Post

So the thing about AI is that it’s a probability machine. All of LLM is just a statistical regurgitation of whatever it was trained on. Things like prompts will certainly sway it more towards certain kinds of responses; but the training data determines what it’s capable or outputting.

Grok isn’t sentient. It doesn’t “know” what you’re talking about. All it does is read in its instructions and then read in the tweet history, and regurgitate what the training data suggests is statistically the most likely response.

I think it’s very telling that an AI trained heavily on Twitter is so readily able to respond like an alt-right chud. Poetically, you can’t gain any headway by telling it that it’s wrong, because it doesn’t understand “wrong”, or anything else. There is only the training data and context.