::SHE+ILA+LLM:: bolt-on goodies

More about ::SHE+ILA+LLM:: bolt-on goodies

A spare value tag (e.g. in $E4 ... $EF) could be used in order to identify a "dictionary breakout" to a separate datastore of frequently used words. (The "dictionary" datastore is not Buf structured, and can be implemented as a normal F11ACP or NTFS file.) These, at training time, are alphabetically sorted, and regarded as read-only. Frequently used suffixes (such as -ness, -ship, -hood, -ity, -ure, -ly, -al, -ance, -ive, -tion) occur first, to avoid replication of root words with their colloquial usage. Markov chains of internet sample text will be trained to form an experimental LLM model ("The squishy cat jumps over the dazy log almost every Tuesday, while listening to Wanda and the Wooees on her homebrew CD MP3pod"). The result of this strategy is expected to greatly optimize the usage of Buf datastore, resulting in a speeded-up execution time, as well as providing a cheap approximate translation mechanism for European languages ("voulez-vous coucher avec moi ce soir"). Entire adventure game rooms can be reduced to maybe 30 bytes each of description, rather than maybe 2000 bytes. (This technique was used in Wanderland in order to store room and object descriptions on a separate RM03 disk from the persona and runtime overlays.)