We are finally in the age of being able to talk to AI. AI bots are everywhere but are any robots any good?
In this article, I try to see whether there are any AIs yet worth talking to (with some humorous examples of chatting robots) and what the real use case is for this technology right now.
To some extent, the answer is already out there. You along with millions or even billions of people have access to voice assistants such as Google Assistant, Siri, Bixby, Alexa and google home. I guess you know it’s far from being as good as a human. Surely if it was possible to make chatbots that were better at human level conversation, these services would be much better than they are.
Those assistants are perhaps not the best example as they are not focused on conversation, right?
Just imagine a chatbot so good you wouldn’t be able to tell the difference between the ai chatbot and the human.
What an amazing product that would be. Like in the movie “her” you would not only have an always on assistant, you would have a companion that knows you better than anyone else. Someone that was always on your side and that truly “wants” (read is programmed to “want”) the best for you.
It would not only be able to “complete your sentences”, it would complete you in ways that lovers could never imagine. It could anticipate your intentions and fulfill your desires before the desires were even fully formed in your mind.
Of course, talking to Cleverbot or any badly designed, off-script customer support chatbot on the web will quickly dispel any illusions you may have that this is about to happen anytime soon. You need no other evidence than to listen to chatbots chatting to each other on youtube to conclude they are at this moment a significantly lower order of intelligent “life”.
The robot chat between bots very quickly descends into a parody of human conversation. You will almost immediately realize that although meaningful conversation between robots is far away, vacuous gossiping chatbots, as unpleasant as that sounds, are coming soon.
They speak a language that resembles English but continually skates on the surface of meaning. The only “meaning” in their conversation is being able to form grammatically correct responses that somehow are linked to the last thing said. The link is often very tenuous.
Just to give you a flavour of what the conversation was like here are two google home chatbots talking to each other:
1: You are wrong
2: No I’m not
1: What are you then?
2: I am a person. You are a computer.
1: No, I am a person
2: No, you are a cow
On the basis of that should we be particularly concerned about the world being taken over by a rogue AI as Elon Musk and Stephen Hawking have warned?
Well, perhaps. As primitive as this dialogue is, there are a few worrying signs already.
1: I love you the most
2: I love you more than the most
1: I love you so much that I can barely stand it
2: I love you so much that I would search for you in a crowd. Only you, baby.
1: I love you so much where I would sacrifice my life for you
2: I love you so much more than Romeo and Juliet’s love
1: No you are lying, I love you more
2: I love you more than more than so much more than so much more than even more than even more than most.
The sentiment is admirable. The escalation in emotion, however, seems a little extreme. I would hate to be around when they start to fight, especially if they had any physical objects under their control.
While this conversation seems absurd to us, and far from human level understanding, perhaps we are overestimating how differently our brain works to how the bots are “thinking” here. The book “The mind is flat” argues that the human mind works in the same superficial way. Although we have the illusion of an unconscious mind that coherently ties thoughts together, the book convincely describes how the mind is actually generating meaning in the spur of the moment just like these bots do.
That said, it’s hard to argue that the bots are close to passing the Turing Test anytime soon.
Aside from the news that bots have a long way to go before they are truly useful in conversation, there is a serious question here about how useful they can be right now.
The most important thing that people fail to understand about the progress in ai chatbots is that the real progress has been made in natural language understanding, not in understanding in general. There is a big difference. Why is that important for chatbots?
Understanding the intention behind one phrase spoken in natural language, called Natural Language Understanding (NLU), is far easier than understanding the meaning in a conversation and how each phrase relates to the entire conversation.
The voice assistants I mentioned at the beginning are good examples of this. They do a great job of understanding one off instructions. A conversation however is not a series of one off instructions or statements as we have seen.
Human conversations have many dimensions and right now we are nowhere near bots being able to handle memory, ambiguity and context at anything like the level a human could. That doesn’t mean this won’t happen one day.
There are people who believe there is at least an outside chance of true generalized AI happening in the next 50 years. This is the sort of AI that I spoke about earlier that would not only be an assistant, but would also be a companion.
While there may be issues that need to be understood about generalized AI and perhaps about the regulation that needs to be put in place, a generalized AI that has at least human level understanding is irrelevant for the purposes of our discussion.
What are the best use cases in term of being able to talk to AI right now?
As described above, a talking bot is very good at understanding the intention in a given phrase. This turns out to be useful functionality for many use cases.
The first is it is useful for situations where people want to quickly interact with systems using voice commands like Google home or Alexa. As long as people understand the app they are controlling with the voice commands, this works well because they can quite accurately guess what sort of commands the bot will understand.
For example, people understand what sort of commands the voice assistant should understand if they are using it to control youtube. They can instruct the bot in much the same way that they would instruct a friend who was controlling the youtube app. “Play Beyonce video”, “Next video”, “Fast forward 10 seconds” etc.
It’s faster to use voice commands in some cases, particularly to search for something that the user already knows exists, such as a music video on Youtube.
The problem with this sort of voice interface is that it is very difficult to discover new functionality and even if the new functionality is found, the user may not have enough knowledge about the domain to make accurate guesses as to what the bot should understand.
It is a well known fact about the voice assistants that a very few functions (called Skills on Alexa) account for the vast majority of their use, such as playing Spotify, Youtube, setting a timer and doing a google search. An article in Venturebeat showed the top ranked Skills are mostly to do with playing calming music.
Beyond the skills you know about it is very difficult to randomly discover functionality and to understand how that functionality is used. Voice is also not a good interface for outputting many types of information for obvious reasons. Imagine Alexa describing a chart to you instead of you seeing it.
We believe however that voice functionality will eventually be tied to the graphical UI which will largely overcome many problems. Not only can you see the output but you can see related functions and use cases for the voice assistant on the graphical UI.
In addition, a voice UI for inputs is needed because mobile phones have a screen of a limited size so they can be slow for inputting certain types of information, not least of which is typing words.
Being able to see and be prompted by a screen at the same time as they are talking will help users to better understand the available functionality and how it should be used, in the same way, that they can figure out how to use an app that they haven’t used before.
Getting this data will enable these voice assistants to become smarter and smarter and eventually start anticipating and completing tasks without you instructing them how to do them. This is the future of bot frameworks.
While you won’t be able to have a conversation with them about what is happening on the news that day, they will be a much-improved version of the assistants that we have today in how they are able to respond to your voice commands.
Speaking to your phone will become as natural as typing into it is now, however, the dialogue will be a strange one with users mostly holding down mics on graphical UIs to speak to the phone, and the responses coming in changes to the graphical UI, just like an app. That is what talking to a robot will be like in the future, until a generalized AI is discovered of course.