190-8 Developing a Domain-Specific Large Language Model (LLM) for the Spanish Peaks in South-central Colorado

Abstract:

Traditionally, the path for a MS and PhD dissertation is to publish a few articles.  Some make it to the internet; most don't.  The author moves on and the papers are forgotten except by a few specialists in the field.  

In 1994, the website, The Igneous Petrology of the Spanish Peaks, came online during the infancy of the internet. Based on today's standards, it was crudely constructed using HTML on a Silicon Graphics workstation.  At the time, it was unusual because it presented all the information from my eponymously named PhD dissertation which included sample locations, geochemistry, etc.  The website is a repository for the data collected from the dissertation.   

Technology has evolved.  While a portion of the original dissertation involved using Neural Networks to map dikes in satellite imagery, nothing prognosticated the direction machine learning would take or the role it would play today.  Currently, we are in the midst of a revolution of Large Language Models (LLMs).  The first of which was ChatGPT from OpenAI. Now there are hundreds of such models.  

These models are quite dynamic, providing all sorts of information.  Some of it is purely made up because of the completion problem, i.e., very few models respond with IDK, so they just make stuff up. Every query has to answered. Part of the problem is that when an LLM is queried, it has to go out and collect information.  It starts from zero, every time.  Several people have recognized this limitation, one of the first was Andrej Karpathy.  He suggested something different.  What if an LLM created a wiki previously discovered and/or synthesized that could be referred to later instead of recreating that information?

Taking this idea, a Spanish Peaks wiki was created.  All the information about the Spanish Peaks was fed into CoPilot (another LLM) including the dissertation, abstracts, published articles, excel files, etc.  These data were organized as a Karpathy Wiki.  Since there are other models available, e.g., GeminiLM, and Perplexity, they were fed the same information.  The amalgamated information is available for interactive queries from each of these models by the author.  In this way, the information on the Spanish Peaks is preserved in context.  This can provide an easily accessible repository for one's entire knowledge on a topic.   

Geological Society of America Abstracts with Programs. Vol. 58, No. 5, 2026

doi: 10.1130/abs/2026AM-18887

© Copyright 2026 The Geological Society of America (GSA), all rights reserved.