Lately, I've been thinking a lot about what LLM's might be telling us about what thinking is...
Today I'm pondering the slightly dated observation that early LLMs had so many biases on account of their training data (like facial recognition doing a better job on white male faces since it was trained on predominately white male face data sets). Even though we're doing a much better job with our training sets for modern LLMs (perhaps obtained illegally but that's a tangent for another day) - there must be biases always. The nature of human generated data is going to be biased, no matter how large the data set.
Now I'm thinking about how much I love the occasional K-Drama or Tom Clancy novel. But if my entertainment diet consisted strictly of these unsophisticated foods, who would I be? What would my thinking look like if I had only had such a diet my whole life? We have to make the effort to consume an occasional Moby Dick or Seven Samurai every now and then or else our own biases will default to another not-so-useful LLM spewing trite nonsense.
If we're going to have biases, let's bias higher rather than lower. Eat well my friends!
No comments:
Post a Comment