> I think I don't really understand the benefit of data portability in the situation.
Twitter was my home on the web for almost 15 years when it got taken over by a ... - well you know the story. At the time I wished I could have taken my identity, my posts, my likes, and my entire social graph over to a compatible app that was run by decent people. Instead, I had to start completely new. But with ATProto, you can do exactly that - someone else can just fork the entire app, and you can keep your identity, your posts, your likes, your social graph. It all just transfers over, as long as the other app is using the same ATProto lexicon (so it's basically the same kind of app).
Exactly. Anything that's ever been public on the internet is never really gone anyways, and it's unsafe to assume so. This is similar to publishing a website or a blog post. Plus, from a practical (non-opsec) point of view, you can delete items (posts, likes, reposts, etc.) on ATProto, and those items will disappear from whatever ATProto app you are using - usually even live. You need to dive into the protocol layer to still see deleted items.
Out of curiosity, does that mean that GPUs are indeterministic only in their exact execution run-time behavior, or also in the actual output they produce? If the latter, is it something to do with the indeterministic order of floating point operations causing small output differences, or something completely different?
Twitter was my home on the web for almost 15 years when it got taken over by a ... - well you know the story. At the time I wished I could have taken my identity, my posts, my likes, and my entire social graph over to a compatible app that was run by decent people. Instead, I had to start completely new. But with ATProto, you can do exactly that - someone else can just fork the entire app, and you can keep your identity, your posts, your likes, your social graph. It all just transfers over, as long as the other app is using the same ATProto lexicon (so it's basically the same kind of app).