AI bot edges closer to human-level proficiency: Google’s Meena

Image: Olhar Digital

Chatbots are all over nowadays. Their applications range from the world of entertainment to more useful applications such as customer care. The major challenge with most current chatbots is that they are domain-specific.  

In a recent paper released by Google research titled “Towards a Human-like Open-Domain Chatbot”, the researchers managed to do something that has never been done before. To begin with, the researchers created a metric for measuring the quality or humanness of a conversation. This metric is called Sensibleness and Specificity Average (SSA) and this captures basic but important attributes for human conversations.  


Meena, their latest open domain chatbot was trained on a 2.6 billion parameter neural conversation model using 341 GB of text sifted from the public domain. The specific model used for training the model was discovered using a method called evolutionary neural architecture. The method takes a leaf from evolution and selects the model that evolves to have the best-desired properties.  

Meena comes very close to human-level conversations in all domains as measured by the Sensibleness and Specificity Average (SSA). 

As Artificial intelligence becomes better at some of the most important tasks for humanity, industry, and governments should always have a strategy of harnessing these new capabilities for good. The rate at which these developments are taking place is astounding and everyone should be involved in policy. formulation so that as humans we are not rendered obsolete by our own creations.  

In our analysis, based on the challenges being faced in the driverless car domain, the remaining 7% to get to human-level proficiency will be the most difficult. The difficulty will increase as we get closer to human-level ability. 


Please enter your comment!
Please enter your name here