> Unlike tokenization, BLT has no fixed
vocabulary for patches.
iiuc this means: the vocabulary of patches is not known prior to training.
I guess once training has established a vocabulary of patches, that same fixed vocabulary is used for inference (if this is not true I don't see how it could work).
Imagine if they added something like this to UE5 licensing:
If your game has not been updated in N years...
1) Internet Archive can distribute it for free
2) Let people distribute modified versions that does not need license key or whatever copy protection.
Harder but extra cool: To get a UE royalty discount, put source code in escrow set to release it if game not updated in N years.
> He slammed the door in my face, loud and sharp, like an acoustic lemon.
I think its not really about the sound, it is about the response you might make after a big hit of lemon flavor: puckering up, wincing, pulling your neck back - I can definitely imagine responding to a distasteful and loud sound in the same way.
Also, I felt this implies the proper-historically-correct-techniques are much better than the wrong-modern-impression-techniques, and would win in combat.
Makes sense, but it would be cool evidence to see simulated combat to test this.
Because 1) WWII was more of a total war scenario for the US, where a country is dedicating as much resources as possible to win (US is not doing this for Ukraine), and 2) developing technological supremacy takes lots of time, and can't be rushed, so it must be done ahead of time (and continue in peace time to maintain supremacy).
I would also add that after Vietnam war America became more reluctant to put Americans in harm’s way. Is almost impossible politically to institute military draft in America: USA did not do it when it was fighting two wars at the same time (Afghanistan and Iraq).
This helps explain at least partially why America is so keen on military tech.
I would also add to that the public perception that a wonder weapon (nuclear bombs) ended WWII, despite the fact it was only one of multiple more mundane factors that essentially guaranteed an Allied victory (and that's only counting the Pacific, in the European theater the "Wunderwaffe" lost the war)
I wish I knew more about how one would stop instances full of malicious spam bots from flooding the services, but as for DMs, my understanding is the fact that everything is public through the AT protocol is probably why it doesn't yet have DMs (though maybe through public key crypto, one could have something). Maybe in the future, BlueSky will allow for messaging by tying accounts to some messaging protocol or services.
Can a server hosting an account read all of the data belonging to that account?
The feature I'd like to see from federated social network like this is encrypt everything to keys only held client side (in the style of keybase), so servers can't read content (only some surrounding metadata so they know which other servers to communicate with about the content).
iiuc this means: the vocabulary of patches is not known prior to training.
I guess once training has established a vocabulary of patches, that same fixed vocabulary is used for inference (if this is not true I don't see how it could work).
Right?
reply