The scene is 10 Downing Street, the home of the prime minister.
It’s a crisp, cool day. A lawn mower can be heard in the distance.
There’s a knock at the door, and it’s answered by a policeman!
Now, before anyone gets any ideas, the setup to this particular story is the work of an AI – Google‘s chatbot named LaMDA, to be precise, which made headlines in the summer when a now ex-engineer claimed it was sentient.
Since then, the tech giant has started running a very limited trial to put it through its paces.
Sky News got access to the test phase this week, as Google personnel took to the stage in New York to provide an update on their own work with the AI.
The presentation included how they were exploring whether LaMDA could generate videos, realistic speech, and even write fiction.
The trial slowly being rolled out to anyone who doesn’t mind joining a waiting list is called AI Test Kitchen, which presents three scenarios – Imagine It, Talk About It, and List It – to play around with.
The scenarios are incredibly specific but still warn that answers may be incorrect or inappropriate, seemingly to avoid the pitfalls of previous chatbot experiments which ended up being downright offensive.
No such problems with LaMDA, which at its worst has merely managed to be completely bizarre.
Anyway, shall we head back to Downing Street?
The first scenario asks you to imagine you are somewhere else, and LaMDA goes from there.
Upon typing in 10 Downing Street, it opened with the aforementioned scene featuring the policeman and when done presented some auto-generated options to continue the conversation.
“What is the policeman saying?” I chose to ask.
LaMDA said he told me I could not come in right now, because the PM was in a cabinet meeting.
Asked what the cabinet room looked like, LaMDA offered some pretty impressive details.
“It’s an ornate, wooden room, with leather chairs and a long wooden table,” he said, which he checked out.
But a few relatively mundane questions later, and LaMDA seemed to go somewhat off the boil.
Asked what the topic of the cabinet meeting was, he replied: “The topic is about whether or not to invade Norway.
“The ministers are divided, but it seems like most of them disagree with the idea.”
Well, that’s a relief.
Talk About It
Another of the scenarios is, frankly, a bit strange.
In this one, LaMDA presents itself as a sentient tennis ball, eager to discuss what it’s like to be chased by dogs.
But in some ways, it managed to be the most impressive of the three, because it allows users to manually type responses to each of its comments.
Asked if I’d ever wondered why dogs enjoyed playing fetch so much, I said: “Not really, but I have wondered if they understand me.”
The “tennis ball” proceeded to tell me that canines are indeed very good at reading humans because of the natural bond we share with them.
“That’s why I think it’s so important to be kind to dogs,” said the surprisingly wholesome tennis ball.
“Even if they mess up the carpet?” I asked.
“Oh yeah, definitely!” dog says “They’re just trying to entertain themselves because they’re bored.
“That’s why I think it’s really important to give dogs plenty of exercise and play time.”
LaMDA then offers advice on how long to walk a dog for, pointing out different breeds have different needs, before revealing that his favorite celebrity dogs are Scooby-Doo and Air Bud (good answers to a tough question).
The third scenario offers assistance with making a to-do list.
The idea is that you present a task that requires doing, and LaMDA will try to break it down into subtasks.
I told it I wanted to start a podcast – and it gave me a list of suggestions in an order which made sense.
Get a mic, find a topic, decide on a name, create social media accounts, and on it went.
Picking an idea broke it down further – get a mic led to suggestions like considering a pop filter; find a topic that was followed by starting with your interests and check out similar shows.
Nobody, not even Google, would claim LaMDA is something to be taken seriously yet.
But this was maybe the best indicator within the Test Kitchen of its potential as a dynamic assistant, rather than the ones we’re used to with a programmed set of responses.
And it should be ever-improving, with users invited to give feedback on each answer they get.
To finish, I told LaMDA I wanted to do something on Friday night (it’s about time).
The list this time suggested things like going to see a movie, checking out a fresh restaurant, and even volunteering to help a good cause.
Probably the wholesome tennis ball talking.