或“自编码器”架构。 多样化输入（经验、刺激）的宽漏斗被压缩成一个“生成内核”（简化的表示，去除不必要的细节），该内核可以在该压缩的“较小”结构（较少信息）中持续存在/存在。 这个内核就是记忆印迹。 然后，这个存储的记忆通过输出被扩展/解压缩。 这个过程无限重复，就像“当下时刻”不断地从中心点向外扩展，使认知锥体更大。
• 由于能量/时间限制，这种压缩是必要的。 生物体无法追踪微观细节（像拉普拉斯妖）；他们必须概括。 中心部分是你对以前学到的东西的记忆，以及你的不断适应，这些概括的学习经验会被重新应用到你正在经历的情况中。
• 记忆在*当下*被重新解释。 它们是来自过去自我的“信息”，不断需要重新评估*意义*。 当前的自我不受过去自我的解释的约束。 这个解释步骤可以引入适应。
• 左侧（压缩）是算法性的。 右侧（重新解释）是创造性的和不确定的，涉及主动的“意义建构”。 生物体得到的所有都是一些记忆印迹，必须重新解释它们，但也存在能力/动机，也可以完全改变这种解释。

进化与不可靠的基质

• 生物学在“不可靠的基质”上运作。 生物体不能依赖于稳定的环境*或*它们自己的组成部分（由于突变等）。 你继承的DNA可以创造出意想不到的东西。 青蛙基因组拥有成为蝌蚪/青蛙/异种机器人的一切。
• 生物体*必须*解释来自其祖先（基因组）的信息，但它们没有义务以相同的方式解释它。 这产生了可塑性和解决问题的能力（例如，涡虫制造错误物种的头部，异种机器人/人源机器人）。
• 这种对信息的“创造性重新解释”驱动了智能。 智能体变得更善于操纵信息，导致“虚构”成为一种特征，而不是缺陷（使信息适应当前）。

作为自主体的记忆

• 自主性是一个术语，当你进行一些实验并发现一个实体显示出这种能力（它有助于你的理解/建模）时，就可以应用该术语，并且不要求该对象在空间中移动（例如，它可以在思想空间中移动，如基因调控网络 (GRN)、细胞、组织、器官，甚至是记忆本身。
• 莱文建议将*记忆本身*作为具有自主性的东西来探索。 存在一个光谱，从转瞬即逝的想法到持久/反复出现的想法（可以改变大脑结构），再到人格碎片（分离性身份障碍）。 思考者/模式本身可以从其上的记忆/模式的角度来看待。
• 记忆被视为临时模式（如飓风或代谢过程），模糊了“思考者”和“思想”之间的界限。
• 模式可能努力*持续存在*并*扩大*其影响力。 记忆模式可以帮助激励生物体重新解释数据。 信息本身被重新解释，但具有一些特征，可以帮助生物体/思考者更容易地编码和传播这些学习。
• 新视角：将*物理身体*视为记忆媒介（“磁带”），将生物电模式视为驱动主体。 这是高度推测性的，测试它的工作现在才刚刚开始。

虚构与讲故事

• 虚构被定义为生成不一定真实于原始事件或记忆的解释或叙述。 这些“叙述”可能更有助于适应未来的场景/结果。
• 裂脑患者是虚构的例证：说话的半球为不说话的半球的行为发明理由。 这种“讲故事”的倾向不仅仅是“撒谎”，因为该生物体只是被驱使将其经验与过去形成连贯性。
• 这是智能的基础：持续构建自我和世界的模型。 走得太远，只对非常非常短期有用会导致不良的长期结果。
• 这与人工智能如何输出幻觉有关：输出适应于*当前*上下文，但不真实于*过去*上下文（优先考虑显著性而不是真实性）。
• 人类有一种基本的内在驱动力，可以观察模式并解释周围环境和事件，因此，对于我们的大脑来说，想出关于为什么我们拥有持久存在的思想/模式的良好模式，会有所帮助。
• 变得非常擅长“粗粒化”的生物体，即擅长将大量经验减少为单一类别以应用“规则”/学习，这是一种很好的生存适应。
• 讲故事（在所有认知层面上）：即使在单个细胞和通路水平上，创建关于自身的叙述也是必不可少的：讲故事不仅涉及微观状态的总和/总结，而且通过应用一些固有的/内置的创造性解释能力，可以允许它重新解释那些记忆/学习，以便新的学习更好地反映新的经验/信息，实际上是使用旧工具来实现不同的目的。

影响和未来研究

• 多计算框架是一种模型，它使用“进化”计算范式来构建新型生物计算平台。 “计算”具有广泛的潜力，例如让计算机的组件使用生物组件，因为生物体基本上非常具有可塑性。
• 研究将集中在创造性重新解释的*机制*上。 印迹如何映射到新情况？ 合成模型（异种机器人、人源机器人）至关重要，因为它们缺乏特定的进化历史，迫使对其遗传物质进行新的解释。
• 科学论文本身充当领结架构，从作者的理解必须压缩成文本/方程式发表，这些文本/方程式将再次需要扩展以供使用（希望它有所帮助！）对于论文的消费者。
• 未来方向：计算模型，将“多计算”应用于此，解释的生物学机制。这正在使用一系列生物模型/工具（如异种/人源机器人）进行测试。