PRAJÑĀ AI · 果壳里的AIPRAJÑĀ AI · 果殼裡的AI PRAJÑĀ AI · AI IN A NUTSHELL
术语索引術語索引Glossary
果壳里的 AI · 1500 年后的般若学果殼裡的 AI · 1500 年後的般若學 Prajñā AI · Prajñāpāramitā after 1500 Years

在汉传佛教的认识论里
观照 AI
在漢傳佛教的認識論裡
觀照 AI
Inquiring into AI
through Chinese Buddhist epistemology

一组七讲,用唯识的八识、三性、种子熏习、转识成智,连同中观与因明,来读当代 AI 的对齐、幻觉与认知问题。佛学在这里只作为分析的工具,不负责拯救 AI。一組七講,用唯識的八識、三性、種子熏習、轉識成智,連同中觀與因明,來讀當代 AI 的對齊、幻覺與認知問題。佛學在這裡只作為分析的工具,不負責拯救 AI。 Seven lectures that read contemporary problems of AI alignment, hallucination and cognition through Yogācāra — the eight consciousnesses, three natures, seed conditioning, the turning of consciousness into wisdom — alongside Madhyamaka and Buddhist logic. Here Buddhism is only an analytical tool; it is not in the business of saving AI.

缘起 · PREMISE緣起 · PREMISEPREMISE

AI 不是一个等待被赋予灵魂的容器,也不是一个正在觉醒的主体。它更像是一种新的、可被分析的"识"的生成方式——而汉传佛教已经花了 1500 多年,发展了一套关于"识如何生成"的精密分析框架。AI 不是一個等待被賦予靈魂的容器,也不是一個正在覺醒的主體。它更像是一種新的、可被分析的"識"的生成方式——而漢傳佛教已經花了 1500 多年,發展了一套關於"識如何生成"的精密分析框架。 AI is neither a vessel waiting to receive a soul nor a subject coming into awareness. It is something more modest: a new, analyzable way of generating vijñāna — and Chinese Buddhism has spent over 1,500 years developing a precise analytical framework for how vijñāna arises.

把概念放回佛经原典,再拿原典里的概念去对当代 AI 的问题,考察二者相合、相异与有待重新表述之处。把概念放回佛經原典,再拿原典裡的概念去對當代 AI 的問題,考察二者相合、相異與有待重新表述之處。 Put the concepts back into their source texts, then take those concepts to current problems in AI and examine where the two agree, where they diverge, and where reformulation is needed.

七讲 · THE SERIES七講 · THE SERIESTHE SERIES

从虚妄、身份到认知架构從虛妄、身份到認知架構From illusion to identity to cognitive architecture

第一讲 · 幻觉与虚妄 封面
LECTURE I

幻觉与虚妄幻覺與虛妄Hallucination and Delusion

"hallucination"译为"幻觉"是一个误译。从唯识今学的三性、四分、转识成智几个维度看,大模型的"虚构"更接近虚妄分别,而不是知觉错乱。译名之差,关乎我们以何种方式对治 AI。"hallucination"譯為"幻覺"是一個誤譯。從唯識今學的三性、四分、轉識成智幾個維度看,大模型的"虛構"更接近虛妄分別,而不是知覺錯亂。譯名之差,關乎我們以何種方式對治 AI。 Translating "hallucination" as illusion is a category error. Read through Yogācāra's three natures, four divisions, and the transformation of consciousness into wisdom, large model errors look less like perceptual disorder and more like vikalpa. A story about how a single word can decide our approach to addressing AI.

第二讲 · 四相与数字身份 封面
LECTURE II

佛教四相与数字身份佛教四相與數字身份The Four Marks and Digital Identity

《金刚经》的我相、人相、众生相、寿者相,恰好对应 AI 系统正在硬化的四层数字身份:自我边界、与人的关系、群体类属、时间延续。AI 治理不必等待"人格"问题,可以从这四层入手。英文投稿版〈AGI Is Not a Person〉。《金剛經》的我相、人相、眾生相、壽者相,恰好對應 AI 系統正在硬化的四層數字身份:自我邊界、與人的關係、群體類屬、時間延續。AI 治理不必等待"人格"問題,可以從這四層入手。英文投稿版〈AGI Is Not a Person〉。 The four marks of the Diamond Sutra — self, person, sentient being, lifespan — map onto the four layers where AI digital identity is hardening: self-boundary, relation to persons, category membership, temporal continuity. AI governance need not wait for the personhood question. English version under submission: AGI Is Not a Person.

