The COVID-19 pandemic has dramatically accelerated the adoption of conversational AI due to unprecedented spikes in customer service traffic.

When the world returns to a degree of normality, adoption of digital channels, including conversational AI technologies, is likely to continue at a much higher rate than before the pandemic. Companies already on the way to adopting these technologies will therefore have an advantage.

Google executives presented the last research and technologies of the big firm. One of them stole the show: LaMDA, a conversational AI capable of having human-like conversations.

LaMDA stands for “Language Model for Dialogue Applications.” Following from previous models such as BERT and GPT-3, LaMDA is also based on the transformer architecture, open-sourced by Google in 2017. This architecture allows the model to predict text focusing only on how previous words relate to each other (attention mechanism).

In that sense, LaMDA is similar to other existing models. However, there’s a crucial distinction between this system and other chatbots. LaMDA can manage the “open-ended nature” of conversations. As VP Eli Collins and Senior Research Director Zoubin Ghahramani explain in their blog post, human conversations have this distinctive chaotic feature. We can start with one topic and end up in a very different one a few minutes later. We tend to derive conversations by connecting topics in the most unexpected ways.

By tackling these situations, LaMDA will revolutionize chatbot technology completely. A chatbot with these abilities could perfectly engage in natural conversations with people. We could ask for information or consult the internet more naturally.

LaMDA was trained in dialogue, the same as its predecessor, Meena, another conversational tech that Google presented in 2020. Meena proved that chatbots could talk about virtually anything. It was trained to minimize a training objective they called perplexity, a measure of how confident is a model in predicting the next token. They found perplexity correlates very well with human evaluation metrics such as the SSA - sensibleness and specificity average - which is very useful to evaluate the quality of the chatbots.

However, LaMDA went a step further. It excels at detecting sensibleness - whether a sentence makes sense in the context of a conversation -and is better able to keep its responses specific. As the authors note in their post, a response like “I don’t know” could be always sensible, but very useless nevertheless.

An expert says, it has developed with all collected data from us from all the chat boots Google is connected too, Google search, Google Assiatance and Google AI.

During the conference, Google’s CEO Sundar Pichai displayed a demo of LaMDA’s versions of Pluto and a paper plane in two different conversations. They didn’t need to fine-tune LaMDA to change it from impersonating Pluto to a paper plane (although Collins and Ghahramani acknowledge that it can be fine-tuned for better performance).

LaMDA is the next big thing in conversational AIs. We’ll have to test it ourselves to see the degree to which it appears to be human. But from what we’ve already seen, it’s promising