First Chatbots then Artificial General Intelligence.

Which current applications are sowing the seeds for the first traces of Artificial General Intelligence to germinate ? My bet would be on chatbots.

As you all know, there has been a Cambrian explosion of startups that build chatbots in recent years (Here are just a few). The chatbots themselves have focused on a range of enterprise verticals including Healthcare, Customer engagement, automotive, real estate, among others. On the consumer side, we have seen Alexa, Siri, Google, Cortana, Bixiby in the form of smart speakers, and they are also increasingly embedded into various household gadgets to reduce friction and encourage interactivity. My theses is that chatbots will be the primary route through which we will get to the first developments in Artificial General Intelligence.

Where are we today ?

The chatbots of today have been built using the remarkable developments in ‘Weak AI’ that have been brought about by the improvements in deep learning neural nets in the past decade. ‘Weak AI’ deals with building machine learning models for very specific problems such as speech recognition, object detection, language translation among others. These developments in ‘Weak AI’ can be stacked in interesting ways to create automated interactive experiences for end users. For example, you could use a speech detection neural net combined with a context identification neural net and finally have pre-defined patterns to build an application that lets users book their hotel rooms, process a store return etc. ‘Strong AI’ is the much harder problem of building an AI that can make logical inferences, have common sense and when the chatbots of today hit these boundaries then they often transfer to a human agent or blank out. ‘Strong AI’ is what we would call ‘General Intelligence (GI)’ or ‘Artificial General Intelligence (AGI)’ and while we currently do not have this technology, we seem to have set the perfect snowball of economic incentives rolling to get us there.

Why do chatbots make fertile ground for developing AGI ?

Text is a much easier modality to deal with than audio, video, images or lidar as many of the problems associated with preprocessing , object detection , and the sheer data volume that are encountered immediately with richer data sources do not arise with text. Also, chatbots are meant to handle fairly well defined problems , as you are not going to ask your real estate chatbot about Nietzsche but it would be nice to chat with it about the weather in the area, the crime rate, the neighbors and other houses bought at that location. Whenever , a chatbot hits its limits and gets transferred to a human agent we are collecting valuable data about the nature of ‘Common sense’ and ‘concepts ’ that are currently outside the purview of ‘Weak AI’. This data when combined with the expense of having a human in the loop provide a great economic interest that is propelling technological development that will make chatbots become better at understanding concepts and at having ‘common sense’ when chatting. This is a perfect recipe for relentless progress from ‘Weak AI’ to ‘Strong AI’ or ‘AGI’.

What are the bigger forces at play in the field of AGI ?

Governments have increasingly realized how AI can be a tremendous advantage and there has been a concerted push at the academic level by the likes of Darpa (https://www.darpa.mil/news-events/2018-10-11) to be able to teach machines common sense. Thanks to Open AI, we have seen how even a well trained ‘Weak AI’s can be guided to generate fake narratives that seem to have been created by humans (Link) which again forces military research dollars (the kind of dollars that come from the very deep pockets of uncle Sam) to be funneled in this direction. Also, An AI with a small amount of common sense can take over a wide range of manual tasks such as taking orders or appointments over the phone , which would bring a windfall to companies like google and they are already on this case as shown here

Given where we are, we have setup the pipeline of incentives from the universities to the tech companies to be able to get to an AGI so now it is a waiting game to see when the break through will happen. At the same time, we have to remember that we can never predict when new scientific breakthroughs will happen and the first AGI breakthrough could very well be decades away or just 5 years. If it were the case that an AGI breakthrough would happen in 5 years, I think the chatbot realm seems like a fertile area for it to germinate.