format - board is missing in lookups: [dobrochan]
what we have [Array
(
    [0] => >>dobrochan/1254
    [1] => dobrochan
    [2] => 1254
    [3] => 
)
]
format - (>>/malform/35) thread is missing in lookups: [dobrochan][1254]
format - board is missing in lookups: [dobrochan]
what we have [Array
(
    [0] => >>dobrochan/1255
    [1] => dobrochan
    [2] => 1255
    [3] => 
)
]
format - (>>/malform/35) thread is missing in lookups: [dobrochan][1255]
- Endchan Magrathea
thumbnail of FjeBMQpagAEjSZR.jpeg
thumbnail of FjeBMQpagAEjSZR.jpeg
FjeBMQpagAEjSZR jpeg
(46.88 KB, 769x740)
thumbnail of FjeBMQqaYAE_3pM.jpg
thumbnail of FjeBMQqaYAE_3pM.jpg
FjeBMQqaYAE_3pM jpg
(155.9 KB, 771x771)
>>/dobrochan/1254@603
I get it. These are fair remarks, a good stenographer is better than a whisper, because he can moderate a discussion in incomprehensible situations, and actively jump into its context, and load himself with a specific lexicon, and use multimodal data and generally predict the following phrases. Cool. And Whisper can write her son instead of Kherson, I don't deny it.
But it's like chess. Many grandmasters offered their brilliant heuristics, and perhaps they could be useful, but in the end, games with full information crushed MCST, and then stupid self-play, because it is easier and cheaper. When the whisper paradigm, that is, the transformer encoding-decoding the log-mel spectrogram in the general case, comes to the plateau (and until it comes out, they have done a little bit more and I see that there are fewer errors), either new heuristics and interaction channels will be screwed up, or all this will be irrelevant.

>>/dobrochan/1255@603
Tashchemt can already be arranged
https://threadreaderapp.com/thread/1600890243452137472.html