Innovation in Numismatics is Rapidly Increasing - Pure Technologies
treybenedict
Posts: 429 ✭✭✭✭
Hey everyone. This post is a bit different, I wanted to publish a state-of-mind while we have been building technology products within the industry. I hope that some people will enjoy my perspective, and it can inspire others to build cool projects to help assist the innovative sector that we are now experiencing.
Background
I was sort of inspired to make this post a couple of weeks back when I met Sam Altman, CEO of OpenAI, at the Y-Combinator HQ in San Francisco. He talked about the early days of OpenAI being a research-first company and pushing into the commercial sector to try to develop their eventual goal of making AGI (Artificial General Intelligence, "self-reasoning"). In many ways I feel that I resonate with the level of research and data first, it just makes sense to me to provide facts and quantify things to be able to describe them in new ways. With strong resources and research, people can build really cool stuff that could have not been possible in the past.
Aside from the scary aspects of AI "taking over the world" there are actually many amazing use cases that could be used to transform our lives for the better (eg. Advanced Healthcare). Obviously, Jobs will be lost, jobs will change, it's the cycle we have faced over the last century with as much innovation has been pushed out.
Similarly related, We are in a really awesome time for innovation within the Numismatics sector, with a new grading company emerging, a new collector-based social network arising, and bounds of new technology being pushed. I believe that the opportunity should be grasped and pushed forward by the community.
Now to the actual meat of what I wanted to write about:
Current Problems with the Industry/Hobby
There is a massive information curve for learning about Numismatics.
Lots of people ask why the Sports Cards craze and the Pokemon craze were able to capture so many people within the collectibles industry? Simple answer is that the information barrier is far easier to grasp, with less grades, more streamlined collections and digitized records. PSA did a really good job on this, promoting the hobby and ramping up hiring top talent to help boost their operations as well as revive a once-dead industry.
On the other hand, Coin Grading, Authentication, CAC Stickers, are all complex layers on top of actually understanding what coins are worthwhile collecting and even picking a series. To make it even more difficult there are different "looks" for each series that collectors/dealers identify as choice, and often take large periods of trial and error for people to be able to build incredible collections. This choice paralysis for new collectors is hard to understand, and everyone who knows the information asks for $$$ in return to "help" out.
These steep information curves cause crowding out problems as well as market inefficiencies, it's simple economics. This allows a smaller market of highly knowledgeable people to be able control the narratives of consumer buying as well as pricing. You are likely one of the people in this circle if you are reading this post.
However, the core side-effect of this is that new generations of collectors and dealers will repeat the same process, with far less participants. Obviously, I was not born for the year of the coin dealer, or the peaking market of 1980, but I can tell you that the new generation of collectors is far less plentiful, despite 2023 being one of the best years for coins in quite some time. Yes, there are great programs to try to expand to YN's and new collectors, but frankly they cater to those already interested in the field, and are likely generational influences that get them into the subject. We need more new people in the hobby.
I am far more optimistic than pessimistic on this subject, but I do believe that everyone will have to be on the same page to be able to push forward to a new generation.
Why using Technology in Numismatics Makes Sense
I do want to clarify that the majority of "AI" described is a mix of Computer Vision pre-processing along with Neural Networks to be able to comprehend and process the image data at hand. Luckily I do thank our hosts for putting together the largest database of Images & Variety guides which is the stepping stone to building great tool for the future of the hobby.
There are a couple of incredible use cases I can see for technology being involved to lower information barriers...
First is variety detection, being able to identify and analyze existing die patterns as well as new die varieties that we may not have known to have even existed in the past. Being able to overlay many thousands of images to see die states and progressions to get a deeper understanding of the particular coin at hand. Typically a lot of this is done by hand at the moment, and many great databases and books have been published to help collectors and dealers alike. New die varieties come out each year, and previously graded coins can be retroactively attributed even in their old holders with the right image processing. This is incredibly important for Foreign coins as well since the resources and documentation of many series is highly lacking compared to that of Morgans or Lincoln Cents.
Being able to digitize these records along with computer vision analysis is an amazing tool to be able to share with all collectors, not just new and upcoming.
The second is defining better standards for modern grading. This is likely the most obvious use case that could help free up lots of time and clarity within the process of grading "70s" vs "69s".
For Instance... Tell me which one is a 70? or are they both 69's? or both 70's? Imagine this is your job as an entry-level grader! You'd likely be scared to give out bad grades or over-grade coins.
