Andy Halliday

Yeah.

So it's for too long.

Don't watch.

Yeah.

TLDW.us.

Yeah.

And I think what I gathered from it was it's analyzing the transcript for the sections that actually have some quality substance.

And then it takes the timestamps for those and pulls that piece.

By the way, I think, I don't know the details on it, but it was built by a single person.

Let's just see what happens if we all just shut our browsers.

Yeah, Jude, call me if you have an update on what happens after we're gone.

I'm going to hit on a small one first before I bounce back to you.

And it is something that actually was published last week.

but we haven't talked about it yet.

And I think it's an interesting insight into how the context data that goes into an inference run affects the model's responses.

And it is a study at Stanford that found that AI chatbots don't have the ability to really distinguish between facts and beliefs when they're expressed by the user.

So if the user says, I believe,

literally those terms, I believe that blah, blah, blah, blah, blah.

It takes that as a fact.

And you could, by prompting the model, prevent that to some degree by saying, look, anything that I say, I don't want you to take as a fact.