I created Numi: The World's Most Powerful Coin Grading and Identification AI, powered by ChatGPT4.
AlbumNerd
Posts: 182 ✭✭✭
Ever since Compugrade tried [and failed] to use Artificial Intelligence back in 1991, coin collectors have wondered if we could use AI to accurately grade and identify coins. So I thought I would give my hand a try!
I created Numi, a fine-tuned version of ChatGPT4 specifically made to provide objective grading estimates and coin identification. [Works on both mobile and desktop] [FYI you will need to have a ChatGPT account]
You can upload up to 10 photos of a coin and the AI will run a combined analysis. Below are some info you will get back details such as:
- Short Description
- Coin Type
- Year
- Mint Mark
- Estimated Grade
What's cool is that you can continue the conversation with the AI to learn more about your coin.
- Want to know who designed the coin?
- Want to know why a certain symbol was used?
- Just follow up and ask!
Caveats
- This is an alpha build!
- I recommend you use wifi as uploading images will be faster
- OpenAI's servers are under heavy load lately, so responses may be currently slower
- The model isn't perfect, but it is quickly improving
Features Under Consideration
- eBay integration to see recent sold prices based on the estimated grade
This app is in the alpha stage. I made this app because I wanted to help the coin community out. I know a lot of people are looking for help with identifying and grading coins.
Looking for feedback and thoughts. Have fun finding ways to break it!
Latest Updates
v1.30 Alpha is now live!
Latest Update Notes Here
Video Demo
Comments
Mods, please let me know if I should take this down or edit this post! This is a personal project and I have no way of making money off this. My goal is to see if I can make something to give back to the hobby that has given me so much.
I went ahead and downloaded the ChatGPT app, but it’s only free with version 3. ChatGPT4 is $20 a month. I’m not sure if I want to pay that much, but I’ll think about it.
Mr_Spud
It's a bummer that OpenAI puts these models behind their paid subscription
But here's a video demo in case you want to see what it looks like 
I want to see what it looks like but I don't want to down load anything or open random links...
Maybe a screenshot or pic?
I wouldn't put ChatGPT on my devices even if it wasn't $20 a month.
Successful BST with ad4400, Kccoin, lablover, pointfivezero, koynekwest, jwitten, coin22lover, HalfDimeDude, erwindoc, jyzskowsi, COINS MAKE CENTS, AlanSki, BryceM
Perhaps this can help
Title
Part 1 Response
Part 2 Response
For grins
Have it grade this based on photos
@johnny010
Here you go!
$5 Gold Liberty Head - 1880
Here are also other coins I've ran through
Jefferson Nickel - 1980 S Proof
Lincoln Wheat Cent - 1948 D
Mexico One Peso - 1971
Odd response.
You'll be using it sooner than you know.
@AlbumNerd
Thanks
I’ll give you two more coins to evaluate then I’ll post the grades for all three for transparency. See below
—
@johnny010
Thanks for the additional examples!
$5 Gold Indian Head Half Eagle - 1914
AI Grade Given: XF 40
Full AI Response
$1 Gold Indian Princess Head - 1862
AI Grade Given: AU 55
Full AI Response
As promised.
Thanks @johnny010!
The AI was shockingly off. I wonder if it would have helped if more photos were included. If you don't mind, could you take and send photos of the obverse and reverse of a coin at various angles?
I suspect the AI would do better with more info/angels.
I’m in air right now. The images attached were professionally taken by @messydesk
I’ll see if I can grab some quick photos when I get home here in an hour, but I promise they will not be nearly as good as what was provided above.
No rush! The photos you provided so far are amazingly high quality. But I suspect a big help for the AI is getting various angles with different lighting.
I set up the app so that it can ingest up to 10 photos and interpret them all holistically. It'll be interesting to test and see what the AI's new response will be.
Pedantry insists that I mention that Compugrade did not use AI back in 1991. They used classical ("linear") algorithms and formulas to generate a scores that would lead to a grade. There was no network learning anything, rather the algorithms were programmed.
Also, the data acquisition part of the system was tightly bound to the grading algorithms. There was no other image input available and the images were reproducible across grading events. Attempting to do AI grading without first being able to control the input image attributes will only lead to a lot of frustration.
Keeper of the VAM Catalog • Professional Coin Imaging • Prime Number Set • World Coins in Early America • British Trade Dollars • Variety Attribution
Want to give this one a try @AlbumNerd? Should be a nice variation of lighting and photos.
Founder- Peak Rarities
Website
Instagram
Facebook
Damn….I grade that $3 as sensational!!!
MY GOLD TYPE SET https://pcgs.com/setregistry/type-sets/complete-type-sets/gold-type-set-12-piece-circulation-strikes-1839-1933/publishedset/321940
No I won't be.
