Allie K. Miller

I am a perfect little unique thing.

No one's ever lived this life.

And it helps me think through that problem.

So the accuracy of these systems right now, the best models have a hallucination rate.

Yeah, good call.

A hallucination rate?

It's like taking ayahuasca or mushrooms.

Just making things up.

So when we talked about how these systems are trained, right?

We said, give it tons and tons of like millions of gigabytes of information.

So the first thing is that these systems were not trained to be factual regurgitators.

So the fact that it's so accurate all the time, even with these couple mistakes, the fact that it gives answers that outperform PhDs is actually pretty miraculous, right?

The remainder of it, when it does hallucinate, we're getting to the point where models have a about 1% hallucination rate, meaning like you ask it 100 questions and maybe 1% of the time it doesn't answer it on the first or one of the first 50 tries.

different benchmarks, whatever.

It's just, it's when AI is spewing incorrect stuff that it's just like, maybe Mel has a Lamborghini.

And there are ways to increase the fact so you can give it access to the internet so that things are cited and you can check the sources.

You can then check the source of the Lamborghini thing and prove that it's not true.

Whether you knew it or not, you just said something that took researchers years to figure out.

What do you mean?

So we're just now seeing research around this space of why do we get things wrong, okay?