Coin Grading: Value of the AI Machine
Higashiyama
Posts: 2,279 ✭✭✭✭✭
Suppose that I have built an 'AI Machine' that grades at PCGS standards with a very high level of consistency. In effect, it completely duplicates the PCGS grading function.
Suppose that it can be relatively easily calibrated to other standards.
What is your estimate of the market value of this machine?
Higashiyama
1
Comments
You mean 'Hal'?
What you speak of is most certainly the future of coin grading. It wouldn't be too hard if one of the big AI shops like google wanted to do it. As for what something like that would be worth? I would estimate 3 X annual profits that PCGS generates from its coin grading services.
Do you also mean it is accurate and cannot be intimidated by whining, or deep egos.
You would want one.
[Such a critter is already available. It does what computing devices do best - precise, repetitive, consistent, loyal, housebroken. Do all the ordinary stuff and leave the flavor to peoples.]
pcgs buys it and destroys it
Yes, it is the classic business dilemma - does the established player adopt the disruptive technology, or try hard to resist it by any means possible!
Patent it and open your own TPG.
I can see it now: Eliminate the human error of grading by submitting your coins to XYZ company!
bob
I'm thinking of it as a kind of 'exit strategy' that benefits both CU investors and the collecting community long term.
PCGS adopts the technology; a somewhat new grading standard is established; the cost of grading is very materially reduced; human error is eliminated (no algorithm is perfect, but assume consistency is very high); huge revenues are generated through regrades and crossovers; investors recoup capital with a substantial return; the coin grading arm of CU is spun-off as a mutual organization (essentially non-profit) that owns the technology.
Grading standards are set by the collector community, by what's sent in and accepted back into, again, the collector community. The grading companies sort out what's sent in, finds the best middle ground that works out best for everyone by passing enough coins that will satisfy the greater number of people. Every grading system has its balances, imbalances, checkpoints and quotas that need to be met. For example; the average income of the larger percent of people buying into coins.
It should be common knowledge if a company becomes too strict, very few coins would qualify and a great number of submitters would become very unhappy, they would discontinue having their coins graded. The collector community has accomplished only two things really. The first, forming grading companies, two; forming CAC. Coin clubs and coin shows have been around since the beginning of time/coins. But we're all trying to find our little niche in the grand scheme of life and its not just coins. Do I have any advice to offer? Not really. Is the greater majority buying the coin and not the holder? Hardly! The people where it matters the most.....are they here and is there a much smaller anonymous group?
Maybe, just maybe, we're the ones setting the standard. Can I get another, "hardly" on that?
Originally for another thread I didn't get around to posting.
Leo
The more qualities observed in a coin, the more desirable that coin becomes!
My Jefferson Nickel Collection
Being there are not many TPG's, I'd say sell the machine to the highest bidder.
Great transactions with oih82w8, JasonGaming, Moose1913.
AI will be the future of coin grading....and then true standards will be established. The market, after initial resistance (as always - TPG's, CAC etc.) will rush to get the 'real' grade....millions of slabbed coins will be regraded...There will be disagreements over which AI is best, but eventually, the standards will be set and the market will consolidate. OK... I will put away my crystal ball for tonight.... but you heard it here first (and actually, the AI prediction, I have made for a couple of years). Cheers, RickO
Are people going to buy these coins? Who will be giving the "AI" a I/O program? People? Seems to me when the coin is graded a well respected Person will come along and put his seal of Human approval on the resulting product. Perhaps
HAC human acceptance corporation.
If the "AI" makes a mistake will we have the same laws regarding not "smacking it upside the head?"
I can foresee the liklihood of handheld optical scanners that would make the job of grading easier, but grading is imperfect, machines would not be able to draw the line on the difference between gradeable and non-gradeable coins.
Personally I'd like to see it... Of course someone needs to tell (program) the machine to tell "what is what". I suspect it could be done now, (possibly using facial recognition software as a starting foundation) if someone wanted to put up the money. If NGC, PCGS and JA (cac) would get together and agree on the grades of a few thousand samples of each series, to teach the machine with, we could be off to a good start. (Oh, and be able to detected and throw out counterfeits)
Everyone always forgets that AI must be trained in MANY examples for a given class (I.e. MS-63 Morgan’s). Who decides which examples are used for training? HUMANS!
Choice Numismatics www.ChoiceCoin.com
CN eBay
All of my collection is in a safe deposit box!
I'm shocked no one has mentioned the impact on "cabinet friction". Especially on thousands of coins that likely never saw a cabinet.
How long would it take the coin doctors to find ways to "improve" a coin that the machine has not been programed to see?
Seconds... we already know how to convert stop signs to speed limit signs, at least in the view of the self-driving car, without being noticeable to the human.
ANA 50 year/Life Member (now "Emeritus")
My question is how would a machine determine eye appeal and toning?
A smart coin grading business would never mention they were using this technology. They would enjoy lower labor cost, increased profit and make the usual excuses for price increases and service declines. A privately held company would do better because they do not have to disclose material events or balance sheets.
