My guess is that they found a bug with their implementation of the model using the weights Google released. These bugs are often difficult to track down because the only indication is that the model is worse with your implementation than with someone else's.

Taking the Amazon example in Part 2:

For e-books (simpler), Amazon gets 30% for running the store, doing advertising, etc. and then authors get 70% [1].

For print books, I'm a little less clear but it appears Amazon buys the books for roughly 50% of list[2] which for Hachette in 2025 is $26.50 so Amazon pays $13.25 to the publisher and then Amazon retails the book for $14.84. So for $100 of books sold on Amazon, $89 goes to the publisher and $11 goes to Amazon. It appears that the cost to produce these books is maybe $2/book (though I'm very unsure on this, this is a guesstimate from public data) and then the rest flows back to authors, advances, etc.

Amazon.com (not AWS) has a 7% profit margin in North America (FY25), so of that $11 they get in revenue they get $0.77 in operating profit.

Ok and this also annoyed me: you say $1.7T/y is $10.5k/worker, which is accurate. but then you say for the average household it's $26k/y. This is not true. There are 134m households in the US [3] so it's $12.6k/y for the average household. Maybe you meant something else like the median household but it seems more likely you just said ~2.6 people/household and multiplied the number of people/household by cost/worker. This is obviously wrong and you should have caught errors like that earlier.

[1]: https://kdp.amazon.com/en_US/help/topic/G200644210 [2]: https://www.readersfirst.org/publisher-price-watch [3]: https://fred.stlouisfed.org/series/TTLHH

The point I was trying to make with "RL is only necessary once" is that you can embark on a single self-play loop getting better and better, and this will get you to something close to the frontier. Once you're at the frontier, the frontier doesn't move very much, so you have quite a while (decade?) where it's totally fine to distill from the RL games.

On correction histories -- imo I correctly described what they do. Cosmo was annoyed by the word "adapt" but what I described was the adaptation.

On SPSA -- you don't have a gradient! you don't do backprop! this is what i was trying to get at.