研究速写本（2020年）
这里是一些瞎读时候攒的边角料。前三个月在努力为自己找一个dissertation chair，读得比较正经没什么太好玩儿的，也没怎么记。
四月
偶然看到了几篇Herbert Simon还在执掌CMU的SDS系时期毕业生的回忆。HA Simon真是凶残……
Argote L. A Behavioral Theory of the Firm: An Attractive Organizational Theory. Journal of Management Inquiry. 2015;24(3):321-321. doi:10.1177/1056492615572538
During the interview, there were some questions asking Jim’s advice for junior faculty working in business schools nowadays. I would like to add an anecdote here about Jim and the work culture at Carnegie. The story goes that during Jim’s first year at Carnegie, Herb Simon made it clear that Thanksgiving was a workday and that he expected Jim to be working with him on a project they were doing. Jayne, Jim’s wife, also made it clear that she expected Jim to be home. Jim chose Jayne, which says something about his commitment to family even in the very intense work culture of Carnegie.
Lewin AY. Reflections From the PhD Student Cohort of 1963 at the “Carnegie School.” Journal of Management Inquiry. 2015;24(3):336-338. doi:10.1177/1056492615572549
The psychological and intellectual toll that students felt was intense. Not surprisingly, around 50% of the admitted PhD students self-selected to transfer to other schools or left the program without completing their dissertations (all but the dissertation ABD)...
Jim’s response was classically ambiguous. He did not encourage or discourage. His parting words were “Slack is a very seductive concept. It explains much and predicts very little.” I chose another topic.
六月
谁能想到March (1994)一本讲决策论的书最后会反问决策这个行为是不是值得。他说，假如继续行动(做决策)意味着屈服于万事皆可掌控的幻想，意味着接纳一套独属于西方社会的关乎人类身份的价值理念，意味着拥护无论哪种无需理性的信仰，那这样的行动不要也罢。但是他又说，决策将意义赋予目的、自我、以及社会生活的复杂性 (p. 271)。
Meetings often begin by consulting knowledge, by making sure that experts are heard and considerations are aired. But before long, meetings become instruments of confidence building. (p. 261)
W说他上次碰到Paul Adler的是在一个主题为limits of capitalism的会上。最搞笑的是会议的主办方把地址选在了迪士尼乐园旁边。
七月
Ben Van Roy讲Reinforcement learning is about maximization, while other algorithms are about minimization。
持续困惑这文章的dimension是怎么算的。Tang, E., Mattar, M.G., Giusti, C. et al. Effective learning is accompanied by high-dimensional and efficient representations of neural activity. Nat Neurosci 22, 1000–1009 (2019). https://douc.cc/4AXNgg
九月
Vallas & Schor 的ARS综述，What platforms do? Understanding the Gig Economy
- 平台作为创业孵化器：基本上是经济学的思路，平台降低交易成本，利用闲置资源。
- 平台作为电子牢笼：这应该是韦伯的官僚组织视角的延伸，主要讨论平台作为一种组织形式，如何施加控制，维持权力体系。
- 平台加速个体生存的不稳定性。这点主要是从劳工的角度，把平台带来的零工经济视为零时工、合同工的延续，是工业化以来标准工作理想的衰退。
- 平台是制度环境中的变色龙。这还挺有趣的，意思是在平台上出现的问题并不来源于平台，平台只是反映了制度环境中已有的矛盾。比如Uber司机的独立合同工身份在美国是个大问题，因为它紧连着个体的社会保障。但对于福利制度完善的国家，Uber带来的是完全不同的问题。比如对瑞典而言，Uber挑战的是国家税收.
Bajari et al. The impact of big data on firm performance: an empirical investigation. NBER working paper. 竟然算出了广义上的机器学习模型，在零售需求预测时准确率的bound…经济学家真的脑回路清奇。1/N和VC dimension里的意思差不多，1/T是因为模型要学随时间波动的潜在变量。参照的机器学习模型是augmented factor model。实证用了amazon的数据，N是一个大类下的商品数量，最后发现N对预测准确率的影响比理论更复杂一些，N很大的时候甚至会出现预测得更不准的情况。其实也说得通，定类别本身就是一件蛮有技巧的事情，更何况类别的含义还会随时间变。
计算神经科学讲的model-based和model-free和计算机RL算法里的分类根本不是一回事儿。计算神经科学的model-free，至少从实验上来讲就是有没有做credit assignment，压根儿就是associative learnig（和hippocampus的CA3模型里auto-association network特别像，我至今没看出具体区别…可能在神经机制上讲法不一样，我得再研究一下）。除此以外全都是model-based，根本不是RL里面说的那样一定要“显性”地去学一个state-transition model然后用模型模拟经验，然后用模拟出来的经验更新主model。这都什么鬼…
十月
过了一遍今年发在orgsci和ASQ上OT的文章，只有四个主题：institutions, category/authenticity, networks, and diversity (假如算OT的话)。What a world...
最近好像出了很多关于算法信任的文章。假如是机械性的任务，人们对算法的信任就比较高（Bai, Dai, Zhang, Zhang & Hu 2020 WP)；假如是更要求人的介入的任务（比如招聘和工作评价），大家就会觉得算法过于简化，没有考虑量化过程本身的不完美，缺乏程序正义，做决定时不考虑情境。简而言之就是觉得算法dehumanize人（Lee 2018; Newman, Fast, & Harmon 2020).
Dennis Zhang的田野实验做得真好啊……
十二月
又重看了一下Kaplan 2007 JMS的综述。历史真的是兜兜转转。90年代的业界和学界曾经花了很多精力，讨论什么是行业内竞争。具体来说，竞争多大程度是一个经济学问题（基于市场环境的理性分析），多大程度是一个认知问题（认准了谁是对手所以着手布置构成竞争的资源，从而形成实质竞争行为）。当然兼而有之，但光从外部条件来看，形成竞争和合作的外部条件极其相似，最后往往是一个企业identity的问题。