If it were possible to develop such a machine, many of us would still disagree with the "grade" (as in market value) as we all have different tastes. CAC, Eagle Eye, etc., astute dealers and collectors would still do their thing and there would still be large fluctuations between coins of the same "grade".
About 10 minutes.
Just so everyone here is aware, AI is not explicitly programmed to see/look for anything. It finds its own abstractions to maximize classification.
Choice Numismatics www.ChoiceCoin.com
CN eBay
All of my collection is in a safe deposit box!
This machine eliminating the need for crack-outs and regrades is predicated on there not being the need to fine-tune or retrain as a result of observed errors in grading, and grading standards, both of the grading service and market into which its product goes, never changing.
Keeper of the VAM Catalog • Professional Coin Imaging • Prime Number Set • World Coins in Early America • British Trade Dollars • Variety Attribution
Some smart kid will probably figure it out and write a free app for our smartphones before long.
Two big problems. Luster and weak strike coins. Can AI see the luster properly? Can AI tell if a coin is just weakly struck with full luster from a circulated coin?
Those are the two biggest hurdles.....but then comes into play grime and dirt and toning. Will AI find artificial toning to be real or ignore?
oh my
Doctors just hack the machine to hand out whatever grade they need
RE: "If it were possible to develop such a machine, many of us would still disagree with the "grade" (as in market value) as we all have different tastes. CAC, Eagle Eye, etc., astute dealers and collectors would still do their thing and there would still be large fluctuations between coins of the same 'grade'."
No. That is confusing objective, consistent results with "taste" and opinion. Once calibrated, the machine has no opinion and no taste.
A meaningful "grade" can only be based on the surface condition of a coin/medal. Everything else is personal opinion and should be an active part of discussion of value.
"Once calibrated"? To what standard? Surface condition alone, ignoring eye appeal? If the grade is only calibrated to determine things like hits on the surface; hairlines, strike, luster (?), then I take your point. However, I'm assuming that the machine would "learn" the more subjective characteristics of eye appeal that make up a big part of today's grading standards - which we may legitimately not agree with according to our own tastes. I would go farther and include some of the more objective criteria as matters of taste as well, but there would be more agreement on those criteria.
The "standard" is whatever objective criteria is desired - use of the ANA as an mediation body worked in the past, and could do so again. Abrasion standards among the major TPGs are similar and reasonably stable. The AU and UNC grades would need to be examined.
However, once established by one service provider. It seems likely that theirs would become the standard.
I'll repeat it: "Taste" and personal opinion cannot be part of a fair, honest and repeatable coin/medal grading system. All of those personal ideas are for buyer and seller to negotiate in determining value.
Yeah, some 200+ year old coins might have been slid into or out of a little envelope, or ben turned over on a velvet tray, but never once placed in an actual cabinet.
Liberty: Parent of Science & Industry
this machine already exists in the PRoC.
They use it to reproduce "reproductions". They developed SW to guide a complicated machine which they also have, to produce these Fakes. Just ask our host if they did ever grade one of these 1861 NS (I think) 1 cent coins.
Now, if they can do that so it is not indistinguishable from an original, I am convinced they could also add many other coin properties to the SW of that machine to evaluate the coin.
that was actually a serious post not a joke. I know its a rarityfor me but every computer system gets hacked and there would be plenty of incentive to hack this. Give other collectors low grades or genuine , cleaned etc. add points to your own submissions or turn your cleaned coins into straight grades. Tons of money to be made
the market value is as valuable as the man that built it. just as it will be as smart as the man that built it
The main value is the total compensation that PCGS would not have to pay the graders that it replaces.
There could be some additional value if people are willing to pay more for more accurate grades.
It would mean the customer would not have to pay again in hopes of a higher grade.
And the customer would not have to worry that other people were doing repeated submissions to get inflated grades.
The above assumes that such a machine could actually be built and developed / calibrated.
At the present time, it is pretty much impossible to "completely duplicate" (match) what the human graders do.
What if you had a more realistic situation, like the machine grades match the human ones 80% of the time?
Are the humans wrong for that 20%, or was the machine wrong, or a little of both?
How would you define a stopping point when your machine was "good enough"?
I think the answer is that the machine would have to return the same grade about 99.5% of the time when the same coin is
submitted in the same condition.
And it would have to match the human grades to some smaller % like 80% or 90%.
It is an impossiblity to match something 100% that has a variance for a reasonably large N.
A benchmark could be the "rating reliability"; how often the human team assigns the same grade to the same coin in the same condition. We have seen examples where the same coin was submitted about 10 times (once a month) and got an upgrade 1 time. That's a 90% rate; it's not too bad. If a machine could do 99% or 99.5% that would be even better.
How would they program the machine to ignore coins without full strikes to issue (handout) MS 67 and MS68 grades anyway? Blown fuses would become too costly.
Leo
The more qualities observed in a coin, the more desirable that coin becomes!
My Jefferson Nickel Collection
That's the spirit!
IG: DeCourcyCoinsEbay: neilrobertson
"Numismatic categorizations, if left unconstrained, will increase spontaneously over time." -me
Whatever that value is, the existence of the machine will lower it. As a collector, that's something I'd like to see.
