|
|
|
|
|
by msp26
53 days ago
|
|
Not necessarily with speculative decoding. Whitespace would be trivial to predict and they would petty much keep using the same amount of compute as before. I don't think that's their primary motive for doing this but it is a side effect. |
|