Will AI take over Discussion Boards?

Typical OF AI response: [Picture of Cats]

So this is an interesting snake-eating-its-tail problem. Current LLM models can only 'learn'[0] by reading human written text. However, when an LLM trains on text that is LLM generated, it reduces its 'knowledge' [0].

Up until ~2023, they were able to train on non-LLM data just in the general internet. Anyone who was a web host at the time probably saw a TON of scrapers going around with no real direction as all the AI companies were just trying to consume ALL of the internet. However, once OpenAI spilled the beans by publically releasing ChatGPT, their gravy train was over!

The internet, almost overnight, became significantly AI generated. You can see this by googling things and realizing that 90% of google results for some things are just the same regurgitated garbage over and over, as LLM generated websites are 'it'. So, AI doesn't really have a PLACE to train anymore, or at least, training data is VERY tough to find, because again, consuming LLM generated data harms their training. AND, if they can't keep consuming data, they will be 'stuck' historically at whatever time they stopped consuming! For a while, ChatGPT froze in late 2023 for exactly this reason.

IN this modern internet, curated 'human only' content is getting more and more valuable. There is a reason the AI companies are doing everything they can to get into active 'human only' communities, like reddit. They want to be able to consume that data!

All that to say: Forums like this unbelievably valuable to LLM models, particularly with the curation that mods do to keep AI out. And if a web forum gets overrun by LLMs and AI companies don't recognize it, they can effectively self immolate 😀

As far as the value of an AI forum: There is none. The purpose of a forum is for various people to learn and interact with eachother. AI doesn't benefit from that, and if you are just looking for an answer to something from an AI, just ask your model directly.

[0] LLMs actually can't 'learn' because they don't 'know'. They are just really fancy random-number generators with large state graphs.

Slightly off topic, my brother-in-law is a professional writer who recently interviewed for a new position. The entire interview was with AI. Other writers had experienced the same. He thinks the interviews were ploys to teach AI how to better imitate human writers.

If true, very distasteful.

This is honestly a horrifying reality. A lot of creative types are seeing exactly this. On top of that, there are companies 'training' for AI interviewing software that are getting folks to interview just to train/test their model with no job at the end.

Last fall I had an idea for a book with a tentative title 'A child's history of AI.' Which is a play on one of Dickens books 'A child's history of England.'

Most of my texts and references are 40 to 50 years out of date. Which is fine since it is supposed to be a 'history.' I tracked down some old programs, what really do not have anything to do with the modern hype nonsense. One was a game called Oregon Trail. Another called Eliza/Doctor. Others were programs what used a time share BASIC statement called MAT. That was removed from personal computer BASIC as it requires efficient floating point abstractions.

I spent time over the weekend working on the outline. Most of what came out from my own point of view related to the early works that created the internal tables, Factorial, Taylor Series, Chebyshev polynomials etc. The other part of the outline came out as biased autobiographical examples.

I have yet to actually see what ChatGTP is about. Or even see what writing prompts entails. Now this seems like a good idea to avoid that nonsense and time sink. I am sure someone could prompt such a thing into one of these systems. Such would probably come up with results that were engaging. I can do the same thing without the expense of having to create yet another set of social media accounts and tokens. (Google search is bad enough)


One sideline of these text books is a strange work nicknamed the 'Voynich Manuscript.' Which was used in an old textbook on letter frequency pairs. This is cataloged as Beinecke MS408. Since the text was written by a Yale prof, it made sense as an exercise for the student.

The manuscript only has about 100 years of physical providence. Internally it seems to date to the late 15th century. Most of the 20th century research linked it to famous alchemists and philosophers in 16th/17th century Prague. The seller who sold it to the 1920s US code-breakers was a Russian rare book dealer, who's wife's works were popular in Russia. Note that these code-breakers worked for the US treasury and were spying on rum running networks in Canada and the southern Americas.

The history of this document is complex. When found it was jumped upon by researchers who were trying to find Binary codes in Shakespeare. These names were the top code breakers of WWI and WWII. A few of them were still around in the late 1990s where they had a simple (non Usenet) mailing list. An informal study group so to speak. Some were cryptographers others linguist.

It was quite fascinating. I used the data and created a font while at Apple since it was a way to generate pages of nonsense test, and test the font mechanics with an unusual set of glyph that are completely independent of any other writing systems. Testing of the printers mechanical systems involved sending tons of paper through them. Not unlike the prayer machines in 'The Handmaiden's tale." What was cuious was that none of the online researchers could ever agree on how the glyphs mapped to the Roman character set.

Supposedly this was the first use of one of the first IBM computers constructed in the 1940s after WWII. I used Friedman's mapping, which was considered obsolete by the 1990s.

Ironically I noted that one could abstractly date things by how hair is styled. The figures, which are mostly female are drawn Nude. Still they have hair. My theory is that one can date films and such to a decade given how hair is styled. Even the costume dramas. This is hard to see at the time of creation. It does become more obvious as the years pass. That got a lot of attention, and I was interviewed for a PBS style documentary.

With the advent of the net, the mailing list got taken over by less serious researchers. Someone as a Joke pointed the Necronomicon at the MSS 408 catalog number. This had the effect of removing all seriousness from any study of this work.

I am still subscribed to the dead zombie mailing list. Many of us who have been researching it for the last few decades feel it is a forgery. Since it is associated with intelligence and disinformation agencies it is fodder for the conspiracy theorist.

Recently there has been activity on that zombie mailing list. There is a seminar planned. I think for this weekend. I guess there is a new generation who are creating papers to present on it. Most of the titles however re hash the same stuff, using the same number theories what generate the same illogical results.

One abstract though was interesting. This was to use AI to look at it. The AI however hit the Turing halting problem. Would claim to have found a solution, but more information was needed. That possibly it was illoian/Greek transcriptions of a lost pre Colombian language. Not much better than the 60 year old textbook which said it the closest statistical match was Hawaiian.

The Twain/Harte devils dictionary probably has the best sum up. There are lies, damn lies, and statistics.

I do think AI is some sort of bubble. That the system is not sustainable.

In my semi auto graphical outline I noted that the early personal computing days (Homebrew Computer club era.) that the new stuff felt like a blank sheet of paper waiting to be written. AI on the other hand feels like a Graffiti covered wall with no room for more expression.

I also notice that what I have written here is actually a bit of a chapter or section of this potential book. That by posting it here I give up certain rights to this content. I suppose I can formally claim copyright to it. As in that I am sharing this as a quoted fair use of an unpublished work. Or in other words, that no one else can claim copyright to this post and this post only.

I don't know about artificial intelligence in the forum, but maybe OF could have "artificially idiotic" answers to some common questions (Does this look too small on my wrist? What do I do about this scuff on my clasp? Why is my brand new luxury watch running 0.5s slow per day? etc.).

No algorithm involved, just a simple deterministic look-up table.

Any post from a new member that includes the words "this beauty" should be answered, "Unfortunately the dial has been poorly repainted."