Grant Harvey

So, like, through, I forget if it's... Genetics.

What's the word for it?

Basically the equivalent of AI genetics.

Yeah.

It can pass on, like, if it secretly loves owls, it can pass on its secret love of owls to another model, is, like, the famous example.

So, you know, what secret weird stuff is in there that people didn't take the time to actually skim through and look at when they first built the original one because they just tried to put a ton of data into one system?

You know, there was a pretty good article about this that was basically saying there's a lot of really hateful, bad stuff in some of these large language models that people just never edited out because they didn't look at it deeply enough.

Right.

Like the only way to control for this is to slowly build it up over time by being very deliberate with what data you put into it.

Of course, you don't get the broader generalization that you get doing it this way.

Anyway, my tangent is the unknown unknown that keeps me up at night is whatever the next version of this is, what it's going to think about us.

Yeah, I think Navarina is totally right about that.

The model that is the safest will be, to a certain degree, the one that's most used.

The example I go back to is self-driving cars.

The safest self-driving car will be the one that most people use.

That's why most people are using Waymo these days versus Tesla.

The sample size of Tesla is much smaller, but a lot of people are trusting Waymo because it is one of the safest.

To be fair, Anthropic's enterprise adoption is growing.

In fact, it's growing faster at this point than OpenAI's is.

Anthropic's main problem, in my opinion, is that they don't have the breadth of models and the breadth of use cases that OpenAI has.