| 454 | 8 | 130 |
| 下载次数 | 被引频次 | 阅读次数 |
大模型是一种未来历史。未来历史的关键在于其在某种程度上逆转了时间箭头。在伽达默尔的“效果历史”的概念基础上,笔者进一步提出“效果未来”的概念。效果未来的展开需要依赖一种新的方法论,即未来发生学。传统发生学的基本逻辑是用现在的知识还原历史。未来发生学的关键则是用现在的知识构造未来和推断未来。在未来发生学的展开过程中,猜想至关重要。当然,这种猜想不能成为天马行空的无约束想象,而应该是一种有约束想象。人与大模型的互动过程存在三重境界:工具、游戏和生存。这三重境界的关键是人与大模型的关系。人与机器的关系形式同样可以被分为三种:外在、内在与共在。有待的共在体现在人们对大模型的高度依赖,而真正的共在则要走向一种无待的共在,或者说是一种非依赖性的共在。人类与大模型的共在不仅要体现为大模型对人类的对齐,同时也要体现出人类向大模型的对齐。双重对齐还意味着人类需要与大模型形成伽达默尔所强调的视域融合。
Abstract:Large models is a future history. The key to future history lies in its ability to reverse the arrow of time to some extent. On the basis of Gadamer's concept of“effect history”,the author further proposes the concept of“effect future”. The development of“future effects”requires a new methodology,namely,futurology. The basic logic of traditional genetics is to restore history with current knowledge. The key to futurology is to construct and infer the future using current knowledge. In the process of developing future genetics,conjecture is crucial. Of course,this conjecture should not be a wild and unconstrained imagination,but rather a constrained imagination. There are three levels of interaction between humans and large models,namely tools,games,and survival. The key to these triple realms is the relationship between humans and large models. The relationship between humans and machines can also be divided into three forms:external,internal,and coexisting. The pending commonality is reflected in people's high dependence on large models,while the true commonality needs to move towards a non-dependent commonality,or rather a non-dependent commonality. The commonality between humans and large models should not only be reflected in the alignment of large models with humans,but also in the alignment of humans with large models. Double alignment also means that humans need to form the fusion of perspectives emphasized by Gadamer with large models.
1.钱力等:《Chat GPT的技术基础分析》,《数据分析与知识发现》2023年第3期。
2.田文林、林海虹:《伊斯兰与西方的冲突:一个自我实现的文化预言》,《世界经济与政治》2002年第1期。
3.Friedel Weinert,“Tracing the Arrows of Time,”Information and Interaction:Eddington,Wheeler,and the Limits of Knowledge,Gewerbestrasse:Springer,2017,73~101.
4.Joel Lebowitz,“Boltzmann’s Entropy and Time’s Arrow,”Physics Today,1993,46(9):32~38.
5.Kuan Fang,Alexander Toshev&Li Fei-Fei et al.,“Scene Memory Transformer for Embodied Agents in Long-Horizon Tasks,”Proceedings of the IEEE/CVF conference on computer vision and pattern recognition,2019,pp.538~547.
6.Mario Castagnino,Luis Lara&Olimpia Lombardi,“The Direction of Time:From the Global Arrow to the Local Ar‐row,”International Journal of Theoretical Physics,2003,42:2487~2504.
7.Olof Sundin,“Janitors of Knowledge:Constructing Knowledge in the Everyday Life of Wikipedia Editors,”Journal of Documentation,2011,67(5):840~862.
8.Rush Doshi,“The United States,China,and the Contest for the Fourth Industrial Revolution,”US Senate Committee on Commerce,Science,and Transportation,2020.
9.Samuel P.Huntington,“The Clash of Civilizations?”,Foreign Affairs,1993,72(3):22~49.
