Ask HN: How do you distill a frontier model?

[npollock]

Is it just obtaining a distribution of the next token predictions, or is it more complex?