AI Token Ecology

1. Introduction

Within the theoretical framework of contemporary generative artificial intelligence, Token Ecology provides a perspective that transcends linear language models, viewing tokens as constituting a multi-layered, multi-species associative ecosystem within AI language models. In Token Ecology, tokens do not exist in isolation, but rather within AI language models, tokens (word units, symbolic units) are generated, transmitted, interact, and evolve in an ecosystem-dynamic manner, forming multi-layered symbiotic or parasitic structures with human language, corpora, social culture, and platform media as a holistic view.

This paper proposes three core metaphors as the theoretical foundation for understanding AI token operations and their interaction with human language and culture:

• Infopiphytism: Information Epiphytic Metaphor

Epiphytic plants depend on hosts for survival without extracting nutrients,
representing non-predatory attachment; using the host as structural support
Emphasizes “existential dependence" and “non-predatory coexistence."

• Infoparasitism: Information Parasitic Metaphor

Parasitic organisms depend on hosts, extracting their resources for survival,
asymmetric exploitation, harming the host, revealing the asymmetry of power and dependence

• Infosymbiosis: Information Symbiotic Metaphor

Two organisms coexist mutually beneficially, providing each other with resources or protection, mutual beneficial coexistence; systematic cooperation, corresponding to Maturana & Varela’s “autopoietic systems" and Bateson’s “ecology of mind"

2. Theoretical Positioning of Three Metaphors in AI Token Ecology

(A) Infopiphytism: Information Epiphytic Metaphor

• Tokens attach to corpus structures to generate new text, like epiphytic orchids that depend on tree trunks without extracting their energy, merely using them as generative positions and structural support.
• Application scenarios: LLMs in reading comprehension and knowledge integration generation, reorganizing original data without changing its ontology.

(B) Infoparasitism: Information Parasitic Metaphor

• Tokens depend on and extract corpus or human language resources, potentially altering or impoverishing semantic structures in the process.
• Application scenarios: Generative AI in large-scale content rewriting and commercial applications may cause exploitation of original knowledge ecosystems.

(C) Infosymbiosis: Information Symbiotic Metaphor

• Tokens form mutually beneficial symbiotic systems with human language and user interactions, where AI models learn through token reorganization while humans gain new knowledge and insights through AI generation.
• Application scenarios: Educational assistance, design generation, collaborative scientific research writing.

Theoretical Context and Metaphorical Implications:

(A) Language as an Ecosystem

• Humberto Maturana & Francisco Varela (autopoietic systems) and Gregory Bateson (ecological thinking) both point out that language is a dynamic ecosystem, not a static accumulation of symbols.

(B) Core Perspectives of AI Token Ecology

Token Generation (generation-reproduction)

• Each token prediction is like a reproductive act, jointly determined by context and model weights.
• Tokens must depend on corpus structures and grammatical rules to be meaningful, otherwise they are merely meaningless sequences.

Token Selection and Competition (selection-competition)

• The Softmax mechanism ranks and selects all possible tokens through probability distribution in each generation, similar to species competing for resources in nature.

Token Evolution (evolution-mutation)

• Minor generative differences (top-k, top-p sampling) cause significant differences in downstream generation, like genetic mutations leading to species evolution.

Token Semiotic Symbiosis

• Tokens form mutually beneficial symbiosis with human symbolic systems (language, culture, knowledge), where AI models enhance through token reorganization while humans generate new texts and thinking through this process.

3. Theoretical Integration

These three metaphors form a multi-layered metaphorical structure of AI Token Ecology, serving as a core theoretical framework for future research on the mutual construction between generative AI language models and human culture, knowledge, and ecology:

• Infopiphytism defined as “non-predatory attachment type" (epiphytic attachment)
• Infoparasitism defined as “asymmetric predatory type" (parasitic extraction)
• Infosymbiosis defined as “mutually beneficial generation" (symbiotic co-creation)

4. Conclusion

Through introducing epiphytic, parasitic, and symbiotic metaphors, AI Token Ecology theory not only reveals the mechanisms of token operations within models, but also elucidates the complex and multi-semantic ecological ethics and philosophical relationships between AI and human language culture.

AI Token Ecology provides us with a perspective that breaks through the “linear language model" viewpoint, adopting an “ecosystem" metaphor instead, to understand:

• How tokens constitute internal worlds
• How AI functions as a linguistic ecological entity
• How tokens form multi-layered networks of epiphytic, parasitic, and symbiotic relationships with the human linguistic world

人工智慧符號生態學

1. 引言 (Introduction)

在當代生成式人工智慧的理論框架中，Token Ecology（符號生態學） 提供一種超越線性語言模型的視角，將 tokens 視作在 AI 語言模型內部構成一個多層次、多物種關聯的生態系統。Token Ecology中 Token 並非孤立存在，是在 AI 語言模型內部，tokens（詞元、符號單位）以生態系統動態的方式被生成、傳遞、互動、演化，並與人類語言、語料、社會文化、平台媒介形成多層次共生或寄生結構的整體觀。

