SnoopJ :python_logo:🫠 on Nostr: npub19d9p0…kf02p npub1jzgxd…ehc3a These models will do poorly with just about any ...

npub19d9p04u4xfysdy92fycw947jrca3xve2gnsauysshzewxvmz8dms6kf02p (npub19d9…f02p) npub1jzgxd3ylxrpqcr90zss3ttyuk20cr6y9r7ewz2hyauvr4n0xzs8sqehc3a (npub1jzg…hc3a) These models will do poorly with just about any task whose principal input involves repetition of hard-to-distinguish symbols, because the output is just statistics soup based on the training data. There's gonna be plenty of stuff in the training data that looks like ([(([]))])

Another sample task: having the model manipulate a given stream of bytes and do fairly simple transformations. Mistakes that no human would ever make are easy to come by.