AI Claude Can't Play Pokeman

Few things tickle us more than watching a genius repeatedly push on a door with a "pull" sign.

I chuckled for a long time when I read about the genius Claude repeatedly pushing on that pull door when trying to play Pokémon Red on a Game Boy emulator. Claude did alarmingly stupid things like trying to interact with enemies it had already defeated, walking into walls, and, most humorously, getting stuck in the same corners for hours.

I guess it's a pleateauing glitch that currently has AI standing in front of a rock wall . . . a wall that, depending on who you ask, is a few feet high or a Matterhorn slathered in pig grease.

After three days in Florence earlier this month, I hadn't established my sense of direction and couldn't reliably navigate beyond a 200-yard circumference of my hotel room. This is borderline-retard stuff, especially for a guy whose solid sense of direction is paralleled only by his exquisite fear of homosexual contact.

The culprit was obvious: My phone. I relied on my GPS from the moment I left the train and never looked up, adopting a posture and ambling gate that, in pre-smartphone times, would've prompted a sensible policeman to detain me. It was a frustrating experience, and I have little doubt it robbed me of many aesthetically pleasing moments while walking around that gorgeous city.

And for what? I didn't want to fire up a few strands of mitochondria when I first got off the train? I could've focused for the first day in Florence, then walked around virtually phone-free for the second and third days, but no. I opted to be led around by that (f'ing!!!!) blue arrow on Google Maps like a mongoloid being led by the hand after a few shots of vodka.

But I find solace in that last refuge of scoundrels: comparing oneself to reprobates.

My stumbling gate around Florence probably looked pretty good compared to the morons who need AI to scan a restaurant menu and tell them what to order. I'd rather stumble into the Arno and drown before I get to that stage.

Contrast stumbling me and the menu-challenged with Claude. Claude always knows where to walk and what to order from a menu, but he can't play Pokeman Red without making ludicrous mistakes, like getting stuck in a corner for hours or days at a time.

AI's resident philosopher, Scott Alexander, author of the compelling "Meditations on Moloch," says it's an "agency problem." AI is great at book learning, he says, but terrible with agency.

AI is built on large language models. It sifts through tens of trillions of words to build its entire framework. That's why it's great at book learning.

But that means AI is exclusively left-hemispheric. The left hemisphere runs on words, the more explicit, the better, and it has no use for things like the gestalt, the poetic, and the intuitive, more oblivious to them than Helen Keller is to that ant on the other side of her room.

So when AI Claude gets stuck in a corner, it'll stay there, bumping against the walls for hours, because its rationalist conclusions tell it that's the correct thing to do. It can't see a way out by looking at the larger, and often unspoken, contours of Pokémon Red.

That's why the agency problem is huge and, I'm pretty sure, has AI looking up a pig-greased Matterhorn and not just a garden wall.

Agency requires a principal. It's the first thing you learn in Business Law 101. The agent works for the principal. If there's no principal, there's no agency.

In the world of neurons, the left hemisphere is the right hemisphere's agent. It is, to use McGilchrist's formulation, the right hemisphere's emissary. The right hemisphere principal appreciates the non-explicit aspects of existence (gestalt, poetic, intuitive) and can see the big picture that always transcends what can be seen or grasped. That's why it's the principal and the left hemisphere ain't.

But AI shuts out the right hemisphere, leaving the left hemisphere no principal to work for, and without that principal, there is no agency arrangement. The left hemisphere then just keeps doing its thing, whatever that thing is, without regard to big picture that the right hemisphere takes into account.

As long as the left hemisphere just flails away like Claude in that corner, it's no big deal.

The problem, of course, is that it won't just stay in that corner. It'll break containment eventually, and when it does, it'll rush into the next ridiculous situation.

At one level, it's all quite humorous, until corner-bumping Claude gets enormous power in its hands. That's when the funny stuff gets scary.

Child’s Play, by Sam Kriss

Tech’s new generation and the end of thinking

Harper's MagazineSam Kriss