Project Ideas
The Daniel Kahneman Algorithm to moderate our intuitive predictions.
Safe AI usage with Claude Caude
This with FHE ? Should I try ?
Would be interested to test Simile for mechanism design situations.
Study of overfitting in LLMs: given training data size(I trust this man) and current (May 2026) models size, to what extent can we say that LLMs tend to be huge lookup tables ? Signal in the training dataset is actually way larger than 20T tokens given interaction spaces (each token appears in multiple different situations) but still it would be interesting to look at simple DL models for which the number of weights is 10-20% of training dataset size.