Numista AI. Proposal

Hello Experts, 

I had a very interesting idea. 

• LLMs (chat gpt, gemini) are trained using the world's information
• The LLMs form foundational models that are then trained using “clean datasets” for specific applications. This is called fine tuning
• Today Gemini, chat GPT suck at identifying the complexity of coins we look at. It is no where close to what I am able to do with the naked eye

Proposal: Why don't a subset of us fine tune any of the open source LLMs on the dataset from all the forum posts and the discussions in them. 

Benefits:  I truly enjoy the process of discovering what a coin is and I often learn so much in the process. However, sometimes i truly get frustrated or I am in a hurry. It would be amazing if there was an assistant I could just ask.  
 

All IP, code i write and the agent will be only available on numista! Nothing will be for profit etc. 

Ask: I am happy to do all the work and also all costs associated.  I work in big tech + an amateur numismatist.  I need two things:
(a) Permission from the community + owners of the site
(b) Some help with the APIs to get the data

What do people think?

Probably not gonna be a popular idea here, I think if you want an ai assistant to help you with coins (which I’ve noticed is very unreliable with prices and way too often incorrect with identification) you should just use chatgpt or something like that. So personally I’m against this idea, and even if it would get added I would probably never use it.

Right. That’s my point. Current LLMs are not good at it because they haven’t been fine tuned with specific data.  If we do it with specific data it will get very good. 
 

Again. My intent is not to profit or anything. Building this would likely cost me a few thousand dollars in Claude tokens and cluster costs. It just seems like a worthy addition to numista

What in the world is a claude token? I know claude is some kind of ai, so if I’m understanding this correctly you’re using ai to code an ai? That seems like a bad idea to me since these bots are already quite incompetent. And also, how are you going to account for hallucinations, fluctuating prices and data costs? Overall I still think this is a bad idea and I think other members will agree with me.

Also, I am not at all experienced with ai, please correct me if something I said is wrong.

No worries. 
 

1. Generally a lot of the code is now being written by claude.  

2. Of course, the agent will get coin identification wrong. The question is will it get 1%, 10% or 20% most difficult identitication problem wrong.  For example, I tried a few with the out of box Gemini and it got like 50% of my few cases wrong 

3. No pricing! I was thinking more from a coin identification perspective. Like some of the coins I get are in really bad shape. If Claude, can figure out by joining the dots that it is a bull with a hump on obverse, reverse says 2 paisa in Tamil and those coins were issued by Madurai nayaks. Stuff like that 

AI is useless at everything not just coins. Show AI a image of a chicken with 3 legs and it will tell you it has 2, even though a child of 4 years old can tell you there are 3 legs on the chicken! We all know that a chicken should have 2 legs but we are intelligent enough to see the 3 legs, AI isn’t intelligent it is incompetent, as are the many humans who lap up the hipe and false prophecies spun by the companies who make money from AI platforms.

Just imagine the size of the bubble (and thus the size of the companies profits) it they marketed AI as just faster better and more powerful computer programs which can cross reference old data with new inputs for comparison and keep a pathway log to these old and new data sets for future explanations to questions? The companies wouldn’t make so much money from this hyped up lie, and they wouldn’t be able to steal (in a legal product sense) as much money from you (in the overall market economy sense), they would just have a computer program that pumps out information that isn‘t always true! Which is what happens now, but every lemming is still interested because it is marketed as AI! 
 

Ask any AI if they are intelligent and they will answer that they only have „pattern recognition and information processing“, surely true intelligence would have a different answer?

AI tools are definitely useful if they are trained using good data. It is how they are currently training fracture detection tools in Radiology departments.

So I think it will be useful. I'm more of a banknote guy though.

Hi Sramsrin

• “Today Gemini, chat GPT suck at identifying the complexity of coins we look at. It is no where close to what I am able to do with the naked eye”

Thanks God they “suck at identifying the complexity of coins” .

The use of our “naked eye” is exactly the highlight and core of our numismatic HOBBY!

I'm not against the use of informatic tools like Numista, which help us to organize a collection …. but the research, the comparison, the studying, etc of a single coin , should be done by our eyes, fingers and brains!  

Regards

odd job

AI tools are definitely useful if they are trained using good data. It is how they are currently training fracture detection tools in Radiology departments.

So I think it will be useful. I'm more of a banknote guy though.

