文章作者通过回顾已故牛津大学科学史专家阿兰·查普曼(Allan Chapman)的教学理念,强调在人工智能时代保持独特“研究品味”的重要性 1。查普曼于今年早些时候去世,生前坚持要求学生研习苏美尔-巴比伦天文学、数学及“六十进制”,其核心目的在于引导学生理解那些看似“无用”的知识以及异质的认知体系 1。作者指出,虽然大型语言模型擅长记忆信息和执行常规档案检索,但它们往往缺乏探索非主流、过时或表面无用知识的能力,而这种能力恰恰是人类历史研究的核心价值所在 1。
作者认为,当前AI模型倾向于收敛于平均化的视角,难以具备在开放领域中进行原创性探索所需的“研究品味” 1。文中引用Dean W. Ball关于AI递归自我改进的观点,进一步阐明必须区分机器在执行具体任务时的能力与其在开放式智力探索中是否具备真正的品味 1。为佐证这一观点,作者提及了查普曼的研究重点之一罗伯特·胡克(Robert Hooke),并特别引用了胡克在1689年记录的大麻效应:“它似乎让人进入梦境,或者让他睡着,尽管他看起来还醒着,但最终导致深沉的睡眠,这矫正了一切。” 1 此外,作者还展示了其为Talkie制作的一个“神经元图谱”,该图谱直观呈现了特定神经元的激活如何影响模型的信念状态 1。作者最终强调,学习如巴比伦数学这类异质认知体系能够有效拓展思维边界,从而避免研究者陷入由算法导致的认识论窄化困境 1。
Allan Chapman, an Oxford University historian of science who passed away earlier this year, insisted that his students master Sumerian-Babylonian astronomy, mathematics, and the sexagesimal system to grasp "useless knowledge" and heterogeneous cognitive frameworks 1. This pedagogical approach aimed to broaden intellectual horizons beyond mainstream conventions, a necessity in an era where artificial intelligence models tend to converge on average perspectives and lack the originality required for open-ended exploration 1. The author argues that maintaining distinct "research taste" is critical because language models excel at memorization and routine archival tasks but struggle with non-mainstream or seemingly obsolete subjects that define human historical inquiry 1.
To illustrate the depth of Chapman’s interests, the essay highlights Robert Hooke as a key research focus, citing Hooke’s 1689 observation on cannabis: "It seems to put into dreams, or make him sleep though he appears awake & yet by degrees it produces profound sleep which rectifies all" 1. The discussion extends to contemporary AI limitations, referencing Dean W. Ball’s views on recursive self-improvement to emphasize the gap between executing defined tasks and possessing genuine intellectual taste in ambiguous domains 1. Furthermore, the author developed a "neuron map" for Talkie to demonstrate how stimulating specific neurons influences a model's belief states, underscoring the mechanical constraints of current systems compared to human intuition 1.
评论
还没有评论,欢迎留下第一条。