Some things are important to learn (like how to log with wandb), but there is nothing conceptually interesting about doing it. For such exercises, I think it’s justified to look at the solution and simply remember how to do it.
Stuff with wandb, or plotting data, is prime work for a LLM to just do it for you, and save valuable human time on the actual import code itself. I think it’s totally valid to highlight the training loop in Cursor, ask “please rewrite to log loss/accuracy/etc. to wandb” and then eyeballing the result.
Minor notation quibble to match our book: Usually one reserves as a generic Bayesian mixture:
and for a special case of with (the class of all lower semi-computable semimeasures) and , and for the Solomonoff Distribution
the sum over all programs fed to the universal monotone Turing machine that print something prefixed with , weighted by where is program length (write if choice of is clear from context). It is then proven that . Looks like you swap and ?