In 2018, I trained a recurrent neural network—a type of artificial intelligence/machine learning (AI/ML)—on all the plots of Star Trek (TNG, DS9, and VOY). The proposed goal was to have the network create new and exciting plots based off the characters, places, and themes of the original episodes. After training for 24 hours, I asked the neural network to generate new text based on what it had “learned.” As you can read in my previous post, the output was more or less Star-Trek-Themed nonsense. While the network output words in an order that approximated sentences, there was no sense of plot or structure in the output. However, at the time this was indicative of the cutting edge AI/ML technologies available. And in a world without new Star Trek it was an entertaining nostalgic diversion.
Since that prior post two things have happened. First, in 2023 the media-consuming public are in the midst of something like a Star Trek renaissance with several series—Discovery, Picard, Prodigy, Below Decks, and Strange New Words—all bringing new ideas to the Star Trek canon. Second, more recently and more crucially, ChatGPT has been released by OpenAI. This new tool makes generating text using AI significantly easier and more intuitive.
My original recurrent neural network experiment was extremely cumbersome. It required 24 hours of training on my personal laptop. It required Python code to extract the plot synopses, train the model, and request the output. After all that the model produced output letter-by-letter that was not usable for the desired purpose. By contrast, the ChatGPT tool is both more useful and more usable. ChatGPT is a generalized Large Language Model that the OpenAI team has pretrained on enormous servers, giving it far more training than my laptop ever could. It can generate responses to complex prompts that incorporate prior context and knowledge from the internet. On top of that, it can be accessed by anyone for free via an easy to use chat-like interface instead of complex Python code.
To demonstrate how much better the ChatGPT output is, here is one sample synopsis that ChatGPT produced for me when given the prompt, “Write a 100 word plot synopsis for an episode of the television show Star Trek Deep Space 9 that has never existed.”
“Title: Shadows of the Past
In the unexplored Gamma Quadrant, the crew of Deep Space 9 investigates a planet trapped in a repeating temporal loop. They discover the Zynthians, a technologically advanced species, who made a pact with a malevolent entity, leading to their downfall. The crew must break the cycle and free the Zynthians, facing temporal distortions and moral dilemmas. Unity and forgiveness become the key to redemption as they race against time. This thought-provoking episode explores the consequences of choices and the power of hope, reminding us that even in darkness, understanding can pave the way for a brighter future.”
This response demonstrates some key improvements over the text that my recurrent neural network generated. The plot follows conventional story structure and includes the same alien species throughout (the Zythians). It generated a relevant title based on the plot. ChatGPT knew that Star Trek: DS9 takes place in the fictional Gamma Quadrant and a longer version of this synopsis included the names of the main characters of the show. Recognizing that Star Trek episodes usually have a moral or philosophical theme it emphasized that this episode focuses on “the consequences of choices and the power of hope.” Overall, this is a monumental improvement over my previous attempts and demonstrates that utility and usability of ChatGPT in generating text.
In 2018, leading experts in the field were effusive in their praise for recurrent neural networks and how great they were. Now, five years later, the AI community has taken a dramatic leap forward. Despite these improvements, the output is far from perfect. Longer versions of the response had conflicting themes and ChatGPT's output are necessarily derivative of existing Star Trek works. At the time of writing, there is an ongoing writer's strike. It is important to recognize that while AI has improved dramatically, it cannot replace real writers. It remains a tool, that is useful but not creative or emotive. However, given the strides in development, it is not unreasonable to think that the next update to this blog post in 2028 will feature AI-generated plots, scripts, audio, and video all at the push of a button. Whether those are any good, remains to be seen. But, at that point it may feel more like we are living in the Star Trek universe rather than writing fiction about it.