"Computers are not as intelligent as you think": the effort to tell AI stories

Computer scientist Dr. Lara Martin tells us about her research to tell stories and play dungeons & dragons in artificial intelligence.



Dr. Lara Martin needs to learn how to tell a story and how to speak it well.


Lara is a post-doctoral fellow at the University of Pennsylvania for computing innovation where she teaches ID to create natural and human stories and language.


She shows why we must train machines how to tell a story and what Dungeons & Dragons have to do with everything.


Why are we going to tell stories by machines?


Since before we were able to speak, we have been telling stories. If computers could also say and understand stories, we could interact more naturally with them.


We start to adopt personal assistants in conversations as a business – such as Alexa or Siri – but these machines are still unable to speak to anyone. The best and most effective way for people to talk is to tell stories.


So teaching an IA to told our life and technology could improve?


Many people don't know how much we say is a story or how much we can be structured like a story. I like to imagine you can just speak with your personal help and find something with you.


You could plan a birthday party for your child and you say "Hey, for Gina's 10th birthday I'm planning a Party. Can you assist me?" And this party will build a story: "Any good party begins with cake. In the local grocery shop, you could get a cake and then buy a ballon while you're there. After the decorations have been put up..." And so forth. And so on.


This party story could collaborate with you until you are satisfied with it. I believe here there is much cool possibility to cooperate with human-AI.


Where are you beginning? And how do you have to create the layers to teach an IA about the story?


There are a few ways to begin. The majority of modern methods are based on a lot of stories. You pick or find a bunch of stories and run through an algorithm that saves patterns in the stories, like battling the dragon, usually until, for example, you save the princess. You—the human person—will come up with the first phase of the tale and spit out the rest.


These systems generate brand new, grammatic English phrases which are very sweet, but they just wander around and forget about it a little while later.


The previous methods – which are still being used by some people – need a lot more effort. These investigators sit down and find all the potential plot points in a world of stories and how they are linked. In order to construct the narrative, the machine must plan a way to take these plot points.


The language is not as focused on itself as it can only show a sentence or two if the machine selects the story. They concentrate on the cause and effect of the story's plot: 'Veena has to get the sword, for instance, before she can kill the dragon.' Notice that this differs subtly from current approaches.


My doctoral thesis focused on integrating different concepts from the two approaches. I will take the text of the latest techniques and use certain rules and restrictions, as in the older ones, to ensure that the next sentence of the story is actually possible.


What type of story computers able to tell?


Older methods produce real coherent and informative stories, but only with the handmade stuff that you can tell rich stories in a single-story universe, can you tell stories. One of my favorite examples is the Façade game developed at Santa Cruz, University of California. You can also think of such stories as playing a gigantic open-world game - like Skyrim or The Witcher - that creates and maintains all the branches of history rather than human ones, using AI.



The newer approaches are very interesting, but they lose coherence quickly as they are so difficult to manage. A strong and well-known example is the AI Dungeon series, where the story begins and you take turns like an old text adventure game.


However, these are both interactive story-generations. Instead of themselves, they are sharing the story with the human, which is very different from just a machine telling the story, but they give you a clear idea of the kinds of stories AI can tell.


You worked as a post-grade ID, how could play the Dungeons & Dragons role-playing game, how does that fit in?


My work focused mostly on making the AI tell consistent stories. Having an AI play Dungeons & Dragons is more a goal, something that I would like to see the group come together to deal with – and what I have chipped on.


The computer must be able to play the Dungeon master, there is so much. [The Dungeon Master for Dungeons & Dragons manages the game world and creates information about the story] for other players. The AI can understand the story it tells but also the pieces of the story the other players say.


It must ensure that everyone else knows the story it tells. They must be able to build stories that are inherently worthwhile so that players can enjoy a story. Instead of concentrating on external benefits like collecting experiences by killing a lot of goblins, this may involve developing interesting characters.


These are just a few examples of the hard issues that have not yet been resolved in playing dungeons & dragons, it's incredible that people can do it! Even a coherent book chapter can't be generated by an AI yet. Language is essentially the true beast we need to better understand.


When are we going to get an AI master dungeon?


Seeing anything like a human Dungeon Master? I don't know in life – I'm somewhat sceptical it's going to happen. We made pretty fast progress, but it's really hard.


Will an AI ever be able to create a plot, or is it ever capable of using only the ideas that you gave it?


If an AI agent creates something new and interesting, and it answers questions, well, was the work of the AI, or was it the work of the creator of the researcher who made it? There are also legal issues here.


You put your imprint on any AI agent you create, whether you like it or not. The more you go to the rule-based side, the more you end up with this agent. So it's a tough question to ask whether an AI can come up with something imaginative alone.


Computer imagination is very interesting because we just don't know how to answer certain philosophical questions. And I am no philosopher. And I am not a philosopher.


I think we're some way away from winning the best screenplay AI with the Oscar?


While I think it would be awesome to see innovative AIs made, as described earlier, having an AI won by Oscars just has too many problems.


The machine is not live, and I believe it is necessary for people to realize that computers are not as intelligent as they think they are. They are not people, they have no agency. they're not people. It is just tools used by other people to work on these items. I, therefore, believe that the best way to use creative AI is to increase human creativity.


Computers look very well at vast quantities of data to create stuff you have never before seen, never thought of as related. However, people are really good at taking and working with the ideas that the computer may convey to them.


I told one of my earliest systems to write a story about the next sentence. It came up with random, strange things most of the time that didn't work. But then came the idea that a horse would become an entrepreneur in the lawn chair.


The computer does not know what it means, it only spits things out. But you could take it and run along, perhaps they will go and make this horse a story – that's fantastic.


People have this capability, and I believe that this is a good symbiotic relationship that must be used more. These are the things AI has to do with.





To help their work, Newsmusk allows writers to use primary sources. White papers, government data, initial reporting, and interviews with industry experts are only a few examples. Where relevant, we also cite original research from other respected publishers.



Source- Science Focus


13 views0 comments