- Crick's comma-free genetic code was elegant only for a noiseless channel
- Simple isn't cheap: Kolmogorov complexity has no clock
- Lowering a language model's loss means compressing the text better
- MDL counts the model's length because a fitted model has to be sent
- Memorizing is the right compression of a fact stated only one way
- Solomonoff's predictor can't dodge its bound by oscillating
- Saying which hypothesis wastes bits, so a Bayes mixture beats the two-part code