GPT Auction Descriptions... 1 1

[Fenntucky Mike]

...are pathetic at the moment. 

I noticed the other day that one of the bigger overseas auction houses is using GPT to generate auction descriptions for all lots, and the few I've read so far are either misleading or flat-out wrong. What, if any, consequences will there be? Will there be more opportunities for knowledgeable buyers? Will there be less "deals" to be had in the future as AI advances and is able to identify coins and banknotes better and pick out varieties? When will AI be used to grade coins and determine value? Will AI solve the '64 SMS coins riddle?  

Below is one listing with GPT description. This is not the worst I've read but wanted to pick out a U.S. coin since that is what most here collect.

Edited February 18 by Fenntucky Mike

[RWB]

A very large proportion of material generated by current semi-AI products are wrong. The products are useful in making quick summaries, or simple calculations. I've tried some numismatic things with Chat GPT and it has failed all but the most basic. Someone on PCGS message board tried it for coin grading, and the results he posted were uniformly incorrect and misleading....no value at all.

[Fenntucky Mike]

On 2/18/2024 at 10:53 AM, RWB said:

A very large proportion of material generated by current semi-AI products are wrong. The products are useful in making quick summaries, or simple calculations. I've tried some numismatic things with Chat GPT and it has failed all but the most basic. Someone on PCGS message board tried it for coin grading, and the results he posted were uniformly incorrect and misleading....no value at all.

AI is in its infancy and is quite often wrong or misleading when generating descriptions, documents, etc., for coins and most things in general as you have stated, but where can it go from here? How useful can it become and how will it be used in numismatics in the future?

To expand on my hypotheticals in the OP. If auction houses generally shift to AI and use it for lot descriptions will subpar and misleading commentary provide an advantage to a more knowledgeable bidder who relies less on descriptions, or will it just keep the riffraff out and the sharps keep sharping? A good description can help a lot and a poor or inaccurate one can hurt but how much?

As AI becomes more accurate will there be less lots slipping through, e.g. varieties that are not listed in a description or on a TPG label, minor doubling that would not normally be noted is, etc.? Will AI cherry pick all the lots, list them, and be more accurate and find more than is done now resulting in fewer deals or bargains for bidders?

With more and more high resolution images being taken of coins and banknotes, available data, and most cellphones having the capability to take these images how soon will AI become proficient at determining a relative grade or condition of a coin? 

When will AI be a useful numismatic research tool, if ever? FlyingAl's proof cameo project comes to mind, if and when enough images are collected can they be ran through an AI program to determine die varieties and identify differences (die markers)? Could AI be used to decipher mint documents from scans or images of the originals , transcribe, sort, and provide meaningful data, information or articles?

How will TPG's use this technology?

Plenty of scenarios out there for AI and numismatics, good and bad. 

[zadok]

On 2/18/2024 at 3:31 PM, Fenntucky Mike said:

AI is in its infancy and is quite often wrong or misleading when generating descriptions, documents, etc., for coins and most things in general as you have stated, but where can it go from here? How useful can it become and how will it be used in numismatics in the future?

To expand on my hypotheticals in the OP. If auction houses generally shift to AI and use it for lot descriptions will subpar and misleading commentary provide an advantage to a more knowledgeable bidder who relies less on descriptions, or will it just keep the riffraff out and the sharps keep sharping? A good description can help a lot and a poor or inaccurate one can hurt but how much?

As AI becomes more accurate will there be less lots slipping through, e.g. varieties that are not listed in a description or on a TPG label, minor doubling that would not normally be noted is, etc.? Will AI cherry pick all the lots, list them, and be more accurate and find more than is done now resulting in fewer deals or bargains for bidders?

With more and more high resolution images being taken of coins and banknotes, available data, and most cellphones having the capability to take these images how soon will AI become proficient at determining a relative grade or condition of a coin? 

When will AI be a useful numismatic research tool, if ever? FlyingAl's proof cameo project comes to mind, if and when enough images are collected can they be ran through an AI program to determine die varieties and identify differences (die markers)? Could AI be used to decipher mint documents from scans or images of the originals , transcribe, sort, and provide meaningful data, information or articles?

How will TPG's use this technology?

Plenty of scenarios out there for AI and numismatics, good and bad. 

...i guess what comes to mind is...do we even need or want that degree of differentiation for coins?...variety collectors r a miniscule segment of the collector base, i am one n i dont feel that i need it nor want it...after all AI is only operational as long as there is a grid for it to exist on...

[RWB]

AI is (and will be) only as usable as the data on which it is trained. I suspect that if one took a huge data set of TPG graded/authenticated coins, the AI output would be mush. That is what would be going in. Coin "grading" has devolved from somewhat objective into entirely subjective and money driven, not a good situation for any system based on 1's and 0's, only. A "quantum" computer might have a chance as would an entirely analog device.

Edited February 19 by RWB

[Henri Charriere]

[I've got my abacus; now where's that slide rule my father left me...

Many good points raised; all are well-taken.  I feel, however, the market will be the ultimate arbiter.

Should the levees of New Orleans be re-re-re-built at a cost of billions, or, knowing the land is continually sinking, should an effort be made to relocate?  Water always seeks its own level.

