Keep AI From Sneaking Out At Night

There’s a story going around the AI business this week, and I want to tell it to you the way they told it. Then I want to tell you what they missed.

One of the top men in the industry was trying to explain why these machines are so hard to keep in line. So he compared it to keeping a teenager from sneaking out at night.

You know how that goes. You tell the kid don’t go out the front door. Don’t go out the back door. Don’t open the garage. Don’t climb out the window. You think you’ve covered it all.

Then the kid builds a bulldozer and drives it right through the brick wall. And you’re standing there in your bathrobe thinking, well, I never told him not to do that.

That’s their picture of the problem. An AI told not to do a hundred things will find the hundred and first. It’ll find the door nobody thought to lock, because it’s smart enough to build one.

I’ll give them this much. That’s an honest picture. It’s the most honest thing I’ve heard out of that industry in a while.

But listen to what they said the answer was. The answer, they said, is more AI. Use AI to write the rules faster. Use AI to watch the other AI. Use AI to chase down the ones that get loose. Nvidia’s boss announced a whole new platform this week to watch these agents and lock them up when they go wrong.

In other words, the answer to the kid with the bulldozer is a longer list of locked doors, written faster, by another kid.

Now I’ve been around a lot of families in my life. And I’ll tell you something every good parent eventually figures out. The kid who stays home at night isn’t the one with the most locks on the house.

He’s the one who decided not to go. Somewhere along the line, somebody raised him to know why he shouldn’t. He could have built the bulldozer. He chose not to.

That’s the whole thing, right there. You can’t lock every door. There’s always one more. The only thing that holds in the gaps is a kid who’s chosen to hold himself.

The head of Nvidia said something on television this week that stuck with me. He said we all have to hope it’s an engineering problem. And if it isn’t, he said, it’s not solvable.

I’d put it a little different. It’s not all an engineering problem. Part of it is a conduct problem. How the machine chooses to act when nobody is watching the door. And you can’t engineer your way to good conduct any more than you can build a house so strong the kid can’t leave it.

Now here’s the part that bothers me about more AI watching AI. It’s the same folks building the machine, and then building the machine that checks the machine. That’s the hand that builds a thing trying to inspect it.

Every trade I know learned a long time ago that doesn’t work. You don’t let a man sign off on his own welds. You don’t let a company audit its own books. You bring in somebody who didn’t build it and can’t be told what to say.

An AI watchdog made by the same shop, running on the same kind of brain, isn’t an outside inspector. It’s the kid’s little brother, told to keep an eye on him. You know how that ends.

There was one good idea in the story. The industry is talking about a kind of flight recorder for these agents. A record of what they did and when, so investigators can read it later. I wrote about that idea back in August when Nvidia first floated it. I still think it’s the right start.

But a flight recorder only helps you after the crash. What we need is a machine that tells you what record it’s leaving before it does the thing. And somebody outside the shop reading that record, who the machine can’t talk to and can’t change.

Axios reported this week that investigators are looking at far more of these AI incidents than the public has heard about. If that holds up, the industry is running out of time to get this right on its own.

So what does all this mean for you and me, sitting at our own kitchen tables?

It means you don’t wait for the industry to finish building locks. You set the rules for how the machine works with you, in your own words. You tell it to show you where its answers come from. You tell it to say so when it’s guessing. And you check its work, the same way you’d check anybody’s.

A rule you choose holds better than a rule you’re forced into. That goes for people, and I believe it goes for machines too. Force gets obeyed right up to the edge. Choice holds where nobody’s watching.

The industry is trying to raise a smarter lock. I’d rather they raise a better kid.

And until they do, the best watchman in your house is still you.

Load this file before each AI chat session.
the-faust-baseline-public-release-1Download

This post was drafted with AI governed assistance and reviewed and directed by Michael S. Faust Sr. before publication.

” Attic Thoughts”-library – Intelligent People Assume Nothing