Successful BST with ad4400, Kccoin, lablover, pointfivezero, koynekwest, jwitten, coin22lover, HalfDimeDude, erwindoc, jyzskowsi, COINS MAKE CENTS, AlanSki, BryceM
You will if you use a search engine. I'm not sure what you think AI is or what it's going to be used for. But it will be incorporated in Google. It's already in Bing search. Apple is working on it.
Then you've got your digital assistants like Seri or Alexa. ChatGPT is simply a more advanced version.
Or is it just ChatGPT that you are boycotting?
I know you're not Amish. You mentioned your devices. 😀
It may well be interpreting luster breaks as wear.
Not to be pessimistic but while this sort of technology is possible today, the preparation of the input dataset would be non-trivial. How many images of each grade of each type did you provide for the fine tuning? And were they all TrueView-style images or a range of in-hand or eBay quality pictures?
The challenge with large language models is they will confidently answer based on a hint of a signal whereas coin grading requires rigor. An integration with ChatGPT from a query perspective is straightforward enough but the underlying model could be built using a more "vanilla" supervised image classifier, given enough train/test examples.
Ultimately, the comparatively low volume of graded coins and major value differences in "jump grades" precludes widespread adoption of these sorts of automated classifier models. Grading modern mint products would probably be a reasonable use-case for automation but even then, it's just one segment of the market.
@jmlanzaf I dont use Alexa either and never will. Not a big fan of social media either. FB, IG and whatever else may be out there. I also turn my phone off at the movies. 😀
Successful BST with ad4400, Kccoin, lablover, pointfivezero, koynekwest, jwitten, coin22lover, HalfDimeDude, erwindoc, jyzskowsi, COINS MAKE CENTS, AlanSki, BryceM
I tend to agree, although part of the issue is that ChatGPT was not trained exhaustively for the task. Future AIs that exist in a constant learning mode could probably get as competent as any human using the same photos. Of course, removing the guard rails had other pitfalls.
Well, hallelujah! Lol. I wish people would at least silence their ringers at movies and in meetings.
I don't recall seeing better images of coins on this forum. If the AI can't even get close to the real grade with images of the high quality seen here, my take is it's hard to take seriously.
Great spirits have always encountered violent opposition from mediocre minds.-Albert Einstein
He's already been using it. Just doesn't know it yet. lol
The whole worlds off its rocker, buy Gold™.
Although I don't have a full understanding of this technology, I suspect that the only way for it to work well would be with consistent photography used for the inputs. In other words, it could only be very accurate in a controlled environment.
A company like PCGS could potentially supply consistent enough training data (trueviews) and inputs (also trueviews). So if the AI proved accurate, then maybe it could be a starting point for the human graders to consider. But hard to imagine it could grade accurately even with excellent images. Maybe a very consistent format of hi res videos to get angles / luster / etc.
Instagram
What is the reason for society's push for the development of AI?
In what manner will AI benefit and improve the lives of individual human beings; and in what manner will AI worsen the lives of individual human beings?
That needs its own thread.
There are numerous benefits.
In what manner will the intent benefit and improve the lives of individual human beings: and in what masher will the internet worsen the lives of individual human beings?
I once heard about computers grading coins and how it didn’t work, and it was asked what about now with the new technology and he said it would just make the same mistakes but quicker.
My YouTube Channel
I know for myself, the biggest reason I look forward to it is to have it intelligently create scripts to perform repetitive entries into software applications. This week, I manually updated 300+ records in two cloud software applications with 8 entry fields for each record. It’s not complex work, but complex enough actions between the two apps that there’s no easy way to do it fast. It is mind numbing repetition, makes my hands and neck cramp and takes me away from tasks that add much more value overall. If I can provide a spreadsheet and tell AI exactly what it needs to do for each one and to rinse/repeat, it’ll likely save me from health issues like carpal tunnel down the road.
To keep this related to coins, those type of software operations occur at TPG’s too even if the actual grading itself is not done by AI. Optimizing them would speed up the grading process. The obvious downside is lost jobs for the people doing these jobs, but I would hope it opens up other avenues for productivity and efficiency, better customer service and value for our dollars, etc. A big hope and I have no idea how it’ll all go down, but just sharing one example of a potential benefit.
Try these two maybe it can answer my GTG coin (top).
Try these, I would be surprised if they got them right.
Type collector, mainly into Seated. -formerly Ownerofawheatiehorde. Good BST transactions with: mirabela, OKCC, MICHAELDIXON, Gerard
Thanks for the insight!
An interesting use case could be using AI to count the number of dentils around the edge as this can identify rare varieties. I'm not an EAC collector so I'd love it if someone could confirm this. I would love a high-quality image of an EAC to test on.