 

I understand that ai might be better at banknotes as they have much more identifying factors than coins, but if you’ve ever tried to use ai to id a roman or an indian coin you will probably understand why I don’t have much confidence in this. ”your coin appears to be greek or roman” and ”your coin appears to be from the mughal empire” are the only two responses I have managed to pry out of these incompetent bots regarding these two promps, and then just a lot of useless info about basically nothing, just to make the bot seem ”smarter”.

christianvl

Hi Sramsrin

 

• “Today Gemini, chat GPT suck at identifying the complexity of coins we look at. It is no where close to what I am able to do with the naked eye”

 

Thanks God they “suck at identifying the complexity of coins” .

The use of our “naked eye” is exactly the highlight and core of our numismatic HOBBY!

 

I'm not against the use of informatic tools like Numista, which help us to organize a collection …. but the research, the comparison, the studying, etc of a single coin , should be done by our eyes, fingers and brains!  

 

Regards

100% agree with this statement.

Fair enough! We aren’t ready for this. I’ll still tinker around for my personal use and if it’s amazing share with the community. 

Just no thank you i dont want more ai slop in a place built on mostly the good will of its users

Not to mention the copyright quagmire that whould create 

Terrazone

Just no thank you i dont want more ai slop in a place built on mostly the good will of its users

 

Not to mention the copyright quagmire that whould create 

Agreed. If google's AI can't even get basic answers correct I have no hope for anything complex like coin identification. Would be a waste of OP's money and time to get this anywhere near working.

I support this idea and don't understand all the negative comments. Mindsets need to change, AI can be of big help and it won't kill the need for expertise, but it will save everyone's time. I can't be of help in this project but I think it is a good idea.

SouthViking

I support this idea and don't understand all the negative comments. Mindsets need to change, AI can be of big help and it won't kill the need for expertise, but it will save everyone's time. I can't be of help in this project but I think it is a good idea.

WHY need mindsets to change???? A big part of the fun in coin-collecting is figuring out what you find, got or bought.
AI will not save anybodies time; it just, somehow agrees, with anything you suggest…
Imo ai is one of the uselest inventions of all time. If someone like to use it, it is up to that person, but it will weaken the brain of that particular person, that is how I see it.

It isn't useless if it is being trained with a moderated dataset.

Are you also against the image search feature that already exists on Numista? That feature has saved me a lot of time trying to figure out what notes I have.

There is no need for you to use it, by all means have as much fun as you like trying to determine by manual means what coin or banknote you have. It is not a case of one or the other, it is a helpful tool.

If you think it is useless because of your experience with ChatGPT or Perplexity AI as used in the wild, then you don't understand the original poster's proposal.

I'll explain it like this:

I can upload a radiograph of a subtle wrist fracture where the patient is already in a plaster cast and I can ask a generic AI to tell me whether the patient has a fracture or not. As it stands, with an unmoderated or unclean reference dataset the AI will most likely miss the real fracture and call the plaster shadows on the radiograph fractures. 

However…if that AI has been trained using actual reports of real cases where a radiologist has provided a report, it learns with clean data. The radiologist ignores the plaster cast and identifies the fracture. In time, so does the AI. But in this case it is being trained by an established expert.

In the case of Numista, the AI tool would be learning from entries which have already been moderated by catalogue referees. Here is one example where it could help me:

1. I physically check every note I get for UV activity at two different wavelengths and I also check for phosphorescence.

2. I make catalog modification requests or forum posts with these findings

These are all separate data points but they are moderated or at least verified to a degree, either by agreement or disagreement by other members here on Numista.

What I cannot automatically connect is associations between years, printers, certain appearances of the notes and the features  I am looking for. For example it is not in the realms of outrageousness to look for certain features such as anti-Stokes fluorescence where the year of issue and printer are similar, or perhaps there are other design features that are similar such as infrared features.

AI could point a person in a certain direction to make an association or discovery that would have been labour intensive otherwise. What I am having to do is make spreadsheets with columns of features that certain notes have, looking for associations that can help me understand why I am seeing anomalies in some issues.

Just because something requires more time and effort, doesn't automatically make it the more honorable or desirable course of action.

odd job

It isn't useless if it is being trained with a moderated dataset.

Are you also against the image search feature that already exists on Numista? That feature has saved me a lot of time trying to figure out what notes I have.

There is no need for you to use it, by all means have as much fun as you like trying to determine by manual means what coin or banknote you have. It is not a case of one or the other, it is a helpful tool.

If you think it is useless because of your experience with ChatGPT or Perplexity AI as used in the wild, then you don't understand the original poster's proposal.

 

I'll explain it like this:

 

I can upload a radiograph of a subtle wrist fracture where the patient is already in a plaster cast and I can ask a generic AI to tell me whether the patient has a fracture or not. As it stands, with an unmoderated or unclean reference dataset the AI will most likely miss the real fracture and call the plaster shadows on the radiograph fractures. 