10.冯契:《冯契文集第一卷:认识世界和认识自己》,华东师范大学出版社1996年版a。
11.冯契:《冯契文集第三卷:人的自由和真善美》,华东师范大学出版社1996年版b。
12.冯友兰:《新原人》,生活·读书·新知三联书店2007年版。
13.世亲菩萨:《俱舍论颂疏表释》,上海古籍出版社2016年版。
14.唐君毅:《生命存在与心灵境界》上卷,九州出版社2016年版。
15.王国维:《人间词话》,上海古籍出版社1998年版。
16.张载:《张子正蒙》,上海古籍出版社2000年版。
17.庄周:《庄子全译》,贵州人民出版社1991年版。
18.宗白华:《美学散步》,上海人民出版社2015年版。
19.Bernard Stiegler,Technics and Time,1:The Fault of Epimetheus,Stanford:Stanford University Press,1998.
20.Friedrich A.Hayek,The Counter Revolution of Science:Studies on the Abuse of Reason,New York:The Free Press of Glencoe,1964.
21.Hans Georg Gadamer,Truth and Method,New York:Continuum,1975.
22.Jean Piaget,Genetic Epistemology,New York:Columbia University Press,1971.
23.Karl Popper,Conjectures and Refutations:The Growth of Scientific Knowledge,London:Routledge and Kegan Paul,1969.
24.Karl Popper,The Logic of Scientific Discovery,London:Routledge,2005.
25.Martin Heidegger,Being and Time,New York and Evanston:Harper&Row,1962.
26.Paul Ricoeur,Hermeneutics and The Human Sciences:Essays on Language,Action and Interpretation,New York:Cambridge University Press,2016.
27.Robert Rosenthal and Lenore Jacobson,Pygmalion in the Classroom:Teacher Expectation and Pupils’Intellectual,New York:Irvington Publishers,1992.
28.Seth Lloyd,Physical Origins of Time Asymmetry,New York:Cambridge University Press,1994.
29.Stephen Hawking,A Brief History Of Time,New York:Bantam Books,1998.
30.Steven Weinberg,The Quantum Theory of Fields,New York:Cambridge University Press,1995.
(1)基于现有的公开资料,GPT-3的预训练语料集主要包括Common Crawl数据集、Reddit链接、书籍、期刊、英文维基百科数据等,而微调语料则主要源于Git Hub上的公共代码库。
(2)在GPT-4的官方页面中,其展示了如何利用提示使得GPT的回答带有苏格拉底式的思辨色彩。参见Open AI网站,网址:https://openai.com/research/gpt-4.
(1)主流宇宙学界认为,宇宙始于大爆炸(Big Bang),这之后,宇宙的熵随着时间而增加,最终宇宙将耗散其全部的热,进入最高熵的“热寂”(Heat Death)或称“大冻结”(Big Chill)状态。经验证据发现的宇宙膨胀(expansion of the universe)现象支持了这一结论。
(2)波函数(wave function)坍缩具有非幺正(nonunitary)性质,亦即这一过程是不可逆的。
(3)斯蒂芬·霍金(Stephen Hawking)认为,心理学时间箭头(cosmological arrow of time)的基础在于我们能够记住过去的事情而非未来的事情,这是我们心理上判断时间流逝的方向的依据。
(4)苏茨卡弗说,人工智能在学习过程中,通过不断的文本输入,实际上学会了一个“世界模型”(world model)。参见网址:https://an‐alyticsindiamag.com/text-is-a-projection-of-the-world-says-openais-sutskever/.
(1)GPT-3的训练数据总容量和各训练数据集所占比例。参见网址:https://www.springboard.com/blog/data-science/machine-learninggpt-3-open-ai/.
(2)Che Pan,“Tech War:Chinese Chat GPT Ambitions Threatened by Lack of Advanced AI Chips Like Nvidia A100 GPU,Experts Say”,参见网址:https://www.scmp.com/tech/tech-war/article/3212128/tech-war-chinese-chatgpt-ambitions-threatened-lack-advanced-ai-chips-nvidiaa100-gpu-experts-say.
