One of the hottest topics of the moment is the conversation around artificial intelligence (AI.) If you’ve been paying attention to this conversation, you likely know that Artificial Intelligence programs have made enormous strides forward in just the past few years. These advances have caused disruption and innovation in equal measure in many fields. These advances prompted discussion in the Model Railroader offices, our primary question being, naturally, what can these programs do for us, and for our hobby?
Curious as to the capabilities of these programs as they relate to model railroading, I started to experiment. Using a program known as ChatGPT, a web-based AI chatbot, I began to test the knowledge and capacity of this program. By asking it questions, ChatGPT generates responses based on the specified parameters of the question – more open-ended questions yield more open-ended responses, and more specifically-defined questions result in more specific answers. The program uses a generative learning model, which pulls from a broad spectrum of resources across the web to generate responses without human input.
Getting Started
To start, I asked an open-ended question to see in what direction the program would take me. My first query: “Generate me a model railroad trackplan layout.”
The response I received:
Let’s start with the basics: a 4×8 layout with a rural countryside theme. Okay, maybe no points for originality, but it’s certainly a workable suggestion. From there, the program specifies what goes where on the layout. In the bottom left corner, we have a small village with a train station. Our mainline curves around the outside of the village and crosses a bridge over a river before entering a tunnel. I can easily picture what is being described in my mind, so it’s plausible.
From there, the track splits, with one path diverging into a mountainous region with a mining operation while the other continues to an urban area with a passenger station and a switching yard.
Well, okay. Seems a bit dense for a 4×8 layout to me, but then again, scale was never specified, which makes a difference. Without that having been specified, how was I to know what scale to use on this layout? So, I asked:
A 4×8 HO scale layout with a rural countryside theme. Again, perhaps no points for originality, but some of the details are well-rendered, and I have no problem picturing this layout in my mind. From there, I posed more questions about the layout, including era, benchwork, and what motive power would be appropriate for a layout matching this description.
The suggested locomotives are certainly viable recommendations. There are of course plenty of options available in the model railroading market for late-period steam locomotives, SD40s, and other such locomotives.
Narrowing results
Wanting to see what would change when I specified the scale in my initial inquiry, I entered almost exactly the same question, but this time including my desired scale (HO.) The result I got back:
As you can tell, the result is almost exactly identical to the first question I asked, and the response is structured in the same manner as that which preceded it.
Specifying response structure
Wanting to change things up and further test the capacity of this program, I changed how I phrased my question in a new query: “Generate me an N scale model railroad layout with detailed description including bullet points.”
The response I received:
I found myself much more impressed by this response. By structuring the parameters of the response I wanted, I seem to have more clearly defined for this program what individual details were needed without actually specifying them. The bullet point structure of this response gives clearly defined individual features, including track, lighting and operations.
Finally, to round out this inquiry, I asked “What motive power would be appropriate on this layout, considering the specifications you have provided me?”
The response was as follows:
Other articles about AI programs have pointed out their tendency to uncritically share incorrect information. I shouldn’t be surprised that the program didn’t know that an NW2 isn’t a steam locomotive.
Generating an image
I also wanted to test out the capacity of similar AI programs focused on image generation. I fed the details into a program called Fotor, and from the prompt “Model railroad layout trackplan, birds eye view, HO scale, mountainous, with tunnel, mining operation and city” I received these images:
Now, if you don’t look closely, these are interesting enough images, but hardly trackplans in the style any model railroader would recognize. Upon closer inspection, though, one might notice a locomotive and a building morphing into one entity, or nonsensical track architecture, or bridges leading into the sheer sides of cliffs, no tunnel entrance in sight. Hardly a track plan, nor a logical interpretation of a model railroad layout. From looking at these images, it would seem the program seems to know what to replicate, but not how to do so.
Considering these results, I found myself underwhelmed. That’s not to say that these are useless programs; there are applications. For example, a beginner modeler could use ChatGPT as a tool to generate inspiration and useful information for a prospective layout. However, it’s ultimately just an aggregator. This program, and others like it, pull information from thousands of different places, finds what occurs most frequently, and regurgitates that to the user without any process to verify the information. Essentially, a user is hav a conversation with a search engine. As we’ve seen, there are ways to better define and structure the response you receive, but ultimately, the information is coming from the same place. Again, this is not to say that the program is useless. But perhaps the truly revolutionary program some believe this to be is as of yet unrealized. For now, it’s a lowest common denominator machine.
Interested in learning more about potential AI use cases for model railroaders? Check out the links below! – Ed.
ChatGPT use cases for the model railroader
After a year of change in the field of LLMs, a look at the progress made and what these new developments mean for model railroaders.
Revisiting AI for model railroaders
A comparison between ChatGPT and Google’s Gemini text-based LLM models, and an examination of the image generation results from Microsoft’s Bing image generator.
Designing structure interiors with artificial intelligence
An exploration of the practical application of image generation AI programs, and a practical application of AI generated structure interiors on Model Railroader‘s Fremont Mills layout.