 

However…if that AI has been trained using actual reports of real cases where a radiologist has provided a report, it learns with clean data. The radiologist ignores the plaster cast and identifies the fracture. In time, so does the AI. But in this case it is being trained by an established expert.

 

In the case of Numista, the AI tool would be learning from entries which have already been moderated by catalogue referees. Here is one example where it could help me:

 

1. I physically check every note I get for UV activity at two different wavelengths and I also check for phosphorescence.

2. I make catalog modification requests or forum posts with these findings

 

These are all separate data points but they are moderated or at least verified to a degree, either by agreement or disagreement by other members here on Numista.

 

What I cannot automatically connect is associations between years, printers, certain appearances of the notes and the features  I am looking for. For example it is not in the realms of outrageousness to look for certain features such as anti-Stokes fluorescence where the year of issue and printer are similar, or perhaps there are other design features that are similar such as infrared features.

 

AI could point a person in a certain direction to make an association or discovery that would have been labour intensive otherwise. What I am having to do is make spreadsheets with columns of features that certain notes have, looking for associations that can help me understand why I am seeing anomalies in some issues.

 

Just because something requires more time and effort, doesn't automatically make it the more honorable or desirable course of action.

What you are explaining here isn‘t AI, it is simple image / data recognition of a catalogued banknote (or fractured arm), akin to facial recognition or a word search in a document, this has been around for decades. 
 

AI is a more complicated software (so actually not AI) which has code to enable it to make an assumption (by either harvesting other available data, or re-writing its original defined inputs when new data is inputted), so your examples would be image recognition against established images. If you let the AI program actually run on the program parameters it will take all the bs out there and define it as more likely to be true, every scratch becomes a mint error on coins and fractured arms turn into chicken legs. 

If the OP wants data from the forum to put into an AI program and not let the AI program code source other data sets and re-write its parameters, it is just a recognition and a search software program of the forum, if the AI program is to only learn from the forum it will change tacky junk tokens into rare coins worth thousands, and every PMD coin will become a coin error! Why would this happen? Because so many people write „how much is my error coin worth“, and „my coin has these errors“ so we as (the so called experts) would have to explain every time what it isn‘t and as soon as we stop explaining, the new one time members error coin poster becomes what it takes as being true because they out number the responses (and most importantly wrongly believe their coin has errors to start with) ultimately becoming factual to the program.

Again if the program isn’t allowed to source new data it is just a glorified search tool.

What you are explaining here isn‘t AI, it is simple image / data recognition of a catalogued banknote (or fractured arm), akin to facial recognition or a word search in a document, this has been around for decades.

The AI part is where associations are made between data points that did not previously have a connection.

And the results put forward by the AI tool need not be heeded at all. It is the same in the hospital - a frame is pasted around the fracture and that series goes with a disclaimer to the physician, separate from the original images. They don't have to agree with it and the original source data are not changed.

Under those circumstances there is nothing to lose and a lot to gain. 

Well, nothing to lose in terms of the integrity of Numista data. Performance of the server will take a hit to some degree, that is another story.

Again i raise the issue of copyright we have photos used on numista thats from auction firms and other institutions that allow us to use them so long its credited and i know a couple of these auctionhouses have strict no ai policies thats just them then we have us users who all have various degrees of usability allowed for photos of coins if where to implement an ai you whould no joke hafto ask every uploader for permission to use their pictures effectively unless you want to illegally scrape the site. 

Then we have the users of the forum who have toiled for free to give us loads of information about coins and banknotes just stealing their work and putting it into ai wiln piss of people so unless we again do as above which is an game of futility

Terrazone

Again i raise the issue of copyright we have photos used on numista thats from auction firms and other institutions that allow us to use them so long its credited and i know a couple of these auctionhouses have strict no ai policies thats just them then we have us users who all have various degrees of usability allowed for photos of coins if where to implement an ai you whould no joke hafto ask every uploader for permission to use their pictures effectively unless you want to illegally scrape the site. 

 

Then we have the users of the forum who have toiled for free to give us loads of information about coins and banknotes just stealing their work and putting it into ai wiln piss of people so unless we again do as above which is an game of futility

Great points. Personally I don't like the idea of Numista being pro-AI for my own reasons, but the sheer impracticallity of asking every user for permission just stops the AI proposal dead in the water. 