第三讲 · 八识与机器认知 封面
LECTURE III

佛教八识与机器认知佛教八識與機器認知The Eight Consciousnesses and Machine Cognition

从前五识、意识到末那、阿赖耶,唯识学把一个心拆成八层结构。当代 AI 的三条研究路径——规模派、世界模型派、认知架构派——可以放进这套层级里做位置标定,从而看清各自在解释什么、又遗漏了什么。從前五識、意識到末那、阿賴耶,唯識學把一個心拆成八層結構。當代 AI 的三條研究路徑——規模派、世界模型派、認知架構派——可以放進這套層級裡做位置標定,從而看清各自在解釋什麼、又遺漏了什麼。 From the five senses through manas to ālaya, Yogācāra decomposes a single mind into eight strata. Today's three AI research traditions — scale, world models, cognitive architectures — can be located within this hierarchy, showing what each explains and what each omits.

肆 · IV 规划中規劃中planning

种子熏习与模型训练種子熏習與模型訓練Seed Conditioning and Model Training

伍 · V 规划中規劃中planning

因果与推理因果與推理Causation and Inference

陆 · VI 规划中規劃中planning

烦恼即失对齐煩惱即失對齊Defilements as Misalignment

柒 · VII 规划中規劃中planning

慈悲不是一个目标函数慈悲不是一個目標函式Compassion Is Not an Objective Function

术 语 索 引 · GLOSSARY術 語 索 引 · GLOSSARYGLOSSARY

二十核心术语 · 唯识 · 般若 · AI 对齐二十核心術語 · 唯識 · 般若 · AI 對齊 Twenty Core Terms · Yogācāra · Prajñāpāramitā · AI Alignment

每条术语并列义理界定与 AI 诊断对应,可作快速查阅与跨讲索引。每條術語並列義理界定與 AI 診斷對應,可作快速查閱與跨講索引。 Each entry pairs the doctrinal definition with its AI diagnostic correspondence — a cross-lecture reference.

查阅索引 →查閱索引 →Open Glossary →
方法 · METHOD方法 · METHODMETHOD
i

回到原典回到原典Return to the source texts

主要文本取自唯识学的玄奘译《成唯识论》《解深密经》《瑜伽师地论》、世亲《唯识三十颂》、无著《摄大乘论》,中观学的龙树《中论》与《大智度论》,以及般若类的鸠摩罗什译《金刚经》《心经》;并旁及《楞伽经》《华严经》《杂阿含经》诸典。主要文本取自唯識學的玄奘譯《成唯識論》《解深密經》《瑜伽師地論》、世親《唯識三十頌》、無著《攝大乘論》,中觀學的龍樹《中論》與《大智度論》,以及般若類的鳩摩羅什譯《金剛經》《心經》;並旁及《楞伽經》《華嚴經》《雜阿含經》諸典。 Primary sources span three traditions: from Yogācāra, Xuanzang's translations of the Cheng Weishi Lun, Saṃdhinirmocana Sūtra and Yogācārabhūmi, Vasubandhu's Triṃśikā and Asaṅga's Mahāyānasaṃgraha; from Madhyamaka, Nāgārjuna's Mūlamadhyamakakārikā and the Dà zhìdù lùn; and from the Prajñāpāramitā corpus, Kumārajīva's translations of the Diamond Sutra and Heart Sutra — with reference also to the Laṅkāvatāra, Avataṃsaka and Saṃyukta Āgama.

