undergrad RL at princeton - prereqs are introML, probability, linalg.
ben-eysenbach.github.io/intro-rl

Don't think Safari handles PDF figures well - usually is fixed by switching to .png ime.

Anything by qntm! He has several full-length novels and some free short stories to read on website:

qntm.org/vhitaos

Another good example from mono depth: DepthAnythingV2 uses a teacher supervised only on synthetic data (600k) and students are distilled from its predictions on web images (62M).
Real-world GT is noisy, so fitting to limited, but perfect synthetic data is better for teacher accuracy.

Threw it at o3 and after thinking for 12 min (!!) it gave “O Canada” and A major. You can take a look at the full chain of thought here (chatgpt.com/share/6871b3...). Some highlights:

Maybe I'm not familiar enough with tokenizers, but is this different than just using a very small bottleneck dimensionality in SiamMAE? There seems to be something special about using precisely one token (specifically the [CLS] token?), but it's not immediately obvious why.