I drew a connected graph of 500 film directors based on the replies of ChatGPT. In every prompt, I asked the chatbot to recommend me five similar film directors for a certain name. I started with some well known directors but the graph got connected after some prompts. I imported the results to a graph application called Graph Commons for visualization. I was aiming to inject more data, but the free subscription of Graph Commons only accepts 500 nodes. It’s also nice, at least gave me a closure. I’ll take a look at the results, hopefully in the coming days. If you want to visit the graph and play with it, here’s the link.
Method
In all my prompts, I asked ChatGPT to give me 5 film directors similar to the one I give with the following prompt:
Forget everything we talked about. List top 5 film directors similar to Lav Diaz. Don’t add explanations.
The first sentence was just an attempt to avoid drawing circles based on the earlier responses but I found out that it probably doesn’t have any actual impact. I gathered the responses in a Google Sheet and imported back to Graph Commons.
I selected the film director names randomly but I tried to widen the graph to make it more diverse. My approach was not a systematic one but I tried to give names that are located on the child nodes to start something new or locate these names better in the graph. A practical example: I didn’t query all the Hollywood action movie directors to avoid discovering the outskirts of this genre. Instead, I focused on Japanese or Serbian directors since I’m also more interested in them, especially for discovering new films. But this leads me to the…
Limitations
Disclaimer: please take this graph as a joke or as a delirium since none of the nodes or the edges have any kind of justification. It’s just a dream of an AI chatbot that I intervened with my dreams.
That said, here are some limitations on top of my head:
- There is no clear ending point for this graph. I just stopped at 500 since the tool I use didn’t let me to add more.
- I started from and continued at every step with my unjustified subjective prompts. I asked the names that I know or want to learn. I attempted at positive discrimination at times. At any point, if someone else asks a different question, then the graph would be pretty different. (Just curious, how different would it have been?). Some of the missing directors include John Waters, Shōhei Imamura, Věra Chytilová, and Giuseppe Tornatore.
- As you might have heard, ChatGPT is also a hallucinative liar. With its great rhetorical baggage, it keeps telling lies. When it doesn’t have enough info about a certain director, it just gives the name I prompted in the results. It also returns some author, actor or non-existent names time to time. I tried to fix these when I noticed, but I’m sure some of them leaked to the final graph. One example that I know of is Isabelle Huppert who is an actress, but I couldn’t remove her from the graph. Because. She’s Isabelle Huppert.
- At first, I did some experiments like giving the same prompt for a certain director multiple times. The results share some commonalities, but it also feels pretty random. Many times, some unrelated name popped up. That’s why I tried to give a prompt for each name only once and tried not to repeat. So these are the initial thoughts of the bot. Andrej Karpathy’s walkthrough on building a proof-of-concept GPT helped me a lot to understand the probabilistic responses of the ChatGPT outputs.
- There’s also a limitation of the time period. The data it was trained ends at 2021. For the record, the one I used was “ChatGPT Dec 15 Version” (2022). I also thought that the near past data (last 5 years) is not as good as the earlier times. But how would I know?
- ChatGPT have a lot of biases based on the input it processed. It’s clear that it reflects those. The non-American or non-European directors have hard time to connect to the main spheres. There are only a couple of junctions, the nation-based similarity dominates the graph. Same applies to women directors. ChatGPT just match women with women most of the time.
Motivation
Why did I do this stupid thing? I’ve been thinking about it while I was writing prompts or copying the replies to a spreadsheet. For a few days, I was fully focused on this but I was also aware that it means nothing. I still don’t know, but I wanted to do it, enjoyed it, and also learned about many directors and genres I didn’t know before. I feel that we’ll talk about the subconscious of AI in the short term. Just like we discover artists, authors etc. some people will be interested in AI-generated content or LLM Cultural Studies, that’s my intuition for now.
Top 5
Based on the centrality of the nodes, the ones who have the most connections are as follows:
- Martin Scorsese | 27
- Jean-Luc Godard | 25
- Wong Kar-wai | 19
- Andrei Tarkovsky, Federico Fellini, Alain Resnais | 17
- Agnès Varda | 15
Agnès Varda and Wong Kar-wai are nice surprises.
Tight Junctions
This one felt like a bug at first but again maybe there’s some truth to it. Some prompts got circular responses from ChatGPT where it was finding similarities between 3 to 5 names, always mentioning those when I asked a connection. Here are some closely related directors according to ChatGPT:
