Today’s voice assistants can answer basic questions. But what if they could also hold long, open-ended conversations on any topic?
At Google I/O, the company debuted a prototype AI system that promises to do just that. The system is called LaMDA, and it’s capable of answering follow-up questions that can lead one conversation point to another.
During the event, Google demonstrated LaMDA, but not as a conventional voice assistant. Instead, the company programmed the system to personify the dwarf planet Pluto, and answer questions about its life on the edge of the solar system.
“I wish people knew that I am not a random ice ball,” LaMDA said in the demo. “I am actually a beautiful planet.”
“Well, I think you’re beautiful,” the human speaker retorted.
“I am glad to hear that. I don’t get the recognition I deserve. Sometimes people refer to me as just a dwarf planet,” LaMDA said in response.
In another demo, LaMDA impersonated a paper airplane. The program was smart enough to ask what the user meant by “good” when it was asked “what’s the secret to a really good paper airplane?”
“Well, my good friend. I must first ask you to specify what you mean by ‘good.’ Some may think it’s how far it goes, some may think how straight it goes, and some may think how flat it lands. So what do you think?” LaMDA replied.
“I care the most about distance,” the human answered.
“Then I would have to say a nice wide wing span is what you’ll want. Also, the paper needs to be really firm,” LaMDA said.
LaMDA, which stands for Language Model for Dialogue Applications, is years in the making, according to Google. The technology can decipher intent by reading the words in a sentence or paragraph, paying attention to how each word relates to another, and then predict what words might come next.
“But unlike most other language models, LaMDA was trained on dialogue,” the company wrote in a blog post. “During its training, it picked up on several of the nuances that distinguish open-ended conversation from other forms of language. One of those nuances is sensibleness. Basically: Does the response to a given conversational context make sense?”
For example, imagine a friend saying: “I just started taking guitar lessons.” A sensible reply could include: “How exciting! My mom has a vintage Martin that she loves to play.” However, AI will often struggle to come up with such a response since they’ve been designed to follow a narrow, predefined conversation path.
LaMDA, on the other hand, can pivot the conversation from one topic to another while trying to make the responses sensible and specific. “The natural conversations are generative, and they never take the same path twice,” Google CEO Sundar Pichai said at the event.
But perhaps most importantly, on the backend, LaMDA can hold new conversations on different topics without Google needing to retrain the AI model every time.
“We’re also exploring dimensions like ‘interestingness,’ by assessing whether responses are insightful, unexpected, or witty,” the company added. “Being Google, we also care a lot about factuality (that is, whether LaMDA sticks to facts, something language models often struggle with), and are investigating ways to ensure LaMDA’s responses aren’t just compelling but correct.”
LaMDA is still a work in progress, and will make errors. “Sometimes it can give nonsensical responses, imagining Pluto doing flips. Or playing fetch with its favorite ball, the Moon,” Pichai said. “Other times, it just doesn’t keep the conversation going.”
Another challenge is making LaMDA understand information from other mediums including images, audio, and video. Still, the tech giant is hopeful it’ll one day use LaMDA to power Google Assistant, Google Search, and Google Workspace. “We’re also exploring how to give capabilities to developers and enterprise customers,” Pichai added.
In the meantime, the company is working to prevent the technology from posing a threat to society. For example, it isn’t hard to imagine LaMDA being exploited to catfish lonely singles on a dating app.
“Language might be one of humanity’s greatest tools, but like all tools it can be misused,” the company said. “Models trained on language can propagate that misuse—for instance, by internalizing biases, mirroring hateful speech, or replicating misleading information. And even when the language it’s trained on is carefully vetted, the model itself can still be put to ill use.”
“Our highest priority, when creating technologies like LaMDA, is working to ensure we minimize such risks,” the company added.
