Winds Of Change In The Philatelic World

"Back to the question of what to do about the prospective loss of human expertise in recognizing forgeries, the AI angle I was thinking of was how that technology is being used to revolutionize diagnostics in health care. "

The loss of expertise is a serious issue going forward, and I feel the internet is making it worse. Instead of putting in the work to learn, too many people are willing to just put a question out to the crowd and there are plenty of clueless people willing to give them an answer.

An AI system to spot forgeries would have to be taught by an expert, and of course be updated when new forgeries turn up. Do we have competent experts willing to train an AI system?

"Isac Asnof"s (sp?)"

Isaac Asimov

First post, diving in.

My philatelic inquiries of ChatGPT, the only AI system I've tried, resulted in grossly inaccurate information on several philatelic subjects. On the topic of the grilled US stamps, I pointed out its errors and it corrected/updated its responses. And the system thanked me for supplying the corrections. So in the future, if someone uses that system to get information on that subject, they won't be told that they first appeared in the early 1860s (ugh) or that the POD began perforating stamps as a reuse-pevention measure (ugh #2). So now to eyesonwall's question:


I'm sure there are. But there would need to be a fee attached. That aside, my experience correcting ChatGPT on that topic showed that it was willing to revise its knowledge base without requiring citations or verification for my corrections. It just took my word for it. On one hand it should because I have expertise in this area -- but it doesn't know that. I logged off thinking that ChatGPT could easily turn into a giant Wikipedia of fact and fiction on any subject.

More significantly, I asked it if it had access to philatelic journals and standard references (I named a few common ones) and it essentially said "no". Here's a copy/paste of its response:

ChatGPT: I don't have direct access to external databases or specific publications, including philatelic journals and reference books. My knowledge is based on text from the internet and other publicly available sources up until my last knowledge update in September 2021. If you have specific questions about philately or need information on a particular topic within philately, I can certainly try to provide general information and answer questions to the best of my knowledge. However, for in-depth or up-to-date information, you may want to consult specialized philatelic references or journals.

I would think a more reasonable "first" step would be to train an AI to correctly identify each stamp listed in the Scott or Gibbons or Michel or Yvert catalogs with some measured reliability.

If it cannot distinguish between multiple legitimate varieties, how can we expect to have confidence in its ability to distinguish a legitimate stamp from a forgery or reproduction.

Cali Phil,
Welcome.
Without knowing the exact questions asked and which version of ChatGPT (the limited free version or the more capable paid version) it is very challenging to offer suggestions or recommendations.

All,
When posting AI system experiences please try to include a 'cut and paste' of the actual questions asked and the responses. With this information, those of us with experience communicating with AI systems can offering more meaningful assistance.
Don

Thank you Don, your points noted.

If I understand correctly AI for a specific commercial application would need to be trained for that application. Is that correct? You don't want to rely upon info scraped from the internet that, as pointed out, is full of the correct, the incorrect and the in-between. So, if philately wanted to utilize AI for identification and authentication purposes the effort would need to be funded firstly and vetted information supplied secondly. I assume that at some point AI can utilize what it has been given and "learn" by interpolating and extrapolating data to for instance plate a three cent Washington correctly or identify a fake overprint.

If the way I understand it above is correct I have a few questions.

Who funds the effort and is it ever really "owned?

What ROM in dollars are we talking about for such an effort?

How do you assemble all of the data in public and private hands necessary for the effort?

What of copyright issues for information critical to the AI tool?

Does AI use tools such as a spectrometer? Excuse this stupid question but doesn't AI need humans in perpetuity to be the "hands" that operate things?

My bottom-line question would be is AI ever going to be economically viable for hobby type endeavors or is it going to be a corporate/government tool, a deep pocket tool? When I say "hobby type endeavors" I mean not at the kitchen table level but at the Philatelic Foundation opinion issuing level or Scott catalog level?

I don't think AI would be helpful for me , I prefer RI for developing my forgery collection . AI sounds like a bunch of stuff to view on a computer screen . I rather accumulate the material ,study it ,lay it out and make my own pages, that is what the hobby is all about
. So I will stick with REAL INTELLIGENCE {R.I.}

Guess I don't make a lot of friends here when I tell the people here ,that I went to the A.P.S. and asked to see their FOREGY COLLECTION and they bring me a plastic bin with album pages ,notebook pages ,glassines of stamps and other envelopes of stamps all thrown in a large plastic container ,Oh well thanks for the viewing .

Sergio stuff comes up for auction and what do I see is a few binders {I was expecting 20 or 30 binders } and they write about common well known packet material in those binders and the another lot of what looks like thousand of damaged stamps as a auction lot . I want to known where is his good stuff not the stuff that I been pulling out of junk lots for 40 years . Clearly not mounted and explained . Shees --no wonder people don't like what I say .....need my fourth cup of coffee ,already wired up on 3 cups .

APS or APRL? I thought the major reference collections were housed by the latter (I've never been to Bellefonte, so I don't know the layout).

Floortrader,
You started a thread and titled it


And in your original post asked the question


and then you post this


I have no issues with anyone who has an open mind, learns about a new tool for the hobby, and decides that the new tool is not for them.

But you are posting here, a community which can also be described as "a bunch of stuff to view on a computer screen". What is the difference between using this community to learn more and using AI to learn more?
Don

Exactly right. Readers here should keep in mind that ChatGPT, Google Bard, and the other free LLM interfaces are experimental and have been trained on uncontrolled data, including data that is factually wrong. The next generation of AI systems, which I expect to be more broadly commercially available by early 2024, will use enterprise-specific, often private and curated training data to produce far more accurate domain-specific results. Imagine, for example, a GPT that has been trained only on the content of the APRL and not on WWW content. These enterprise AI systems will be expensive and law firms, accounting firms, FinServ firms, other large commercial entities with extensive document corpus, and government will adopt them first and train them on their closed corpus of enterprise documents. For end users other than employees, use may be limited or not free. Nonprofit hobby associations will have difficulty affording or deploying these systems (someone has to curate the training data) for at least 3-5 years. For example, you can contract with IBM's professional services unit to do the foregoing today, but for a corpus like APRL the cost would be $1-3 million.

I agree with everything cjpalermo1964 posted above and would add that it is easy to poke holes in the broad, free AI systems by asking very specialized questions early in their implementation curves.

AI systems are already being used in thousands of ways that each of us are already using weekly. Material procurement, shipping logistics, and hundreds of other 'behind the scenes' implementations are being used by companies like Amazon, ebay, Google and other Big Business.

On a more direct and personal level, I now use at least 15 AI enhanced software applications every week with my online philatelic development and administration activities.
Don

Sorry, but I think the AI stuff is getting this thread off-topic.

OP asked



and the AI discussion, while interesting in its own right, has relatively little to do with this. AI can enhance certain aspects of philately, but by itself cannot be a replacement for true expertization.

To try and get this thread back on track, the real question is, who will be the next Sismondos of the world, and what should the hobby do now to ensure that experts are still around 50 years from now?

IMO, we are in danger of running out of experts in far less than 50 years.

gvol21,
Sorry, but you are wrong, AI is 100% on topic and is not only part of the topic title but is also part of FloorTrader's original post when he asked,


The new experts will be using AI tools to 'find a scan on the internet'.
Don