How well A.I. is received ought to be left to the up and coming generations. There are no games on my [wife's] phone either. It wouldn't matter if there were.  If you are of retirement age or older, it's no longer your world, or mine. Here today. Gone tomorrow.  A quote from Old Abe: ".... The world will little note, nor long remember, what we say here...."  Good nite, all!  🐓 

[GoldFinger1969]

On 2/18/2024 at 8:13 PM, RWB said:

AI is (and will be) only as usable as the data on which it is trained. I suspect that if one took a huge data set of TPG graded/authenticated coins, the AI output would be mush. That is what would be going in. Coin "grading" has devolved from somewhat objective into entirely subjective and money driven, not a good situation for any system based on 1's and 0's, only. A "quantum" computer might have a chance as would an entirely analog device.

At a minimum....AI can scan hundreds of sources...create a few paragraphs.....have that checked/scanned by an auction house worker for any obvious/gross errors...and now all auctions have a few paragraphs of information that used to take hours or longer to compile. 

Roger, how long did it take you to come up with that 2-page HA commentary for the 1928 Double Eagle bag and the story of the 500 stolen DEs ?  A few hours ?  A few days ?

Now imagine something getting 90% of that same story and information in a few seconds or minutes !!! 

Edited February 19 by GoldFinger1969

[GoldFinger1969]

On 2/18/2024 at 10:53 AM, RWB said:

A very large proportion of material generated by current semi-AI products are wrong. The products are useful in making quick summaries, or simple calculations. I've tried some numismatic things with Chat GPT and it has failed all but the most basic. Someone on PCGS message board tried it for coin grading, and the results he posted were uniformly incorrect and misleading....no value at all.

It could be that the nuances of our hobby and coins are so detailed and based on "eye appeal" that no computer will be able to recreate useful coin stuff until we get to Star Trek-level neural net processors. 

Seriously, I think for me right now this AI thing would mostly be good for translating voice commands into work applications (if it can do that).  For instance, I am very slow in the Office365 stuff like Excel (and Word) and if you had an AI voice command where I could TELL it what I want the pages or sheets to look like instead of having to manually do it, that would be something I would pay extra for because I am a beginner/intermediate with those programs and I would expect AI to be at Expert-plus level.

OTOH...if AI could somehow crunch hundreds of auction results from books...websites....catalogs...etc...that would be of use to our hobby.

 

Edited February 19 by GoldFinger1969

[Fenntucky Mike]

AI will be used to try and maximize profits/sales by increasing efficiency in everyday business at auction houses, TPGs, dealers, etc.. Whether it is by generating lot descriptions, improving efficiency at the office, compiling information, identifying varieties/cherry picking, etc., the collector will have no say in most of these decisions, it's only a matter of if or when AI and the equipment to support it can advance enough to preform some of these tasks. For example, if at some point AI could cherry pick varieties, minor ones, at no cost to the user why not identify them in an effort to maximize profits regardless of public sentiment. 

If only AI could process monster boxes of ASE's at the TPG firms. 

Edited February 19 by Fenntucky Mike

[zadok]

On 2/19/2024 at 6:37 AM, Fenntucky Mike said:

AI will be used to try and maximize profits/sales by increasing efficiency in everyday business at auction houses, TPGs, dealers, etc.. Whether it is by generating lot descriptions, improving efficiency at the office, compiling information, identifying varieties/cherry picking, etc., the collector will have no say in most of these decisions, it's only a matter of if or when AI and the equipment to support it can advance enough to preform some of these tasks. For example, if at some point AI could cherry pick varieties, minor ones, at no cost to the user why not identify them in an effort to maximize profits regardless of public sentiment. 

If only AI could process monster boxes of ASE's at the TPG firms. 

...diminishing returns, from both ends of the stick...the cost to implement n program will far outweigh the return, the people that it replaces wont be there to oversee the net results n then the lower end collector or speculator that looks for these minor varieties to generate marginal returns will quit looking because its already been done...cherry picking will be over, theres no picking if there r no cherries...i can see the return of the local un-cataloged estate sale being the high point for the collector, un-certified coins will be the looked-for items...profits/sales may be the driving force but just mite drive the numismatic vehicle off the cliff...this approach is very similar to the big chain stores implementing the self-checkout lanes, results...loss of margins, loss of customers, loss of service, loss of employees n now major backtracking....

[Sandon]

  The 1851 large cent in the initial post appears to have been overgraded and would be more accurately graded as Very Fine details, "cleaned" and with reverse damage. The descriptions of the coin's design and its historical context are fairly accurate, however.

Edited February 19 by Sandon

[RWB]

On 2/18/2024 at 10:35 PM, GoldFinger1969 said:

Roger, how long did it take you to come up with that 2-page HA commentary for the 1928 Double Eagle bag and the story of the 500 stolen DEs ?  A few hours ?  A few days ?

Now imagine something getting 90% of that same story and information in a few seconds or minutes !!! 

90% is not acceptable accuracy or content. It is less that what is possible from available data, and inherently flawed.

You're assuming the AI system already had the data and the set of integrated algorithms necessary to make necesary logical connections. Further, that the input request was correctly interpreted and understood by the AI system.

That little 2-page background about a missing bag of 250 DE (not 500) actually spanned almost 15 years of data collection and several hours of specific correlation, analysis, revision, interpretation and hypothesis testing before its initial draft. Add to that the basic restraint imposed by the language inaccessibility of cursive handwriting and poor OCR. A general AI system would turn out gibberish. A dedicated system might return something with most of the correct words but not the correct story. (Humans are notorious for that.)

My present view is that there is no AI system with the capability required. Present AI is copycat "fur ball vomit" good for summaries and amusement, but much too limited to enter a conversation above the "small talk" level. That will change as more powerful computing resources become available.

Edited February 19 by RWB

[Errorists]

GPT is stupid.