For new collectors, there is virtually no difference, and many entry level participants are introduced to Vintage coins through the buying of bullion products. Simplifying the bridge is a huge push and I think this could be very well assisted with "graders notes" or pre-screening done with Computer Vision to give the collectors a better idea of what they are buying and what is in the holder.
The Most Important use case is counterfeit detection. Most recently, I was attending FUN and saw a couple of raw counterfeits that the seller believed were authentic (I don't want to name the coins or the seller), but the coins were quite high in value which is quite concerning for newer collectors and even moderately seasoned numismatists to have in the marketplace. Personally, I believe that given a good amount of research, this could be a tool that could be open-sourced for the greater good of the hobby, allowing for the barrier of entry for information to lower for those who want to get involved in the hobby without getting burned.
I'm not just talking about image guides, but more in depth analysis to point out die characteristics, and forensics that can not only dictate but explain why something is not authentic to the collector.
Myths about AI Grading & Data
There are a couple myths of why people think AI Grading is impossible, I'd love to share my perspectives after actually approaching this problem.
Myth #1 "Computers cannot grade as well as Humans"
This is completely false. Neural Networks are trained off of more data than any coin dealer or grader could ever see in their lifetime. With a large enough dataset, the bounds for grades are already predetermined when being tagged and set, making the accuracy only dependent on the amount of data provided. These models understand far more about what dictates the difference between grades than a human because they are trained upon all the possible information known to Numismatists. This makes economic sense for high-volume coins such as morgan dollars or silver eagles, not so much sense for $4 Stellas and Flowing Hair Dollars.
Myth #2 "AI grading will take graders jobs"
This is also completely false. AI grading does not have product market fit. The resources made to be able to do this would cost more money than the return of grading the coins, and manual labor is far less costly than the amount of labeling and training needed to have incredible accuracy.
Models can also be trained and supervised by industry experts to be able to carry on their legacy of grading, eye appeal distinctions as well as variety publications.
Myth #3 "Data is the most valuable resource"
Imagine you own a plot of land that you believe has the chance to strike it rich with gold. If you don't have the means to extract the gold, then it is purely hypothetical. This is the same case with data. Data can only be useful if it provides service in analytics, otherwise is it typically better starting over than trying to re-engineer broken data sources that may not have all the needed attributes.
How to Actually Approach the Information Problem
Collectors posted a great article on pre-screening of cards at PSA through the use of VGG-16, a convolutional neural network that can identify and recognize cards based on their appearance. Essentially PSA has built a large enough database to be able to tag and understand what a card is previous to even manually viewing the card.
A glimpse into how it works, if you want to read more checkout their article: https://blog.collectors.com/image-search/
Similar processes can be done for Counterfeit detection and Die Varieties, given a good enough dataset. This could provide an amazing resource not just internal for TPGs, but for more people to use in their day-to-day collecting lives. It could even serve as a resource for modern grading and speeding up turnaround times!
Keeping in other industries, one of the best resources I've seen is the CGC graders notes for comics. Essentially the grader will designate why the comic graded as such, and what attributes it has or is missing. This is a key insight for comic collectors and is done for every comic that passes through the grading room. What's there to say that this could not be done for coins? It would provide an interesting legacy and detail to the coin that goes beyond just the holder.
Furthermore, Open-Sourcing resources. Information about what a coin is worth should not be paywalled, and the IP associated with that coin should belong to the owner. Guides should be fact based and not pushed out on a whim. Also honesty and integrity should be held at a high standard. Far too many participants are actively deceiving others for financial gain which is highly unsustainable. We must protect the future generations of collecting in the hobby, now is not the time to make deceitful profit.
Anyways, if you made it this far. Thanks for reading, I have a lot more to say but figured I'd have this as a brain save for future releases and articles.
I'll be at the ANA at table 307 if anyone wants to chat 
Expect a couple product releases in concurrence as well.
Comments
IMO, the consistency in both straight grading and rejecting a coin (details) is atrocious. In my short time collecting, I've broken out several details coins only to have them straight grade with another TPG. I don't mind this so much but loathe the fact it's not explained how or why it was details. Based on how it affects value, it might be considered a crime. AI could increase the throughput and provide a text and visual map of a coin with little additional effort. I hope to see this implemented before I'm six feet under. The humans obviously need some help. Thanks for your post.
USAF veteran 1984-2005
Thanks for the comment! I think it is totally feasible to create such aspects. For sure something we hope to approach in the near future.
I am not worried about AI taking over numismatics. In the end, humans have to buy them.
