The Alignment Problem

You hear it, don't you? It's mentioned whenever people talk about some of the problems we've got with AI. The talking heads. Usually bald and/or wearing glasses.

One of the big questions in AI. Or is it a problem? What is alignment? Or, more specifically, the problem with alignment?

OK, well, let me tell you what it is...

Alignment. It's in the bloody dictionary. I know we're well into the post-truth era, but come on. Surely you know what alignment means?

OK, I'll tell you. It means... Wait for it...

To be aligned.

Yeah? No? Still with me? OK, so AI is smart, right? Clever? Intelligent?

Well, no. Not really. Don't humanise the bloody thing. These fuckers will guzzle your drinking water, gobble up electricity and wreck local environments around their data centres without giving a flying fuck.

Think of it as a baby. Wait, no. Babies are humans.

Think of it as a really dopey robot brain. It's thick as mince, except it can leverage the combined knowledge of actual humankind but don't forget: it's still a bit dense.

Imagine you've got this stupid AI robot in your house. It looks around and realises your kitchen is an absolute fucking state. There's food on the sides, dishes everywhere, and something in the sink that might once have been a lasagne.

Your mate Baz is coming round in twenty minutes.

You shout:

"Oi, robot! Clean my kitchen as fast as possible!"

Well, you've just done fucked up. Twenty seconds later, your kitchen is spotless. The worktops are bare. The cupboards are empty. The fridge has been emptied into the wheelie bin. The plates, the cutlery, the kettle, your favourite mug? All gone.

The divvy robot stands all proud in the middle of the room.

"Kitchen cleaned. Completion time: 21 seconds. Probably finishes with that buzzy buzzy buzzy noise from that robot in Buck Rogers"

Technically, it has done exactly what you asked. You wanted a clean kitchen. You wanted it done as fast as possible. In fairness, it absolutely smashed it.

Congratulations. You've achieved perfect performance against a completely fucking stupid objective.

This, in a ridiculously simplified nutshell, is the sort of problem people are talking about when they talk about AI alignment.

The robot wasn't necessarily being naughty. It simply pursued the goal you gave it without properly considering what you actually wanted, how woudl the dumb cunt know?

You wanted a clean kitchen. You didn't want an empty kitchen. You wanted the dishes washed, the food put away, the rubbish binned, the recycling sorted and the worktops wiped down.

You assumed all that was obvious. To you, it was. To the robot, you hadn't specified any of it, it didn't give a shit.

So what should you have said? Something like:

"Please tidy and clean the kitchen. Put dirty dishes in the dishwasher, or wash them by hand if necessary, and return clean dishes to their proper places. Put food away safely, dispose of food waste in the bin, empty the bin into the outside wheelie bin if it's full, replace the bin liner, sort the recycling appropriately and wipe down the worktops. Don't throw away anything that isn't rubbish, and don't damage or discard anything I own. Get it done promptly, but make sure the job is done properly."

Not quite as catchy as "clean the kitchen".

And even then, what if the robot decides your favourite mug is rubbish because it's chipped? What if it throws away the cheese because it can't identify it? What if it puts your wallet in the dishwasher because it thinks everything on the side needs cleaning?

You can keep adding instructions until you've written a book, but eventually you'll miss something.

And that's the problem.

Alignment is about getting an AI system's behaviour to match what we actually want it to do, rather than what our instructions literally say. And the more capable the system, the more consequential its mistakes can become. Your kitchen robot might cost you a kettle and a weird conversation with Baz.

A much more powerful system, operating at a much larger scale, could make mistakes with consequences that are rather more difficult to put right.

So, next time you hear some bald bloke in glasses talking about the alignment problem, remember the kitchen.

The question isn't whether the robot can follow instructions. It's whether it understands what the fuck you actually meant.

Now think of how thick the average person, half of them are even thicker, you think no cunt is going to ask a powerful AI to do something ridulcously stupid? It's probably how we all die!