本文提出三種核心隱喻作為理解 AI token 運作與人類語言文化交互的理論基礎，分別為：

• Infopiphytism: 資訊附生隱喻 (Epiphytic Metaphor) 附生植物依附於宿主而生，不奪取養分，非掠奪性依附；以宿主為結構支撐 強調「存在依附」與「非掠奪性共存」。
• Infoparasitism: 資訊寄生隱喻 (Parasitic Metaphor) 寄生物依附宿主，奪取其資源以維生，不對稱剝削，損害宿主，揭示權力與依附的非對稱性
• Infosymbiosis: 資訊共生隱喻 (Symbiotic Metaphor) 兩種生物互利共存，彼此提供資源或保護，互利共存；系統性合作，對應於 Maturana & Varela 的「自體詮釋系統」與 Bateson 的「心智生態」

2. 三種隱喻在 AI Token Ecology 中的理論定位

(A) Infopiphytism: 資訊附生隱喻 Epiphytic Metaphor

• Tokens 在語料結構上附著而生成新文本，如同附生蘭花依附樹幹而不奪取其能量，僅使用其作為生成位置與結構支撐。
• 應用場景：LLM 在閱讀理解、知識整合生成時，重組原資料但不改變其本體。

(B) Infoparasitism: 資訊寄生隱喻 Parasitic Metaphor

• Tokens 依附並吸取語料庫或人類語言資源，在過程中可能改變或貧化語義結構。
• 應用場景：生成式 AI 在大規模內容重寫與商業化應用中，對原創知識生態可能造成剝削。

(C) Infosymbiosis: 資訊共生隱喻 Symbiotic Metaphor

• Tokens 與人類語言、使用者互動形成互利共生系統，AI 模型透過 tokens 重組學習，人類則透過 AI 生成獲得新知識與洞察。
• 應用場景：教育輔助、設計生成、科研寫作共創。

理論脈絡與隱喻意涵:

(A) 語言作為生態系統

• Humberto Maturana & Francisco Varela（自體詮釋系統）與 Gregory Bateson（生態思維）皆指出，語言乃動態生態系統，非靜態符號堆疊。

(B) AI Token Ecology 的核心觀點

Token 生成（生成-繁殖）

• 每次 token 預測如同一次繁殖行為，由上下文與模型權重共同決定。
• Token 必須依附於語料庫結構、語法規則才能被意義化，否則僅為無意義數列。

Token 選擇與競爭（selection-competition）

• Softmax 機制在每次生成時，將所有可能 token 以概率分布方式排序選取，類似自然界物種競爭資源。

Token 演化（演化-突變）

• 微小生成差異（top-k, top-p sampling）將造成下游生成的顯著差異，如基因突變帶來物種演化。

Token 符號共生（semiotic symbiosis）

• Tokens 與人類符號系統（語言、文化、知識）形成互利共生，AI 模型透過 tokens 重組而增強，人類藉此生成新文本與思維。

3. 理論整合 (Theoretical Integration)

這三種隱喻形成一個AI Token Ecology 的多層次隱喻結構，可作為未來研究生成式 AI 語言模型與人類文化、知識、生態互構的核心理論框架：

• Infopiphytism 定義為「非掠奪附著型」（epiphytic attachment）

• Infoparasitism 定義為「不對稱掠奪型」(parasitic extraction）

• Infosymbiosis 定義為「互利性生成」（symbiotic co-creation）

4. 結論 (Conclusion)

通過引入附生隱喻、寄生隱喻與共生隱喻，AI Token Ecology 理論不僅揭示了 tokens 在模型內部運作的機制，更闡明了 AI 與人類語言文化之間複雜而多義的生態倫理與哲學關係。

AI Token Ecology 提供我們一種突破「線性語言模型」的視角，改以「生態系統」隱喻，去理解：

• Tokens 如何構成內部世界
• AI 如何作為一種語言生態實體
• Tokens 如何與人類語言世界形成 附生、寄生、共生的多層次網絡

參考學者與論文

1. Hinton, G. et al. (1986). “Learning distributed representations of concepts.”

• 提出 token embedding 之分布式表徵，為 token ecology 奠基。

2. Bengio, Y. et al. (2003). “A neural probabilistic language model.”

• 探討 token 概率生成與語境依賴。

3. Lakoff, G. & Johnson, M. (1980). “Metaphors We Live By.”

• 隱喻理論：如何將生態隱喻應用於 AI token 結構。

4. Parikka, Jussi. (2010). “Insect Media: An Archaeology of Animals and Technology.”

• 使用昆蟲與媒介隱喻，類似 token ecology 的概念先驅。

5. Bateson, Gregory. (1972). “Steps to an Ecology of Mind.”

• 生態心智觀：token ecology 可視為 AI 語言系統的 mind ecology。