NUMI: HOW THE AI SAUSAGE IS MADE

Creation Effort: It took about 10 hours to develop an alpha build. The key task was fine-tuning the AI with detailed custom instructions [all written in English] and background documents [also in English] about coin grading, particularly the Sheldon scale.

Example Challenge and Solution: Initially, Numi struggled with identifying 's' mint marks on US coins. I updated the instructions to give it context about proof coins, and now it can better identify proof coins and give a PR grade rather than an MS grade.

Interface and Back-end: The app uses OpenAI's existing chat interface. There was no work done, or needed, to update the front-end user interface. The back-end customization involved only instructing the AI in natural language, without coding. It is possible to link the app to 3rd party APIs so the AI can interact with outside websites. One feature I will build is eBay integration so that users can pull up recently sold listing and see how much their coins [with the estimated grade] is worth.

Field Testing: I went to my local coin shop yesterday and tested the app with the shop owner. Results were mixed, with some estimates being in the ballpark while other estimates were off by as much as 20 points. I'm working on improving this by experimenting with lighting conditions and additional instructions.

The AI is not consistent at the moment. I am testing out various lighting conditions, adding new custom instructions, etc. I'm excited when OpenAI allows videos to be uploaded instead of just photos. Videos have a lot more info and may be the key to successful grading estimates.

Ability to follow up on conversations
One of the best things about Numi is it does much more than identify and grade coins. Users can ask follow-up questions to learn more about their coins. Users can follow up with questions such as:

Another collector asked me how I built Numi and what I think is next. Below was my response:

"I built Numi using only natural language. No coding was necessary and no photos or training were involved. My value add was creating highly customized instructions to try to get a grade corresponding with the top TPG's grading standards. The base core model is GPT4, which everyone has access to.

"My bet is that by the end of this year, the concept of AI Grading will start to be taken seriously by the hobby. By the end of next year, especially if GPT5 is released, AI Grading will be accepted as a valid way to test coins before TPG submission."

I'd appreciate it if readers could send photos of their slabbed coins (8-10 photos, 4-5 per side, with the grade out of view for testing. Once I run the analysis, I'll report back to the reader and then they can give me the grade of the coin [if they know].

I would also love to get feedback from advanced/expert coin graders on Numi's current grading criteria. You can see the exact criteria in this Google word doc and leave comments there.

Readers can contact me at justintsn10@gmail.com. I am also actively updating the app and posting results on my Instagram page for those who would like to follow my progress.

Impressive. It's not the first attempt to automate these tasks, and it won't be the last. We'll look forward to further reports as the technology develops.

In the past I've remarked to people that when I started The E-Sylum I wasn't writing a newsletter, I was building a website - and now our archive holds over 33,000 numismatic articles and 55,000 images. When I began my Coin Library website (before joining forces with the Newman Numismatic Portal) I told some people that I wasn't building a website, I was building a database of numismatic images - with the goal of someday training a computer program to recognize numismatic items visually.

NNP moved more in the direction of digitizing text, but it does have a lot of images. I remember a year or two ago when Len Augsburger reported that bots were driving a ton of traffic to the site - we figured it was just companies offering print-on-demand versions of out of copyright works. In the back of my mind though, I remember thinking, "...or it's artificial intelligence becoming sentient."

I didn't say that part out loud for fear of being thought of as a tinfoil-hat weirdo, and the "becoming sentient" part was a joke (we're still a long way from that). But I wouldn't be surprised if AI companies the likes of OpenAI were sucking up massive quantities of data to train their models. And all the coin images and descriptions, not just on NNP but everywhere, from every auction and dealer site in the world were a part of that. That's an immense amount of information that no one human could fully comprehend, let alone read in an entire lifetime. This could get interesting very quickly. -Editor

For more information on GPTs, see:
Introducing GPTs (https://openai.com/blog/introducing-gpts)