AI just broke through a major bottleneck in coin analysis

Well, if the "grading" AI is as accurate as "siri" or any of the other programs that are in computers or answering services, then I'd say stand by for all the glitches. Have you ever had to repronounce a word into the AI in computers? Talked to an AI computer and have them understand what you are asking? A properly graded coin goes through a minimum of 3 sets of eyes ... AI would be one.

 

Log in or Sign up to hide this ad.

Oh right, that’s what I took the effort to try to explain. So you had an argument with your spouse, and there she blows, there goes the grade! That’s all you got out of it, I’m saying it’s subjective. The weighting is subjective. Your wife could be right, Jeff, you’re a bum. Not saying take it out on the grade. Read it, again. The one before it, you need more.

 

I'm not sure if you missed my point or not. I asked how it could decide weak strike from wear, etc, and you responded that it "will be overcome with live video analysis and real-time lighting adjustments." Again my point is, if someone is experienced enough to know what sort of specific video and real-time lighting adjustment is required for coin grading, it's not a huge leap to think that they'd also have enough expertise to give the coin a grade while they're at it, rather than doing all the necessary work just to have AI grade it. You seem to be saying AI can do it if everything is spoon fed to it exactly how somebody would grade the coin. So, just have that person grade the coin!

 

I think I understand what you're saying now. The value of AI grading would depend on the user, but I agree that it would not be helpful to those who already know how to grade accurately.

A few scenarios jump out of my mind with AI grading.

a) The collector already knows how to grade a coin accurately
- In this case, the primary value of AI grading would be to serve as a sanity check. It could also help the collector identify a variety if they aren't versed in a specific coin series. Overall not that much value.

b) The collector has weak grading skills
- In this case, AI grading can help them in a variety of ways:
-- Serve as a sanity check to help them improve their grading skills
-- Determine if it would be worth it to send into the TPGs
-- Give them a sense of the grade and its value before going into a coin shop and talking with a dealer

c) Online Dealers w/ weak grading skills
- In this case, it can help them give a more accurate description of the coin when listing a coin for sale
- AI grading results can be pointed to buyers to give buyers more confidence

d) Online Dealers w/ strong grading skills
- AI grading results can be pointed to buyers to give buyers more confidence

e) In-person coin dealers
- AI grading results can be pointed to buyers to give buyers more confidence
- Serve as a sanity check when they have to leave the shop and a more junior dealer is manning the shop

f) Third-Party Grading Companies
I have A LOT of thoughts to share on this

I've been thinking a lot about the consequences for the TPGs of the emergence of AI grading. It may impact their business model in many ways. But I've also considered whether the TPGs are even incentivized to adopt AI grading.

I believe that, eventually, the TPGs will use AI to help increase efficiencies in areas like counterfeit detection. But overall, I don't see them replacing human graders anytime soon.

There are a few reasons why I think this way:

• The TPGs value is not grading. They have billion-dollar valuations because of their reputation and the liquidity they provide to the marketplace. They make it easier for dealers to buy and sell coins. It just so happens that this liquidity is provided through their grading services. It doesn't matter that imperfect human graders are used. It works and works well today. There's no need for them to rock the boat.
  
  
• They have no incentive to normalize AI grading. Right now, the TPGs are differentiated through their reputation with collectors. If AI grading becomes normalized, the market may shift to focus on how accurate their AI algorithms are as the main differentiating factor. TPGs are not tech powerhouses. Hiring AI/ML specialists to run an ever-spiraling arms race toward the perfect model will take time and money. It's much safer to avoid this altogether. There is the risk that the TPGs all end up with nearly perfect models. So, people will just submit to the cheapest grader. Or even just grade at home.
  
  
• There is the risk that collectors will balk at the idea. Many dealers remember Compugrade and their "AI" models from the 1990s, and they aren't remembered fondly. It's not worth the reputational risk.
  
  
• The TPGs have very little to gain financially from using AI. Why risk the company's reputation to save a few million on graders? What is the upside moneywise?
  
  
• AI grading would destroy all the money the TPGs get from the breaking slabs and resubmitting game. Why bother resubmitting when you know the AI will remember the same coin and give the same grade?
  
  
• Last but not least is market vs technical grading. Most TPGs do market grading, which requires a deep understanding of the market for each specific coin series. A VF Large Cent in the 1840s may have shifted to XF today because of market sentiments. AI can and will be great at technical grading. However, AI does not have a "gut feeling" about what the market would accept as the "right" grade. Only what a coin can technically grade as given objective criteria.

Note how I did not mention much about technical limitations. It's mostly market forces at play. This was the lesson I learned from my work with AI and coins.

 

Those are all interesting thoughts. I can see a more advanced version of Coinoscope (which works "some" of the time) but imagine the resources required to image every coin ever made and then also in a wide range of grades, strikes, die clashes and cracks, etc. And then for the novice grader/collector categories you outlined, have them take images and videos sufficient for AI to do an accurate comparison. I suspect it's going to be garbage in, garbage out. Is AI going to walk them through taking different images, each time telling them "do it this way, I'm looking for bag marks", "do it this other way, I'm looking for signs of counterfeiting", etc? AI might have to train the user to take pictures as much as it was trained to grade coins. The user throws up their hands, AI says here's the best I can do, and off they go down a variety of different paths. It will be interesting to see where this leads though.

