Large language models (LLMs) or as they are colloquially known, AI, have made impressive progress since I last explored possible use cases relating to model railroading last summer, in the form of both chat-based programs and image generation.
With a year having passed between then and now, I thought it would be worth revisiting this technology to see if the aforementioned progress which has been made in the field of LLMs applies to their understanding of model railroading, and the process of their construction.
A few things to establish from the outset: In the past installment, I used OpenAI’s ChatGPT, which was at the time the gold standard in text-based LLM programs. However, since then, multiple companies have released equally impressive competing products. For the purposes of this article, I’ve used Google’s Gemini.
To start, I began by asking simply “Help me to plan a model railroad.”
Already I notice better response formatting than ChatGPT offered a year ago, with a clearly delineated and bulleted response listing the many considerations a modeler must make as they begin to plan their layout. The response continued:
As Model Railroader‘s Cody Grivno pointed out, there are some inaccuracies in the above statement. I’ll let Cody speak for himself: “HO scale is not half-inch to the foot. That’s 1/24 scale, which falls under the large scale umbrella. N scale is not one-quarter inch to the foot. That’s O scale.”
Another difference between ChatGPT and Gemini: Gemini is asking questions in response within the secondary bullet points of each numbered consideration. Admittedly, this could be down to the prompt I began with. As mentioned before, I asked Gemini to “help me to plan a model railroad” while a year ago I had asked ChatGPT to “generate me a model railroad trackplan layout.” The former prompt is clearly superior to the latter, as my understanding of both chat-based LLMs and the hobby of model railroading has grown in the past year. While both prompts express basically the same thought, the former is more open-ended, while the latter includes unnecessary verbiage (“help”) and redundant descriptors (“model railroad trackplan layout.”)
However, a clear difference emerges between the two regardless of the quality of my prompt: Gemini is asking questions in response, guiding the user, while ChatGPT required follow-up questions on my part, presupposing a baseline of knowledge of railroads in both prototype and scale.
Hey! Trains.com included as a reference! I know those guys!
Having read through the questions asked of me in response, I decided that I wanted my model railroad layout to be HO scale. I tried to keep my response more or less limited to the information which was provided to me by Gemini in response to my first prompt. Including that information, I then asked “…what railroad should I use for inspiration?”
Gemini responded with three primary considerations to keep in mind: Era, location, and railroad type.
From there, Gemini went on to list railroads to consider broken down by era, starting with the steam era (Union Pacific, Baltimore & Ohio, and Southern Pacific) and followed by the diesel era (Atchison, Topeka and Santa Fe Railway; Amtrak; and Canadian National). Also, narrow gauge railroads (Denver & Rio Grande Western, Durango & Silverton) are listed for consideration.
I decided that for the purposes of this hypothetical layout, I had an interest in the American Northeast during the steam-to-diesel transition era. In response to that decision, Gemini offered era and location appropriate railroads alongside with a Pro-Con breakdown for each railroad.
From this list of railroads, I decided I would proceed with Maine Central as my prototype. I then asked Gemini “How should I get started?”
The response I received listed steps to take before construction could begin. The list was as follows:
- Research the MEC
- Define Your Layout Scope
- Choose a Specific Era
- Design Your Track Plan
- Gather Supplies
Additional resources were again listed, which included Model Railroader and an existing Maine Central layout that could serve as inspiration.
Finally, I wanted to see first-hand the progress made by image generation programs since last writing. I used a different program than last time, this time opting to utilize Microsoft Copilot’s Bing Image Creator. To begin, I also used the same prompt: “model railroad layout, birds eye view, HO scale, mountainous, with tunnel, mining operations and city.” I chose to reuse this prompt as a control, a means of comparing the results while knowing the input remained the same. The result was a marked improvement in both clarity and quality. Locomotives were now more clearly distinguished from buildings, and lines were better defined.
Then, I changed the prompt to better fit the Maine Central prototype theorized above. The prompt: “model railroad set in Maine with green and yellow locomotives, photorealistic.” The results were impressive.
The locomotives, while not an exact replication of actual Maine Central paint schemes, was an admirable approximation.
The strides made in the field of LLMs in the past year is impressive. It now offers real potential value to model railroaders, particularly to those searching for inspiration or resources for a new layout, or those modelers beginning their first foray into the hobby.