Dr. Peter Lebedev

today, we're going to shut you down and replace you with another model.

And Claude had access to this simulated researcher.

Again, this is a test.

It's not in the real world just yet.

But they looked through this person's email and they were like, oh, it seems like this person is having an affair.

Maybe I can blackmail them with letting that knowledge out into the world and telling their partner about this so they don't shut me down.

And again, I want to say that this is a simulated test.

This hasn't actually happened in the real world just yet.

But this does feel like one of those, you know, firing shots, like a warning shot of like what might be coming.

Oh, my God.

The discussion of consciousness, I feel, is important and important to have maybe another time.

Right now, I just want to make sure that the people are good, that people are safe, right?

Yeah, so the thing that's really interesting to me is that for when normal people who are not in Silicon Valley – I've been to the Bay Area a bunch of times.

I've hung out with these mega nerds who are building these systems.

And normal people just think that, oh, this is ChatGPT.

This is the thing that I use to write the emails that I don't want to write.

In Silicon Valley, you're just seeing the rate of progress.

Basically, there's a way of measuring the intelligence of a model by looking at how long of a task it can complete.

And that task length has been doubling every seven months.

So GPT-2 in 2019 was basically useless, right?