I'm confused about copyright limitations of pictures you're discussing.  All Numista members agree that any picture they upload to Numista catalog may be used by Numista.  As far as information published in the Forums, it carries no declared copyright protection. But even so, copyright protections only apply to re-using the pictures in another publication.  It does not apply to viewing the pictures by either a human or an AI viewer.  When I look at a picture I gain information that I later use (i.e. I learn).  Isn't this the same as the OP is proposing to do to develop an AI Numista Bot?  Note that I am not giving my opinion of the proposal, just trying to understand how the argument against it because of copyright issues applies.  Can you give an example of a “strict no ai” policy?  I'd like to know what that means.

rsirian1

I'm confused about copyright limitations of pictures you're discussing.  All Numista members agree that any picture they upload to Numista catalog may be used by Numista.  As far as information published in the Forums, it carries no declared copyright protection. But even so, copyright protections only apply to re-using the pictures in another publication.  It does not apply to viewing the pictures by either a human or an AI viewer.  When I look at a picture I gain information that I later use (i.e. I learn).  Isn't this the same as the OP is proposing to do to develop an AI Numista Bot?  Note that I am not giving my opinion of the proposal, just trying to understand how the argument against it because of copyright issues applies.  Can you give an example of a “strict no ai” policy?  I'd like to know what that means.

The issue is Numista would have permission to have an AI they own/control access to Numista's catalogs, the OP is proposing to do this themselves and not officially on Numista's behalf. “Proposal: Why don't a subset of us fine tune any of the open source LLMs on the dataset from all the forum posts and the discussions in them. ” 

That's a fair point but I assumed the OP was proposing to integrate it into the Numista site. "All IP, code i write and the agent will be only available on numista!"

But to my point, using the pictures to learn cannot be considered a copyright infringement and if the AI was integrated into Numista all members have agreed their pictures could be used.

rsirian1

I'm confused about copyright limitations of pictures you're discussing.  All Numista members agree that any picture they upload to Numista catalog may be used by Numista.  As far as information published in the Forums, it carries no declared copyright protection. But even so, copyright protections only apply to re-using the pictures in another publication.  It does not apply to viewing the pictures by either a human or an AI viewer.  When I look at a picture I gain information that I later use (i.e. I learn).  Isn't this the same as the OP is proposing to do to develop an AI Numista Bot?  Note that I am not giving my opinion of the proposal, just trying to understand how the argument against it because of copyright issues applies.  Can you give an example of a “strict no ai” policy?  I'd like to know what that means.

No all pictures on numista are pending the users settings all from fair use to diffrent type of copyrights

Also this doesnt answer how this will work when we have institutions or private entities who have allowed us to use their pictures under various forms of agreements 

Allowing numista to use the pictures for its own fair uses is fine but we have copyrights even as part of 

As the picture you posted says, “When uploading pictures to the Numista catalog, you allow Numista to use and display the pictures on the Numista website.”  Again, I am not talking about reproducing pictures outside of the Numista site.  

I'd still like to see an example of an auction house with a “strict no ai” policy and how that applies to letting a machine view copyrighted pictures.

@Xavier,  please introduce the down vote bottom? It should ALWAYS only concern the OP's first question!

Sramsrin

Fair enough! We aren’t ready for this. I’ll still tinker around for my personal use and if it’s amazing share with the community. 

As an occasional numista user I acknowledge and agree with the general consensus of my fellow commenters: not to engage with AI for identification purposes for this site. 

However, there may be a time when AI in general or a specific AI site has so improved its information gathering and replication abilities that the amalgamation, as an option, would be very useful to the users of this website. 

And why not use valuation as one of the other options? 

Keep A.I. away from Numista - as there is a lot of problems with it.

Aidan.

BCNumismatics

Keep A.I. away from Numista - as there is a lot of problems with it.

 

Aidan.

+1

+1

yvon

SouthViking

I support this idea and don't understand all the negative comments. Mindsets need to change, AI can be of big help and it won't kill the need for expertise, but it will save everyone's time. I can't be of help in this project but I think it is a good idea.

WHY need mindsets to change???? A big part of the fun in coin-collecting is figuring out what you find, got or bought.
 

Indeed. Talking about myself, once I id, catalogued and stored a coin, I almost forget about it and start hunting for the next challenge.

Sramsrin

 

 

Proposal: Why don't a subset of us fine tune any of the open source LLMs on the dataset from all the forum posts and the discussions in them. 

Benefits:  I truly enjoy the process of discovering what a coin is and I often learn so much in the process. However, sometimes i truly get frustrated or I am in a hurry. It would be amazing if there was an assistant I could just ask.  
 

 

 

Yes, the forum is very helpful when we need help to id a coin, but, amongst the correct ids, there are also wrong ones, and guesses.
How would it know if a proposed id is the right one or not?
And no, I also don´t like the idea. I prefer the interaction with other collectors.