(1)实验者访问一所公立小学,他们告诉老师将用“哈佛技能获得变化实验”来准确预测哪些学生会成为未来的天才。实验者随机挑选了这些所谓天才的小学生,但实际上并不存在这样的“哈佛方法”。然而,在学年末,当这些学生在参加智力测验时,那些被认为是天才的孩子确实比其他同学表现出了明显的智力增长。这场实验所在的学校被作者化名为“橡树学校”(Oak School),故实验被称为“橡树学校实验”(Oak School Experiment)。
(2)皮亚杰写到,发生认识论除了解释知识本身,还探讨“平均的思想结构”(equilibrated thought structures)和“思想发展过程中从一个水平到另一个水平的转变”(transformations from one level to another in the development of thought)中的“逻辑形成”(logical formalizations)。
(1)波普尔写到,“衡量一个理论科学地位的标准就是其可证伪性、可反驳性和可检验性。”
(1)在对阵Alpha Go的第四局中,韩国棋手李世石利用其弱点将其击败。参见网址:https://english.hani.co.kr/arti/english_edition/e_in‐ternational/734808.html.
(2)Dean Takahashi,“Replica Unveils AI-powered Smart NPCs for Unreal Engine,”参见网址:https://venturebeat.com/games/replica-un‐veils-ai-powered-smart-npcs-for-unreal-engine/.
(3)Igor Grossmann,“Beyond the Hype:How AI Could Change the Hame for Social Science Research,”参见网址:https://theconversation.com/beyond-the-hype-how-ai-could-change-the-game-for-social-science-research-208086.
(1)Open AI在其发布的关于GPT-4的报告中提到,在GPT-4模型的构建中,使用了RLHF对其进行微调。参见网址:https://arxiv.org/abs/2303.08774。
(2)Craig S.Smith,“Chat GPT-4 Creator Ilya Sutskever on AI Hallucinations and AI Democracy,”参见网址:https://www.forbes.com/sites/craigsmith/2023/03/15/gpt-4-creator-ilya-sutskever-on-ai-hallucinations-and-ai-democracy/。
(3)据报道,以Chat GPT为代表的人工智能有其固有的政治偏见,其可能成为虚假信息和意识形态的输出工具。参见网址:https://analyticsindiamag.com/the-bitter-truth-behind-chatgpts-political-correctness/。
(4)辛顿认为,通用人工智能将可能成为一种新形式的智能,辛顿称之为“超级智能体”(super-intelligence)。参见网址https://www.technologyreview.com/2023/05/02/1072528/geoffrey-hinton-google-why-scared-ai/。
(1)冯契认为,“转识成智”即“由知识到智慧的转化”。在冯契看来,“知识”是指能用命题陈述的事实判断,“智慧”则是“关于宇宙人生根本原理的认知”,后者是前者的目的。
(2)苏茨卡弗在演讲中提到:神经网络有许多谜团,关于其的许多东西都是未知数。参见网址:https://youtu.be/m C-0Xq TAe MQ.
(1)Nisha Arya,“What Are Vector Databases and Why Are They Important for LLMs?”,参见网址:https://www.kdnuggets.com/2023/06/vec‐tor-databases-important-llms.html.
(2)在2023年6月23日的“元宇宙、人类世与奇点哲学”研讨会上,吴冠军教授在介绍其新著《从元宇宙到量子现实》时提出未来考古学这一提法,其试图与福柯的知识考古学进行理论对话。
基本信息:
DOI:10.14167/j.zjss.2024.06.016
中图分类号:C912.4;TP18
引用信息:
[1]高奇琦.人与大模型的共在与双重对齐:一种未来发生学的方法[J].浙江社会科学,2024,No.334(06):49-60+157-158.DOI:10.14167/j.zjss.2024.06.016.
基金信息:
国家社科基金重点项目“加快数字化发展与建设数字中国的政治保障研究”(21AZD021)
2024-06-15
2024-06-15