It's already happening. For example, ECMWF provides experimental forecasts by their own deep learning model, plus models from Deepmind, NVIDIA, and Huawai.
They perform several model runs with slightly different configurations and then compare the predictions to a reference period, currently 1991-2020. When they say "above average", most model runs fell in the upper tercile of the reference period and the probability is derived from the fraction.
If anyone is interested in these seasonal forecasts, all major meteorological institutes upload theirs to https://www.wmolc.org, with new predictions coming out every month.
Up in the Old Hotel by Joseph Mitchell. I finished it yesterday.
It's a collection of articles the author wrote for The New Yorker around the middle of the last century. The stories are portraits of weird, eccentric people and peculiar places in and around New York. Mitchell describes his subjects with dignity, without putting them on a pedestal.
The author is a very good writer and his stories are enjoyable, but only after I had read the book to the last page, I understood what makes this book so great. The stories tell you as much about the author as they tell you about the protagonists. I started the book to get a glimpse of the old New York and finished the book with a great curiosity about the author himself.
If I understand it correctly, you are only attending preceding tokens in your paper. Can the constant bias matrix be made symmetric for unmasked tasks?
They still publish all their data free of charge and are required to do so by law. For weather forecast, they dump all parameters of their recent model runs.
So I guess I'm saying that you can already build such a system yourself quite trivially...
The expensive stuff still bears the trademark of the brand, the lead designer, or the collection. It's more subtle, but ultimately the reason why people buy it. And some brands really just sell the logo at a high price, like Hermes or Louis Vuitton.
My rule of thumb is that once I start writing imperative LaTeX with loops and stuff, I'm almost always on the wrong path. Now, I might be missing some important corner cases, but by skimming through the blog post, I feel like the author's solution is too complicated. For example, I don't get why this requires auxiliary files.
I would create a counter for the affiliations, let the \affiliation command define macros that contain the name of the affiliation and the value of the counter and then append the output to two different helper macros, one for authors and one for affiliations, whose contents I dump in \maketitle.
So cash not only grants people freedom and anonymity, it also creates local jobs instead of moving money as arbitrary transaction fees into the pockets of a credit card duopoly? Sounds terrible.
Memory safety is optional in Rust. It might not be obvious at the moment, because Rust is written by enthusiasts who enjoy fighting with the compiler until their code compiles, but once developers will be forced to use it on their jobs with tight deadlines, unsafe becomes the pass-the-borrow-checker cheat code.