Google is testing a new AI chatbot - we put it through its paces


From the imagination of the prime minister holding a cabinet meeting on war to a guide on how to best start your podcast, Google's LaMDA AI is sure to be flexible.

The scene is 10 Downing Street, the Prime Minister's home. It's a nice cool day.

Now before anyone has any ideas, setting up this particular story is the work of an artificial intelligence: Google's chatbot called LaMDA, to be exact, which made headlines in the summer when a now former engineer claimed it was sensitive.

Since then, the tech giant has started running a very limited test to put it to the test. We got into testing this week when Google staff took the stage in New York to provide an update on their AI work.

The presentation included how they were exploring whether LaMDA could generate videos, realistic speeches, and even write fiction. The test that's slowly rolling out for anyone who doesn't mind joining a waiting list is called AI Test Kitchen, which features three scenarios -

Imagine It, Talk About It, and List It - to play.

The scenarios are incredibly specific, but they still warn that the answers can be wrong or inappropriate, ostensibly to avoid the pitfalls of previous chatbot experiments that proved downright offensive. There have been no such problems with LaMDA, which at worst managed to be completely weird.

Anyway, are we going back to Downing Street? 

imagine 

The first scenario asks you to imagine you are somewhere else and LaMDA takes it from there.

When she wrote 10 Downing Street, she opened with the above scene with the cop and, when he finished, presented some auto-generated options to continue the conversation. "What do the police say?" I chose to ask.

The MDA said he told me she couldn't come in right now because the prime minister was in a cabinet meeting. When asked what the toilet room looked like, LaMDA offered some pretty impressive details.

"It's a decorated wooden room, with leather chairs and a long wooden table," she said, which has been verified. But a few relatively mundane questions later, and LaMDA seemed to be boiling a bit.

When asked what the cabinet meeting was about, he replied: "The question is whether or not to invade Norway. "The ministers are divided, but it seems that the majority does not agree with the idea." 

Well, that's a relief.

Tell me about it 

Another of the scenarios is, frankly, a bit odd. In it, LaMDA presents itself as a smart tennis ball, eager to discuss what it's like to be chased by dogs.

But somehow, she managed to be the most impressive of the three, because she allows users to manually write responses to each of their comments. When asked if I'd ever wondered why dogs have so much fun playing fetch, I said, "Not really, but I wondered if you understand me." 

The "tennis ball" went on to tell me that dogs are 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 healthy tennis ball. "Even if they ruin the carpet?" I asked.

"Oh yes, definitely!" He says. “They are just trying to have fun because they are bored.

"That's why I think it's really important to give dogs plenty of exercise and time to play." 

MDA then offers advice on how long to walk a dog, noting that different breeds have different needs, before revealing that their favorite famous dogs are Scooby-Doo and Air Bud (good answers to a tough question).

List it 

The third scenario offers assistance in creating a to-do list.

The idea is to present a task that needs to be done and LaMDA will try to break it down into subtasks. I told him I wanted to start a podcast and he gave me a list of suggestions in a sensible order.

Grab a microphone, find a theme, decide on a name, create social media accounts and that's it. Choosing an idea divided it further: getting a microphone aimed at suggestions like considering a pop filter; searching for a topic was followed by starting their interests and watching similar programs.

Nobody, not even Google, would claim that LaMDA is something to be taken completely seriously at the moment. But this was perhaps the best indicator within Test Kitchen of its potential as a dynamic assistant, rather than the ones we're used to with a series of programmed responses.

And it should be constantly improving, with users encouraged to provide feedback on every response they get. In closing, I told LaMDA that I wanted to do something on Friday night (on time).

This time around, the list suggested things like going to the movies, visiting a new restaurant, and even volunteering to help a good cause. Probably the healthy talking tennis ball.

Post a Comment

Previous Post Next Post