I am a newer collector (started April 2020), and I primarily focus on U.S. Half Cents and Type Coins. Early copper is my favorite.
Truth. I think that it can only help the growth and information barrier that our hobby has moated.
The grading system would cease to function if grading anything became too perfected. Submissions would drop. Folks are unlikely to submit if not profitable or at least recover their fees. Grading of anything needs to be imperfect so we don't become too dependent. There are many levels of collectors who will seek the lesser quality of graded coins and be content.
Leo
The more qualities observed in a coin, the more desirable that coin becomes!
My Jefferson Nickel Collection
AI is hugely controversial; one obvious case, the Hollywood writers strike. Technology will continue to displace humans, and we need to really consider what the real world consequences are of having fewer and fewer meaningful jobs in our society. Robots will presumably be replacing humans in the box stores and there are many more areas where the displacements continue to be problematic. I believe we were better off with more mindful jobs of a clerical nature, libraries, box stores, more human to human interactions keeping people grounded. Plus the waves of foreign and illegal labor are taking away jobs from locals. Most of the contractor jobs locally are heavily staffed with foreign labor, few jobs now for high school and college kids or casual labor projects. If we have the choice between employing fellow Americans or cut-throat corporate sharks looking to cut costs to service their unhinged profits, we should opt for the first choice every time! Edited to add; a friend sent me a link to a lawsuit due to AI algorithms being used to deny healthcare: https://www.axios.com/2023/07/25/ai-lawsuits-health-cigna-algorithm-payment-denial
I appreciate your work. I think it is in the cards. Good luck. Peace Roy
BST: endeavor1967, synchr, kliao, Outhaul, Donttellthewife, U1Chicago, ajaan, mCarney1173, SurfinHi, MWallace, Sandman70gt, mustanggt, Pittstate03, Lazybones, Walkerguy21D, coinandcurrency242 , thebigeng, Collectorcoins, JimTyler, USMarine6, Elkevvo, Coll3ctor, Yorkshireman, CUKevin, ranshdow, CoinHunter4, bennybravo, Centsearcher, braddick, Windycity, ZoidMeister, mirabela, JJM, RichURich, Bullsitter, jmski52, LukeMarshall, coinsarefun, MichaelDixon, NickPatton, ProfLiz, Twobitcollector,Jesbroken oih82w8, DCW
I would love to have an AI search give me results for what I am looking. Search as it is now is super time consuming.
You make a lot of different statements in your long post which are not supported by evidence or even by good theories.
I disagree with many of them, and I am not feeling motivated to list them all and the reasons why.
In general, I don't think you will ever have enough training data to make an accurate computer grader for all series.
You might be able to do it for moderns, but even that will take multiple photos per coin, especially for proofs.
That data might exist at some point in the future for moderns, but I don't see computer grading of moderns as particularly interesting anyway.
I frankly agree with Yos, classic coins are an outlier they are different than Pokemon and Sports Cards etc. Classic coin collecting is complicated and truly successful collectors will need to become experts leaving less knowledgeable collectors to fend for themselves and ultimately fail and leave the hobby than the other less complicated collectibles out there. Al will fail as well as human graders will fail due to greed. CAC will be the exception to the rule.
I applaud your innovative mindset and approach. You are going to do great things with that natural curiosity. Don’t let old farts in this hobby dissuade you from disrupting the status quo.
Latin American Collection
AI has the advantage of accumulating quite a bit of data and sifting through it. I could see it helping with coin grading where there is a plethora of information / coins available, like the example Trey gave with moderns. I don't know if enough data is available, for say, Flowing Hair dollars.
"Seu cabra da peste,
"Sou Mangueira......."
If PCGS, NGC, ANACS, ICG, CACG, etc could use the same AI algorithms and datasets, it would help to solve industry grading inconsistencies, but that won't and can't happen. AI grading would help considerably with grading turntime. Like some others, I am doubtful about AI grading with early US coins.
Nice writeup and post!
I'd like some AI to do something like to help me find me find all my 1868 coins among all my coin images in my library. Is something like that out there?
10-4,
My Instagram picturesErik
My registry sets
I agree. Grading has too much subjectivity for AI to offer a quantum leap forward in grading. Just as an example,
toning has been both in and out of favor in recent years. I remember when toned coins were downgraded relative
to "white" coins. (Never mind that a lot of those "blazers" had been dipped). Now it seems like toned coins get
a "bump". I love ranibow toners, but their prices are in the stratosphere. Who knows what changing tastes/perceptions
will do in the future.
