612: The Sommelier of Sentences

Uh, about watermarking specifically used by anthropic and Claude as, uh, sort of vaguely mandated by the EU and by say sort of vaguely.

I mean, I don't know if they, they haven't stipulated do it this way kind of thing, but, uh, right.

John particularly picked the words adulterated and perverting the text.

I, uh, Like I said, I'm friends with John and I.

He ran that piece by me on Friday.

And another friend of ours who I won't name because I don't know.

Whatever.

It doesn't matter.

It doesn't matter.

But it's somebody who knows a lot about AI and explained all the math behind it and explained that the way that you pick these things is basically just weighting a random one way or the other.

uh i called out to john what you called out and this was completely independent you didn't we hadn't discussed it no we didn't you didn't we

didn't talk about it until that

after right yeah because well you know if somebody runs a piece by me i will give them feedback but i'm not gonna i don't talk about it like that's not a thing that you do you know so Anyway, I also raised the thing that like you can't adulterate something that's, you know, literally the word adulterate means to make less pure.

And the product of this stuff is not pure because there's no thought behind it.

There is no intelligence trying to express and communicate an idea.

And for me, purity of communication would be somehow it's a impossible because, you know, we're always limited by whatever.

unless you're Professor X, I suppose.

You use the words in whatever art you have the capability of creating in order to try to express a thought as clearly and as purely as you can.

Now, LLMs inherently cannot do that because they don't have thoughts.

They are statistical models, incredibly fancy ones.

But they are statistical models picking the next most likely token in order to achieve an output result.

So you cannot adulterate something that is by its very nature like wind rustling through the trees.

Like it's not, it's yeah.

Right.

Or it's, it's, it's a melange of everything that, that, you know, all the training data and the way.