AirType is a lot faster than it used to be, but still takes a few seconds to load into memory, and init process is too slow.

[sid9102]

Loading serialised bitsets is slow. A better way to do it would be to load the bytes directly into memory, using OpenBitSet would be a lot faster. (Done: 15d192a)

The init process is not good for UX, as the user has to look at a loading bar for no apparent reason. The current encodedTrie implementation also means that inserting a word into the dictionary would require the init process to run again.

[bendbro]

Intersperse words with node bitset, delimiting the words with bitsets, cutting the number of bitsets in half.
Have the initial encoding process sidestep airtrie by implementing airtrie's functions on encodedtrie.