SOTA models of different architectures and parameter scales exhibit highly consistent failure patterns on the same set of hard samples, suggesting that the performance bottleneck stems from shared deficiencies in training data rather than architecture itself.
大多数人认为不同架构的模型会有不同的失败模式和弱点，但作者发现无论架构和参数规模如何，SOTA模型在相同困难样本上表现出高度一致的失败模式，这表明性能瓶颈源于训练数据的共同缺陷，而非架构差异，这一发现挑战了模型多样化的传统观点。
we have all of these huge applications that are gathering all this data uh and it's out there and theoretically is our data sort of but in reality they control it and you can't actually link the data to each other you only link to accessing the data through their application
for - quote - silos - internet limitations - location addressed server architecture limitations - silos - cannot link data from each silo - Juan Benet - IPFS
My UIs are data/store driven. The UI is just a way to visualize the data. Your data could flow through all of of the extensions and the extensions can make decisions (e.g. setting visible to false). Like middlewares in a Connect/Express/Polka app. And the UI doesn't even know about all this, it just updates with the current state and makes sure it's consistent.
The Coerciveness of the Primary Key: Infrastructure Problems in Human Services Work