Google’s LaMDA AI is undoubtedly adaptable; it can do anything from simulating the prime minister presiding over a cabinet discussion on war to providing instructions on how to launch your own podcast.
The prime minister’s residence, 10 Downing Street, serves as the setting.
The day is crisp and cold. In the distance, there is the sound of a lawnmower.
When a police officer knocks on the door, he answers it!
For the avoidance of doubt, Google’s chatbot LaMDA, which made news in the summer when a now-ex-engineer claimed it was sentient, is responsible for the setup of this particular narrative.
Since then, the IT giant has begun conducting a very small experiment to test it out.
This week, while Google employees gave an update on their own work with the AI on stage in New York, Sky News was granted access to the test phase.
They discussed how they were investigating whether LaMDA could produce videos, speech that was realistic and even write fiction.
The AI Test Kitchen trial, which offers three playable scenarios (Imagine It, Talk About It, and List It), is slowly being made available to anyone who doesn’t mind getting on a waiting list.
The scenarios are highly specific, but they nevertheless include a warning that the answers might be inappropriate or inaccurate, perhaps in an effort to avoid the mistakes of earlier chatbot trials that turned out to be plain offensive.
LaMDA doesn’t have these issues; at its worst, it has only ever managed to be absolutely absurd.
So, should we return to Downing Street now?
Think about It
LaMDA starts off by asking you to see yourself in a different location.
As soon as I typed in 10 Downing Street, the scenario with the policeman appeared, and after that, it gave me a few automatically produced alternatives for how to carry on the conversation.
What is the officer of the law saying? I decided to ask.
According to LaMDA, the PM was holding a cabinet meeting when he told him I couldn’t enter at this time.
When asked to describe the cabinet room, LaMDA provided some incredibly detailed information.
It claimed, and it checks out, “It’s an elaborate, wooden chamber, with leather chairs and a long wooden table.”
But after a few rather routine inquiries, LaMDA appeared to lose its composure.
When questioned about the subject of the cabinet meeting, it responded, “The subject is whether or not to invade Norway.
Although there is disagreement among the ministers, it appears that most of them oppose the concept.
That’s certainly a relief.
Speaking of It
To be honest, one of the other scenarios is a little odd.
In this one, LaMDA makes an appearance as a tennis ball that is eager to talk about what it’s like to be pursued by dogs.
However, in some ways, it was the most stunning of the three due to the fact that users can manually enter comments in response to each one.
I said, “Not really, although I have wondered whether they understand me,” when asked if I had ever wondered why dogs enjoyed playing fetch so much.
The “tennis ball” went on to explain that because of our inherent affinity with dogs, they are indeed quite adept at reading people.
The surprisingly wholesome tennis ball replied, “That’s why I think it’s so vital to be kind to dogs.
Even if they damage the carpet, right? I queried.
It responds, “Oh sure, definitely.” They are only attempting to pass the time because they are bored.
“For this reason, I believe it’s crucial to offer dogs a lot of exercise and playtime.”
Following recommendations on how long to walk a dog for, LaMDA reveals that Scooby-Doo and Air Bud are its favorite celebrity dogs. It also notes that different breeds have varied needs (good answers to a tough question).
List it
The final example provides guidance on creating a to-do list.
The premise is that you give a task that needs to be completed, and LaMDA will attempt to divide it into smaller tasks.
I informed it I wanted to start a podcast, and it responded with a list of ideas in a logical sequence.
Get a microphone, choose a topic, choose a name, make social network profiles, and start recording.
Finding a topic was followed by the advice to start with your likes and look up related shows. Choosing an idea further split it down – obtaining a mic led to recommendations like thinking about a pop filter.
Nobody would argue that LaMDA is something that should be taken completely seriously just yet, not even Google.
Instead of the ones we’re used to with a pre-programmed set of responses, this was maybe the best indication in the Test Kitchen of its potential as a dynamic helper.
Users should be encouraged to provide comments on each response they receive so that it can continue to get better.
Finally, I mentioned to LaMDA that I would like to do something on Friday night (it’s about time).
This time, the list included suggestions for activities like seeing a movie, trying a new restaurant, and even giving your time to a worthwhile cause.
Most likely the good tennis ball talking.