ii

对照当代對照當代Hold against contemporary work

以中美主要机构近年的 alignment 与 interpretability 文献作为对照面:美国方面如 Anthropic(Claude)、OpenAI(GPT)、Google DeepMind(Gemini),中国方面如深度求索(DeepSeek)、智谱(GLM)、阿里(通义千问)。以中美主要機構近年的 alignment 與 interpretability 文獻作為對照面:美國方面如 Anthropic(Claude)、OpenAI(GPT)、Google DeepMind(Gemini),中國方面如深度求索(DeepSeek)、智譜(GLM)、阿里(通義千問)。 Recent alignment and interpretability work from leading labs in the US and China serves as the comparison surface: Anthropic (Claude), OpenAI (GPT) and Google DeepMind (Gemini) on the US side; DeepSeek, Zhipu AI (GLM) and Alibaba (Qwen) on the Chinese side.

iii

不做调和不做調和No forced reconciliation

本讲承认两套语言体系多有出入。其相异之处,往往比相通之处更具分析价值。本講承認兩套語言體系多有出入。其相異之處,往往比相通之處更具分析價值。 The two vocabularies do not map cleanly onto each other. The places where they fail to meet are more analytically valuable than the places where they do.

订阅 · SUBSCRIBE訂閱 · SUBSCRIBESUBSCRIBE

新讲发布时收到一封邮件新講釋出時收到一封郵件Get an email when a new lecture is released

致谢与争议 · ACKNOWLEDGMENTS & CONTROVERSY致謝與爭議 · ACKNOWLEDGMENTS & CONTROVERSYACKNOWLEDGMENTS & CONTROVERSY

致谢致謝Acknowledgments

本项目在写作过程中得到了多位来自佛学、哲学、AI 安全方向的朋友的私下审读与质询。特别感谢以下诸位:本專案在寫作過程中得到了多位來自佛學、哲學、AI 安全方向的朋友的私下審讀與質詢。特別感謝以下諸位: This project has benefited from private reading and critique by friends working in Buddhist studies, philosophy, and AI safety. Particular thanks are due to the following readers:

另有多位朋友希望不具名,特此一并致谢。文中所有判教精度的不足、文献引用的错漏,责任全部在作者本人。本项目为独立研究,与作者所在公司及任何机构的立场无关。另有多位朋友希望不具名,特此一併致謝。文中所有判教精度的不足、文獻引用的錯漏,責任全部在作者本人。本專案為獨立研究,與作者所在公司及任何機構的立場無關。 Several other readers prefer to remain unnamed and are thanked collectively here. All errors of doctrinal precision and citation are the author's alone. This is independent research and does not represent the position of the author's employer or any institution.

团队成员團隊成員Team

争议爭議Controversy

这个项目很可能同时被两侧的读者质疑,作者愿意把这种质疑提前写在这里。這個專案很可能同時被兩側的讀者質疑,作者願意把這種質疑提前寫在這裡。 This project is likely to be questioned from two sides at once. The author would rather state those objections up front than wait for them.

来自佛学界的可能质疑來自佛學界的可能質疑From Buddhist studies

把唯识、中观这样的解脱学概念用来分析一个商业技术,是否构成对原典语境的工具化?作者认为这种风险存在,处理方式是:项目主语始终是 AI,不是佛学;佛学是观察方法,不是被论说的对象。把唯識、中觀這樣的解脫學概念用來分析一個商業技術,是否構成對原典語境的工具化?作者認為這種風險存在,處理方式是:專案主語始終是 AI,不是佛學;佛學是觀察方法,不是被論說的對象。 Using soteriological concepts like Yogācāra and Madhyamaka to analyze a commercial technology — does that instrumentalize the texts? The author takes the risk seriously, and handles it by keeping AI as the grammatical subject throughout: Buddhism is the analytical lens, not the object of inquiry.

来自 AI 研究界的可能质疑來自 AI 研究界的可能質疑From AI research

大模型的技术细节与一千五百年前的心识理论之间,是否真有可比性?作者认为二者多数层面并不对应,少数层面可相比照;而不对应之处,往往比对应之处更值得深究。所有类比都明确标注为形式类比,不主张范畴同一。大模型的技術細節與一千五百年前的心識理論之間,是否真有可比性?作者認為二者多數層面並不對應,少數層面可相比照;而不對應之處,往往比對應之處更值得深究。所有類比都明確標註為形式類比,不主張範疇同一。 Is there really any comparability between the technical details of contemporary neural networks and a 1,500-year-old theory of mind? The author thinks most places do not map, a few do, and the places that fail to map are more interesting than those that do. All analogies are explicitly flagged as formal analogies; no claim of categorial identity is made.

