A machine learning write-up assessing whether we could fix a spurious correlation in a dataset by just collecting some examples and fine-tuning on those examples. We also had interest in understanding what changed as we did this and whether the correlation became spurious in another way.
This could have been a paper, but is better as a blog post.
A machine learning write-up hypothesizing how we can use humans in the loop to create unsupervised models with axes of variation about which we care greatly.
A lengthy exploration into important historical papers in machine learning meets game theory. Of particular note is a deep dive into the origins of the Fermi distribution wrt its ubiquitous use as the underlying probability of whether agent A adapts agent B’s strategy. Everyone who used this approach, including Alpha-Rank most recently, can be traced to Blume '93.