Background

The need and importance of conversational AI had long been identified by companies like Google, whose key products are based around natural language processing.

But have you ever tried interacting with Google or your Google home device? Did you found it exactly like interacting with another human?

Here is a sample chat a human has with Google –

Anil — “Hey Google, Can you tell me where is the nearest I can watch Avengers Endgame?”

Google — “Its in the Cineplex theater near Pool plaza, just 1.5 km from your house. Do you want me to book a ticket there?”

Anil — “Yeah. What is the price?”

Google — “The ticket for the matinee show is 100 bucks.”

Probably a 7 year old kid or our grandmothers would find it amusing that a machine responds to our queries, but honestly, the conversations with Google or Google home never are as close as an interesting conversation we have with other humans.

Have a look at a conversation between 2 office colleagues below.

Anil — “Hey Sameer, Can you tell me where is the nearest I can watch Avengers Endgame?”

Sameer — “I think its in Cineplex. You should watch it. I loved Thanos.”

Anil — “What? Isn’t he the shady dark character trying to end half of the World with a snap of his finger?”

Sameer — “Yeah. But I wish I had the same power. I could make our manager disappear with a snap of my finger.”

Anil — “Haha. Sounds like a better world. By the way do you know how much it costs there?”

Sameer — “I guess 100 bucks. But the effects are better with a ticket of 300.”

Which conversation did you like? With Google, You ask a question and it responds with a suitable answer. That is certainly not how humans interact. Unless you are having a conversation with your girl friends father, human beings tend to have conversations based on emotions, not tied to a single subject, often wandering off to a rather humorous or different topics and mostly without the responses being repeated. They might not look like the most optimal way to get answers, but they are emotionally satisfying and refreshing. That is probably why many studies ruled out the possibility of an AI robot replacing a human companion.

Introduction

Google developed 2 transformer based Natural Language Processing models based on BERT, in 2018 and has been using it since 2019 to better understand user searches. They also improved it, altered it to use it for their moderately popular Google Home or Google Nest as it is called today. But the focus has always been on understanding what the user wants to know or search and providing him with the appropriate search results rather than engaging him in a meaningful conversation. Yes, Google searches or Google Home do answer effectively and the conversations are sometimes pretty helpful for someone looking for answers, but we cant deny the fact that its conversational abilities are limited. Try telling Google Home some weird story or ask it some weird question and it will be left perplexed possibly answering its usual response –

“Sorry I did not understand”.

If you try asking the same question twice, you will most likely end up receiving the same response. This is because the language models which work behind have some predefined paths leading to a particular response as an output.

Last year GPT-3 opened the doors to an all exciting world of conversational AI, possibly giving a glimpse of countless possibilities we have in front of us by training a model on an extensive or rather humongous set of data.

LaMDA

On May 18th 2021, at Google I / O, Google announced LaMDA (Language Model for dialogue applications), a language model trained on dialogues or conversations. What is interesting about this model is its ability to collate information and synthesize meaningful responses. Google calls it instilled sensibleness. Similar to BERT and GPT-3 , it is based on transformers, which is a neural network architecture, where the model is trained to find the best next word or phrase suitable for the given context. I believe what it really stands out is the flow of dialogue or conversations it offers which gives it a more humane ability.

Below a small video on how LaMDA was trained to act as planet Pluto and get in a meaningful conversation with the team of Google.

What is notable in the above conversation is that the model not only answered the questions asked about the planet Pluto, but it synthesized some responses to form a meaningful conversation.

How is it different than GPT-3

GPT-3 and LaMDA, both are trained on huge datasets. LaMDA is more focused on conversations or dialogues as against GPT-3 , which has a lot more information to process. This definitely gives an advantage to LaMDA in being more refined and efficient in having interactive conversations, but it could possibly also mean that its knowledge base could be limited. Being a search provider, Google promises to keep the factual precision parameter of responses for LaMDA extremely accurate. It means that the focus of the LaMDA model would be to respond more accurately on any topic along with having an insightful and entertaining conversation.

Conclusion

LaMDA is definitely another step in taking conversational AI to the next level. With Google implementing it in their products, we can be definitely sure that these technologies are not far from reaching us. It should not be too long when we will get rid of those machine-like conversations we have with search engines or home assistants and it would be difficult to distinguish if the last conversation we had was with an AI or human.

Anup