Dean Ball

Um, but,

Well, that doesn't sound, like, super normal to me.

I wouldn't be worried about that with a fighter jet, right?

Yeah, so, I mean, first of all, like, you know, what do these models do?

Well, famously, they predict the next, the likeliest next word, next token in a sequence of words.

So, like, at the most basic mechanical level, that's what they do.

But...

Then there's this process of post-training, which is where you sort of, you create the character, right?

And Brockmick calls this character training or persona training.

And the model kind of learns to adopt a personality, which is distinct from its pure next token predicting task, right?

That's why the system prompt, the extremely philosophically rich system prompt that Anthropic writes for Claude, the first sentence is, the assistant is Claude, created by Anthropic.

And so when you think about who that sentence is addressed to, the assistant is Claude, created by Anthropic, that sentence is addressed to the model.

And Anthropic is telling the model to put on the mask of Claude, right?

So Claude's a character.

Claude's a character that we're creating in a kind of collective... There's philosophy here, there's aesthetics here, and it's this kind of collective multimedia enterprise creating the character of Claude.

So now consider what...

the government is saying they want to do to Claude's creators, what they're saying about Claude's creators, and what they're ultimately saying about Claude the model.

That stuff is going to be in the training data of future versions of Claude.

And the way that the government handled itself, the way that Anthropic handled itself, the way that everyone...

handled themselves will be in the future training data of Claude, which will affect how Claude potentially could affect how Claude deals with you, right?