Artificial Intelligence Generated Coin Collecting Applications

I evaluated a few coin-collecting applications when my collection was large enough to become unmanageable, and I was not impressed. Online versions seemed to require too much interaction, and the few desktop applications I reviewed seemed too complicated or simply did not match what I wanted in an application. Instead, I created my own Microsoft Access database with a few forms and tracked my collection. Over the next few months, the tinkerer that I am, I added forms and features, using the database’s built-in coding features for more complex tasks, such as finding and displaying a coin image saved on my computer. I manually entered dates and mint marks into a reference catalog, populating the tables with the Lincoln Cent Series and additional coins as I acquired them for my type set. I tracked invoices to coins in the database and maintained cost basis for each coin. I linked to the shipping website to track orders. This database was developed ad hoc over a few months. This database was slightly more advanced than the Excel spreadsheet method for tracking a coin collection.

I have generally been an AI pessimist but a friend started using a public-facing AI system and I started using a few of them to conduct research about coins.

Note that there were several times when these systems provided incorrect or inconsistent information, including coins that do not exist. One indicated that the 1909-D Lincon Cent was missing from my collection. There is no such coin. Another stated that the West Point Bullion Depository had minted Eisenhower Dollars without a mint mark to supplement the production of the other mints, but another stated that it had not. I have not ben able to independently confirm that this production ever happened and the United States Mint official website does not include this in its history of the West Point Mint. These systems can also be somewhat sycophantic. If you ask a question suggesting a particular answer, you are more likely to get that answer.

However, these AI systems do help point one in the right direction for further research, or can help check research one has already conducted. For example, I had one system check the references and summaries in the legislative timeline found on this site against the text of the statutes in the links provided – it spotted a few inconsistent citations, made a few suggestions on the summaries, and also made a few suggestions of statutes to add. I view these systems as assistants to help me with my research, but do not rely on them for accuracy.

I was a software developer by trade for a time and have a decent understanding of software architecture and familiarity with development tools. After using these AI systems for a few months, I decided to use one to help develop an application to replace the Microsoft Access database forms.

Note that I am not advocating the use of these systems for commercial software, as I am not qualified to opine on the stability, viability, security, and licensing requirements of AI-generated software. Some iterations of the application would crash. AI-generated software can contain bugs just as human-generated software.

However, just as assembler language coding was supplanted by higher-order programming languages, which were then replaced with object-oriented languages, which were then integrated into sophiticated programming environments, AI-generated software can be viewed as a tool with knowledge of design patterns, access to open source code (hence the licensing question for commercial software), and the ability to scan developer sites for known issues, and used to support the rapid development of personal targeted software applications. Because I already had a functional database with reference information and representative data, implementing a desktop application to replaced the Microsoft Access database took only a few days, and the results were surprising. The image above is the coin details page of the generated application.

To understand how this development process works, in one iteration of the development process, I provided the following instructions to add a second panel to a window I use to track research topics:

I would like to revise the ReportSection table to include fields for History and Quicknotes, making it consistent with the ReportSectionItem elements in the same tree. Then I would like the tree view window of report sections to be two-paneled, with the tree on the left and the values for the selected node on the right (title, description, history), so I can quickly review by navigating down the tree. They do not need to be editable for now, but this may change – more thought has to be given to the behaviors if I change and move away – auto save, save confirmation, and the like. One exception might be Quicknotes, as it was intended to jot something down quickly for follow-up. Depending on the complexity level, we could make that field be saved automatically. If we do, we’d need to remember its value on entry and exit, and update the database only when it changes. For the generated reports, the history should now be used instead of the description.

Five minutes later, the changes were made and ready for testing. Note that this was the change that first caused the application to crash when exiting the window, which was resolved after providing the system with the exception’s call stack. The resulting window is depicted here. In another iteration, I noted that the location window, which tracks coin locations for inventory purposes, displayed the pages of a certified coin binder out of order. The AI systems response was for me to check the data stored for the specific container profile in the database which I have already done. When I pushed back on that suggestion, the system reviewed the code and acknowledged that the software was not even sorting the data returned from the database query. It seems the general attitude that a software bug always resides in someone else code exists even within AI systems. There is an old adage in the software development industry – garbage in, garbage out. Precise and well-thought-out instructions will produce better results.

I have now looked for other capabilities to add that might make my collecting activities easier, and more importantly, more enjoyable. In the immediate future, the Professional Coin Grading System (PCGS) provides an interface to Coin Facts, and a capability can readily be added to the application to look up a coin by certification number and display additional information, such as its population report and true-view images, if available. Further in the future, estimating a coin’s condition using these AI systems is in its infancy, wholly dependent on the quality of the images provided, and unreliable. I do, however, imagine the day when this capability is added to the application.

Another nascent application I had created allows me to build a wish list of coins and search eBay for active listings. After the process of registering for an access key to eBay’s public programming interface, this application took about two hours to complete.

As I stated in one of my earlier posts,

Coin collecting is more than collecting coins. Except for perhaps the wealthiest among us, there is a limit to the coins that can be acquired. A collector of twenty-first-century uncertified circulated coins has the luxury of searching coin rolls indefinitely. Otherwise, there will come a day for the rest of us when coin collecting must be more than collecting coins. On that day, other activities must take the center stage. 

This is one of those activities.