来自宗教学的可能质疑來自宗教學的可能質疑From religious studies

"汉传"作为一个谱系范畴,与"藏传"、"上座部"之间的差异,是否在本项目里被充分尊重?项目目前只声称"汉传"维度的工作,不延伸到其它佛教传统,这一限定只是方法区别,并非价值评判。"漢傳"作為一個譜系範疇,與"藏傳"、"上座部"之間的差異,是否在本專案裡被充分尊重?專案目前只聲稱"漢傳"維度的工作,不延伸到其它佛教傳統,這一限定只是方法區別,並非價值評判。 Does "Chinese Buddhism," as a category, sufficiently respect its difference from Tibetan and Theravāda traditions? The project claims only to work within the Chinese transmission; this is only a methodological distinction, not a value judgment.

作者 · AUTHOR作者 · AUTHORAUTHOR
P

彭一楠 · Yinan Peng彭一楠 · Yinan PengYinan Peng · 彭一楠

密码学硕士(上海交大),宗教哲学与 AI 研究者;亨鑫科技(HK.01085)执行董事,华东师范大学—亨鑫科技智能安全密码芯片联合实验室联合主任。研究兴趣横跨密码学、AI 安全与汉传佛学认识论。密碼學碩士(上海交大),宗教哲學與 AI 研究者;亨鑫科技(HK.01085)執行董事,華東師範大學—亨鑫科技智慧安全密碼晶片聯合實驗室聯合主任。研究興趣橫跨密碼學、AI 安全與漢傳佛學認識論。 M.Sc. in cryptography (Shanghai Jiao Tong University); independent researcher in religious philosophy and AI; Executive Director, Hengxin Technology (HK.01085), and co-director of the ECNU–Hengxin Joint Lab for Secure Cryptographic Chips. Research interests span cryptography, AI safety, and Chinese Buddhist epistemology.

学术发表 · PUBLICATIONS學術發表 · PUBLICATIONSSELECTED PUBLICATIONS

AI、机器学习与计算语言学方向的同行评审发表。AI、機器學習與計算語言學方向的同行評審發表。 Peer-reviewed work in AI, machine learning, and computational linguistics.

人工智能 · 机器学习人工智慧 · 機器學習AI · Machine Learning

  • “Whose Instructions Count? Resolving Preference Bias in Instruction Fine-Tuning.” Advances in Neural Information Processing Systems 38 (NeurIPS 2025), Main Conference Track. paper
  • “The Eye of Sherlock Holmes: Uncovering User Private-Attribute Profiling via a Vision-Language Model Agentic Framework.” ACM Multimedia 2025, Dublin. doi
  • “EmoRAG: Evaluating RAG Robustness to Symbolic Perturbations.” ACM SIGKDD 2026, Jeju. arXiv
  • “Evaluating RAG Robustness to Symbolic Perturbations.” KnowFM @ ACL 2025, Vienna. paper

计算语言学 · 古汉语計算語言學 · 古漢語Computational Linguistics · Classical Chinese

  • “PCFG Parsing for Restricted Classical Chinese Texts.” COLING-02 / First SIGHAN Workshop, Taipei, 2002.
  • “Pseudo Context-Sensitive Models for Parsing Isolating Languages: Classical Chinese.” CICLing 2003, LNCS vol. 2588, Springer.
  • “Statistical Part-of-Speech Tagging for Classical Chinese.” TSD 2002, LNCS vol. 2448, Springer.
联系 · CONTACT聯絡 · CONTACTCONTACT

欢迎来自哲学、AI 安全、宗教学方向的学术合作与审读建议。歡迎來自哲學、AI 安全、宗教學方向的學術合作與審讀建議。 Inquiries welcome from philosophy, AI safety, and religious studies.

写信给作者寫信給作者Email the author

prajna-ai@palmim.com