3 ms·
I stumbled upon an eerie discovery deep within GPT
- its-summertime 3y agoThe "Drunkard's walk" is a similar understanding, https://en.wikipedia.org/wiki/Random_walk https://en.wikipedia.org/wiki/Random_walk Given the current location (e.g. the entire text so far), to go one of several directions (the next token), depending on their probability to be the next destination (the new text) Being "drunk" is the key aspect of that personification though: The agent doesn't see further than the current junction and doesn't remember anything of the past (hence why the current location needs to represent the entire text, and not just the current token)
- eternityforest 3y agoYou could use this for a Lovecraftian horror game. Make a text adventure, but every room would have a few dozen variations generated training a model to recover words from embeddings, and randomly corrupting the embeddings. The lower your sanity meter or the closer to R'leh, the more corrupt the embeddings.