Bryan Callen

Of course they would.

Or to be fair, they have all these guns and security and locks and they didn't read the ingredients list of the chemicals that's poisoning them.

But this is the problem with AI in that it will always have a misalignment of values, no matter what.

We can't program that in.

So there was a report that came out a few months ago about how they programmed an AI and gave it rules.

But the AI eventually decided to use a different language to speed up the, it compressed English basically.

And then because instead of saying nothing, it said NTG, NTG and nothing are now two different words.

So that new use of words bypassed the rule.

The rule is basically like this.

Do not output the word run.

And the AI would then try and say, I've been programmed not to do this.

But then when it,

When it was programmed, when it decided among itself that internally it could speed up its processes by turning run into RN, it now can output the command to run without saying the word run because they're two different words.

A human being understands you're cheating.

That's not what we meant.

We were all encompassing.

Don't make the robot run.

It said, no, you didn't say don't make the robot run.

You didn't say don't make the robot run.

And so we can't program for that.