#predictive text

[winterthebeau]

Transphobia check! Type "trans women are" and let your phone finish it for you

[wyrm-hlaf]

I wrote a script to make a rudimentary predictive text generator for Worm:

It's not an ai or LLM, it's just a Markov chain that I built by splitting all of Worm into a list of words, then saving it as a dictionary of each one, two, three, four and five-word sequence that appears in worm, overlapping (the dict is 5M items long) along with each word that comes after each sequence. Then I just use that to generate probabilities and it can spit out something that's almost coherent!

It can (sort of) do dialogue because when parsing the text I had it look for words wrapped in quotes and mark them with a special character. When generating text it removes the character and adds the quotes back in. This means that technically, "the" and "the"(spoken in dialogue) are two different words from the script's point of view (RIP file size)

[mymangamemes]

Predictive text when I ignore it for the 100,000,000th time

From Scapegoat Baby by Makuno

[bunniechuu]

[onecells]

Google circa 2000s:

i bet i can predict what you want to search, if you start typing, i'll predict the rest! *proceeds to be more or less correct sometimes at predictive seatches*

me: aiight :) that's pretty cool :) sometimes 👀

Google circa 2020s:

i bet i can Predict EVERY THING..!!!!👀 ALL THE TIME. HAAAHAHAHAHHA

me: ........n....n-no... 👀...

[krumpkin]

I've had trouble like this with predictive text 🙄

[ymas00]

Predictive Text

[somethingusefulfromflorida]

The novel A Canticle for Leibowitz begins hundreds of years after a nuclear war brought humanity back to the stone age. 26th century medieval monks at a Catholic mission in Utah work dutifully to preserve whatever pre-war knowledge they can find, but with no real interest in what the knowledge was or why it exists. They do not know why old buildings made of stone had metal poles in them (rebar), they think blueprints are works of art that should be copied and illuminated rather than used to build what they depict, and when one of the Brothers finds an old book with half the pages rotten away he invents a mathematical formula to derive the missing text one word at a time. It takes him years, decades, to figure out the most likely string of letters that would follow from the beginning of one sentence fragment and feed into the end of the next. The result is meaningless word-salad made up of statistically common groupings of words which, taken together, form no grammatically accurate sentences. He assumes that's just what 20th century English sounded like.

That is exactly what chatgpt does, but much faster. Chatgpt may sometimes appear to be more accurate than the monk, but only because it has much more data to pull from when inventing hypothetical sentences (when it doesn't lift the sentences verbatim from the source), while the monk had to do all his math with slide rules and handwritten tables curated from Cold War-era little blue books. Cell phones have had this feature for years with predictive text, but that doesn't make it any good.

xkcd 1068, Swiftkey (June 13, 2012)