Artificial Intelligence has seen several applications in the past years with really high success rate in some categories. There are several applications of chatbots and creating a chatbot for a specific task is really simple. The smart home devices by Google, Amazon, and Apple have been in use by a lot of people. Despite that, very few people actually rely on those devices and often use them for basic tasks like adjusting home lights, knowing the weather or playing music. That’s because there are several tasks that cannot be achieved by chatbots because they find it difficult to understand. Open-endedness is the characteristic of having a baseless conversation like the ones we have with our friends. Sure there is a topic to discuss, but the topics are often switched and could be context specific or an inside joke. AI chatbots find this really difficult.
A problem with creating such a chatbot is the dataset required for training. Will it contain everything in this world and every topic? If so, how can you fine-tune that chatbot? It becomes really difficult and nearly impossible to manage such a chatbot. These giants often hire contractors to manually review the conversations for quality control purposes. Amazon also releases some corpus for general access called as Amazon Mechanical Turk. This means there is some of the data available and what we are left with is fine-tuning. Apart from these corporate giants, there is a company called OpenAI which has created a fine-tuned model which can help in generating text data, also known as GPT-3. This can help in drafting a formal email by just a sentence. Despite these powerful models, how far are we in terms of achieving this goal? There are several chatbots that can engage people in open-ended conversations such as Meena by Google, Mitsuku by Pandorabots, DialoGPT by Microsoft, and Cleverbot by Rollo Carpenter. Some of the chatbots like Eliza date as back as far as 1964!
Despite having these many improvements and a well drafted problem that goes back before the moon landing, why isn’t this goal achieved? Actually there have been a lot of evolutions in these models but this is also a never-ending algorithm. Evolutionary algorithms vs. deep learning algorithms is a critically acclaimed debate just as is console vs. PC. If we were to have a fully functional conversational agent that can understand open-ended conversations then think of the applications it could have! The model can help in developing new ideas and also provide ways to put those ideas in practice.
Even if we consider a step back where the algorithm just finds out the open-endeedness, it would be great deal. Open-endedness marks the next big challenge in the field of computer science as it does not have the typical “problem and solution” approach which most machine learning algorithms have. It is the “road not taken” or rather the “road which is not there yet”. It is purely imaginary and can be shaped into whatever reality you want. A truly adventurous path without a clear destination.