Also, just read another thread on price realized vs PCGS Price Guide for 1887 P Morgan MS 63. This is about as generic
as a coin can be. Post showed various examples of this date and grade for sale. Asking prices ranged from $85 to
$150. How will AI affect that? Can't tell from photos, but I'm guessing the $150 example is the nicest of the bunch. There
will always be subtle differences between graded examples of any series in any date, etc.. Imagine the "fun" of differentiating between examples of a rarer date, or super high grade.
I don't think you read my post thoroughly. There is an information problem in the industry. I don't think AI grading is PMF, but I do see it solving a lot of the information gaps. Data Augmentation exists and is in high use of training models.
OCR can do this fairly easily. Although the fonts for each 1868 may be different, this is like a hello world to the larger picture and can be created/found pretty easily.
Thanks!! I will stay heads down building towards these goals as long as I can.
Do you think natural exclusion due to information gaps is a good thing for the hobby 50 years down the line? This is the core reason I wrote this post. These information barriers can be subsidized using technology to aid the new collector to be on a level playing field. This is how we transcend the hobby to new generations
I actually think this is one of my biggest goals. Each coin is unique and shouldn't be catered towards one general price guide or sheet. It should be fairly simple to classify attributes within each coin that separate the value their peers. We are already doing this with toning
I would not dismiss innovation in grading down the road.
AI will assist in machine learning, the chips and tools are here already.
A hurdle, but not significant will be the mechanisms to allow the material to be analyzed at different angles and lighting, with safety and reliability. Now we are talking consistency and quick turnaround.
Lets face it, buyers, sellers and every tier of the market wants and needs innovation.
Literature review is a key part of research. Review the obstacles that Compugrade faced. Locate the book with the chapter titled: Computer Grading Mint State Morgan and Peace Silver Dollars, "The Art Becomes Science".
Compugrade is interesting. They approached the problem in a much more segmented way (literally splitting up the coin into small segments with different values). It seems my problem is more of an information problem. I have read this entry though, and I appreciate your citing.
Do you have any suggested software?
10-4,
My Instagram picturesErik
My registry sets
Apple has integrated OCR into the photo library. Android not too sure.
Actually would be good if there was desktop software that could use to scan a coin image directory and pull the dates from the images. On my Android phone, I have Microsoft Lens which is hit and miss finding the dates. It has a tougher time with toned coins.
10-4,
My Instagram picturesErik
My registry sets
Not sure if any out-of-the-box solutions are free here. Sounds like you'd need to get handy with Python and an image detection library like OpenCV and do some basic typeface training. I could write you a quick one but I'm terribly busy at the moment.
EDIT; I messaged you a script.
In my opinion it's very simple we are all researching and AI is on our shoulder learning and observing, you all think I miss nothing well I think the machine misses nothing, grading aka matching up all data, yes I think possible and personal bias yes the machine will never be able to interfere.
Hints my comments like, "you all would find a reason to go against the AI", which I still stand by today.
I think it is a very awesome post and the bots/spiders are working, what I mean is we are all feeding AI.
Thanks all
Ps to answer if machines will control, shoot I have no clue, just they are here and are learning 24/7 no breaks.
Oh and in my opinion the one on the left is MS 70 the one on the right MS 68 maybe a plus but to me the novice, the toning or slight color change -1 and I think I see a mark -1
Again I'm no grader and this is only obverse assessment and I could be wrong.
Thanks all, I mean to voice my opinions on coins and everything.
Ok and to be honest yes I'm leaning towards AI because the machine and a few humans have been able to see the similarities, just AI took it further and associated mine with what it knows.
Not saying anyone is right or wrong just saying.
Grateful for your insights here @treybenedict. Very excited to continue to see how Pure is able to bring the future to fruition for the betterment of the hobby as a whole! I, for one, welcome our new coin overlords.
Thanks Matt! Always happy to share my thoughts publically. Let's connect again soon I've got more to share before the ANA.
One thing to think about here is that a human can quickly tilt a coin in the light and then zoom in using a loupe so a person can get many views of a coin quickly in hand.
Some computer systems just use a single still image of each side so such systems could be working with a significant deficit of information compared to the humans.
AI grading may not have product market fit, because myth #1 may not be a myth at the moment.
Data is very valuable for AI projects. Open AI was able to train its models on data it didn't license, but the data owners are starting to take issue with this so will be interesting to see how this works out in the longer term. In the PSA example above, who owns those images? For coin images, what would happen if the companies that owned the images, withheld license to their images to AI application developers?