I agree. I suspect OpenAI's data set on coin grading comes from public high-quality pictures from third-party graders. This data only shows one photo for each side of the coin. Even though it's high quality, it doesn't reflect different lighting. I'm exited when ChatGPT allows uploading photos. I suspect the grades will be much more accurate.
I test the app out with the owner of my local coin shop yesterday. The accuracy was all over the place. Some coins had grades in the right ballpark. Some grades were over 20 points off.
@Coinscratch
Here are the results for your two coins
1978 Kennedy Half Dollar [Coin #1]
AI Grade Given: MS 63
Photos Used
Full AI Response
1978 Kennedy Half Dollar [Coin #2]
AI Grade Given: AU 55
Photos Used
Full AI Response
@Ownerofawheatiehorde
Here are the results for your coins
1861 Seated Quarter
AI Grade Given: VF 20 - VF 30
Photos Used
Full AI Response
1880 S Morgan Dollar
AI Grade Given: MS 64
Photos Used
Full AI Response
I think your algorithm will be more accurate if you provide it with pictures of the entire slab.
My Adolph A. Weinman signature
That is interesting and pretty cool. The second Kennedy graded AU so I'm wondering if the lighter and darker shades on the particular color toning threw it. Or did it actually pick up some on some wear, of course we'll never know cause I'm not sending it in now
Have you tried a Trueview image yet. This one is graded MS67 and is going back one day for another look regardless. And for the particular year 1982 I wonder if the AI takes into account the known minting problems.
Same coin slab shots.
Surprising, the AI hit the Morgan right on the nose, but had some trouble with the quarter. I think it's having trouble detecting wear on coins.
Type collector, mainly into Seated. -formerly Ownerofawheatiehorde. Good BST transactions with: mirabela, OKCC, MICHAELDIXON, Gerard
Retracted - I recommended a specific image resource. However, I see from one of your earlier posts that it is not currently possible to upload images in order to fine-tune the model.
The 80 S Morgan looks 63 to me
Successful BST with ad4400, Kccoin, lablover, pointfivezero, koynekwest, jwitten, coin22lover, HalfDimeDude, erwindoc, jyzskowsi, COINS MAKE CENTS, AlanSki, BryceM
This is one of the more interesting threads that I've seen here in a while. I have nothing else to contribute except my opinion.
In person the Morgan is on-par with some of my other slabbed 64’s, though with my crappy pictures I can understand why.
Type collector, mainly into Seated. -formerly Ownerofawheatiehorde. Good BST transactions with: mirabela, OKCC, MICHAELDIXON, Gerard
@Ownerofawheatiehorde What grade is it on the slab?
Successful BST with ad4400, Kccoin, lablover, pointfivezero, koynekwest, jwitten, coin22lover, HalfDimeDude, erwindoc, jyzskowsi, COINS MAKE CENTS, AlanSki, BryceM
It is in a 63 slab, and looking at it under the light next to some of my other 64’s I can’t help but agree with you. It has less heavy marks, but more light chatter which probably lead to it’s grade.
Type collector, mainly into Seated. -formerly Ownerofawheatiehorde. Good BST transactions with: mirabela, OKCC, MICHAELDIXON, Gerard
A little look into the training dataset might be insightful in this respect.
Keeper of the VAM Catalog • Professional Coin Imaging • Prime Number Set • World Coins in Early America • British Trade Dollars • Variety Attribution
>
Where can I buy "AI"-graded coins ?
People like to think they're smart. Those who need it to do their thinking are not thinking.
Well, not deep enough, anyway.
Entertaining thread. Been tried before and never has been successful. Hint: don't waste any more time on it. It would be better to use AI to understand the language of animals so that we can see what they think.... or not.
bob
vegas baby!
It is hard enough for a person to grade from photos. I don't think a computer is going to be even close. If you had the computer scan the coin itself, then you might have a chance. Looking at the photos/grades here, it looks like a lack of bright lustrous surfaces due to toning are picked up as wear.
This might be a neat tool for identifying the various die/parings and variety of coins for the specialist. But someone will have to input and teach the AI what it's looking for.
Maybe it's biggest help would be for those people thinking they have a rare high dollar modern "error" coin. AI can look at the scans and output "PMD parking lot find"... and save us from having to do it instead.
@Batman23 toning shouldn't subdue the luster. Heres an extremely toned dime. Look at the blazing luster.
Successful BST with ad4400, Kccoin, lablover, pointfivezero, koynekwest, jwitten, coin22lover, HalfDimeDude, erwindoc, jyzskowsi, COINS MAKE CENTS, AlanSki, BryceM