 

Market grading is eye-appeal grading. And AI can't assess eye-appeal. It can trick us into thinking it can, but that's only its LLM performing NLP tasks. It can't assess eye-appeal, it can only talk as though it can.

 

This is an amazing idea, and one I haven't thought of before. The AI could have an even stronger idea of a coin's desirability based purely on the sold price and coin image. There is no need to mention its mintage, history, etc. The market has already spoken through past sales.

Data from the market could effectively teach the AI market grading. It's worth investigating further.

My only hesitation is that the AI would be trained with photos. It may not work if the AI is trained with images, but we then ask it to analyze coins using videos. It wouldn't be a true one-to-one comparison.

 

I wonder if these Chess models have achieved a "good enough" status for most people. A model doesn't have to be perfect to reach mass adoption by the market. The arms race for the perfect Chess model could mostly be for clout and bragging rights among engineers.

I wonder if this logic could be applied to AI grading. The TPGs may have different imperfect models, but if their AI grading results converge and remain consistent, the market may treat the TPGs as the same.

This convergence may be significantly more straightforward to achieve than with Chess. Chess has 10^111 possible moves. AI grading only has the 70-point Sheldon Scale.

 

This is simply not going to happen.

Comparing Chess programs to grading a coin is a strawman conversation.

There is a Board. There are a set #of positions on the Board. There are a set number of Pieces and a set number and direction that the pieces can be positioned. Even if another strawman conversation using multi-level Chess is used as an example (Thanks a lot Star Trek for screwing things up) the mathematical solution is finite and there is no gee AI Grading for coins is where Chess was blahblah, is a waste of time and an insult to Einstein.

AI will never be able to "mimic" the subjective nature of a human, and can not be trained/endowed/adapted/programmed/taught the ability, and Mr. Spock The Coin Evaluation Robot is not going to run around the Bourse Floor examining subjective graded pieces pronouncing "that's logical" will make it so.

Maybe a case can be made on some humanoid level, I don't know, but I suspect if that happens, the song 2525 will be the national anthem, and humans will be melting humanoids rather than precious metals.

I should be clear, I never and still don't buy that ANA got in right in the wayback machine and then sold themselves for 30 baubles. BS. A critical review of the the original ANA standards reveal the illogical logic of the conclusions, and THAT is what gave Birth to TPG.

The subject is not even adaptable to the I'm a Fan I'm not a Fan position, because it does not exist, and will never be able to exist, because:
AI IS NOT HUMAN.

My closing statement: mentally imagine an AI with glasses (made in China).

Savvy?

 

AI stands or falls on the data that it trains or "learns" on. If AI trains on bad data, it will produce bad results, no matter how good the learning engine. I work with some AI tools at my job and I've seen both brilliant and terrible results. At this stage, there still needs to be someone there (i.e., a person) to distinguish the bad from the good that an AI tool outputs.

Any coin grading AI tool needs to specify strict criteria for the data it will accept to train on. If the desired output is technical grading and the tool trains on a lump of data that includes examples of both technical and market grading, the results will likely come out very confusing. Some of the subjective nature of grading will likely get missed or "filtered out" over time as well, since some people might give a coin an MS 63 and others may give it an MS64. It all depends on what the engine learns on. It would probably be good to include data that includes two different grades for the same coin to account for some of this subjective variability, but it wouldn't likely ever please everyone all of the time. Human-based grading doesn't always please everyone all of the time, either.

Regardless, I think it's worth looking into AI coin grading, but to also keep some skepticism (or cautious optimism) behind it until real results are seen, verified, and argued over. AI has done some amazing things for other fields (including mine), but it's far from perfect at this point and it's often very oversold. The first phase of AI coin grading would likely be producing results from an engine based on data that had solidly defined criteria, then having human coin graders judge the output. Then evaluate, change the learning engine or the data, and repeat. This may already be happening, but it's a necessary first step in deciding whether this will ultimately work out or not.

 

No, it won't. When it can do what JF does, which it never will, I will concur.

I'm safe in stating Hordes of VC investors ( not even one-because VC investors are not foolish) will not rush to fund such a deep hole. VCs might....might dip their toe in funding a factory or Think Tank Labe that they can control for the payoff end, but it won't be to produce AI to grade coins or stamps or Art.

Mentally imagine two IA Robots discussing the grade for a piece? Form the image in your mind? Good. Now realize you just defeated the illogical logic of the attempt.

truth in editing: changed "to" to "two", so the comment does not read like an AI would type it.

 

.. and that is the failing point of the imagined AI infallibility, because any such "input" is SUBJECTIVE because it is human "input", thereby concluding it is illogical logic to state infallibility of I can be achieved.

 

I forgot:
Signed :
Maynard Krebs.

 

Ya forgot the 'G'........

 

I'll point out that "sold price" alone ignores human factors that AI might be clueless about - how many bidders, how badly two or more bidders wanted that specific coin and why they wanted it, how much ego was involved in having to be the winner instead of valuing that coin logically, how long ago the sold price was and what the general and specific markets were like at the time, whether someone was scammed by an ebay listing, whether the auction timing was good or bad, how much news and attention a particular coin got prior to the sale, etc etc. Then think of fads like toning and lowball valuations, or fascination with trivial "error" coins, which really haven't been around that long and may no longer be relevant in the future. I suspect AI is going to give us "new things to look for that make it more valuable" that don't actually exist.