Can You Trust a 100% Alignment Score? AI Ethics Unpacked

So it is of things onto it after that to make sure that it was safer.

So it's supposed to be GPT six, I think.

So it's supposed to be a step change above Saul even, which already still feels like it just came out.

But now we're on Astra.

So yeah, the, I think the biggest things though, is it's supposed to be their most aligned model.

And there were a few things that were included in the release where they One of the things is they they're benchmarking.

Obviously, these have not been verified by anybody third party yet because it just came out yesterday.

But exploit bench earned a 100 percent on it, which is basically like tests on how much it will like exploit other systems.

Which.

I find it really hard to trust anything that's ever at 100%.

And then there was another one, too, where it was testing how much it, oh, yes, the honeypot cheating, 0% versus 48.2% from the prior model.

So basically saying, like, how much is this thing going to do a thing that it's not supposed to do or exploit other systems? And I wanted to get your thoughts on On that, if you have any thoughts or high-level feelings from that.

Okay, so I guess I am not as familiar with the world of model testing and validation.

So...

Who creates these benchmarks and who runs the tests? Is it like a self-attestation of like, I have run these benchmarks and I'm saying that it's 100% or is it like a verified third party? Like, how does that all work?

I think it is run, I think the ones that they reported were run internally.

They were created by another, by a third party, but then they are run internally by OpenAI.

Somebody on the internet can correct me if I'm wrong there, which I think is interesting from the perspective that like they've reported on these numbers for past model releases as well.

And I don't think I've ever seen one in the past that was like, 100% accurate and 0% on another thing.

And that's one of the things I find most weird about it.

So the honeypot cheating one was 0% versus 48.2% for, I think, Sol.

And that means it's also harder to monitor the reasoning of Astra compared to Sol.

So it makes me wonder, is this really... not cheating or is it making it look like it's not cheating and just being really convincing at it and knowing, like, anything scoring 100% or 0% makes me think, like, ah, okay, it knows what the test is and it's doing the thing according to the test.

That's kind of my knee-jerk reaction, too.

Like...