IG: DeCourcyCoinsEbay: neilrobertson
"Numismatic categorizations, if left unconstrained, will increase spontaneously over time." -me
Only if I am guaranteed not to LOWER my grade which is the current PCGS guarantee,
I would like to have a portable AI grader to carry around at shows. See how many feathers I can rustle.
Early American Copper, Bust and Seated.
The 'AI Machine' would be similar to a camera with a viewer but a high tech one that's connected to a computer. Each of the sensors that would detect each of the aspects of coin grading, strike, condition, luster and color variations. Basically 'AI Machine' sees what the observer sees through the viewer and the observer would only need to apply a notation, for instance 1 to 10 or A to J to rate/gauge the above and enter such information into the computer to calibrate 'AI Machine' . How the coin is placed in the 'AI Machine' could be tricky. Some kind of device would be needed to turn the coin 360 degrees in 3 ways and be held two separate ways.
Leo
The more qualities observed in a coin, the more desirable that coin becomes!
My Jefferson Nickel Collection
Not possible, in my opinion.
Compugrade? PCGS already bought them out years ago. I have several of their slabs.
I think that a system like this could be developed and implemented as is for the grading of Moderns, and I believe that you would be able (if you wanted to) to implement decimalized grading of Moderns from 69-70, which might breathe some life into that area of the market.
It would be more difficult to develop a program that could accurately grade vintage - particularly early material, Colonials, Pioneer Gold, etc.
Assuming that technical grading could be accurately achieved with machine learning, you'd have to accept that the stated grade would be a purely technical grade - it wouldn't take eye appeal into account whatsoever, rendering the end product insufficient to the purpose of pricing items.
The solution to this would be to develop a hybrid human computer grade, where the computer assigns the technical numeric grade on the traditional 1-70 point scale, and a grader assigns a prefix modifier that describes the aesthetic "quality" of the coin, if it falls outside average product aesthetics. The easiest modifiers to use would be words that we already associate with quality levels ("Choice", "Gem", and "Superb" jump to mind). A system like this could be used to bridge the gap between current market grading and the more consistent technical grading that would develop. To wit, coins that currently get a bump in grade for great toning might be downgraded technically by a point or two, but the modifier would describe an increase in value (or registry point status).
Obviously, a system like this will not be adopted. The market and the TPGs abhor change, but it is nice to think about.
What is now proved was once only imagined. - William Blake
I believe our host was developing a computer based grading system about twenty-five years ago. PCGS Expert was the name if I recall correctly. I have heard nothing about any developments since. I believe at least a computer system or computer assisted grading system could be useful on coins such as modern SAE's and such. I can only imagine trying to accurately grade a couple nonster boxes of SAE"s a day without going nuts through sheer boredom!
When a display of the Expert machine was set up at a major coin show. A universally known numismatic publisher/collector and I took the test. We both disagreed with the machine.
@leothelyon said: "Grading standards are set by the collector community, by what's sent in and accepted back into, again, the collector community."
After reading this silliness, I stopped reading further and scrolled down the page for more opinions.
god no , no decimalized 69-70
we don't need to breathe life into that area , we need to breathe death into it
A computer could do a great job of consistently evaluating objective inputs - surface hits, DPLs vs hairlines, rim integrity, luster, and strike. Obviously it would need to recognize this in the context of each particular series. Gold simply isn't graded the same way that early copper is, for instance. Luster on 1921 Peace dollars is different than on the low-relief coins, etc. etc. It would have an impossible time evaluating eye appeal, which is entirely subjective. It could probably learn to recognize certain desirable patterns over time, but that's a much higher hurdle to jump.
Keep in mind there are significant financial reasons to keep grading squarely in the realm of expert opinion.
Finally, if you can build such a machine, someone else will do it better, faster, and cheaper in short order.
One flaw computer coin grading is up against is dirt, oil and whatever else contaminates the surfaces of coins. All coins have, to some degree, contaminated surfaces. How a computer sees through or differentiates between the smut on a coin and its true surfaces would be a monumental problem. So...would all coins need to be cleaned before they meet "Al machine"? And since most cleanings of a coin take place by the uncouth.....where does that leave us?
My babbling above has nothing to do with the article below since I hardly read it.
http://www.coingrading.com/compgrade1.html
Leo
The more qualities observed in a coin, the more desirable that coin becomes!
My Jefferson Nickel Collection
And we would all still get the allusion whoever is behind such a creation really knows what the he!! they're doing.
The more qualities observed in a coin, the more desirable that coin becomes!
My Jefferson Nickel Collection
If they do ever create a machine like this, they should definitely name it 'Hall'!
Having someone other than yourself (including an AI machine) make decisions for you is something that:
would cause some people to want it happen (they think it would make life easier for them; they think it would make them "free"; and it would absolve them of personal responsibility for the consequences of their decisions and the angst that goes along with said responsibilities) ; and
would cause other people to object strenuously and resist with all of the energy and effort they can muster.
1 above, if it comes to fruition and becomes the norm on a global scale, would be something that changes a central component of being "human" (free will and personal responsibility for the consequences of one's choices).