I swear South Park's episode on how "Family Guy" come up with jokes (in 2006) is completely applicable to how LLMs come up with information 20 years later.
[
attachment=15386]
"rhythmic. scanning. pattern"
(01-05-2026, 11:43 AM)Mark Knowles Wrote: You are not allowed to view links. Register or Login to view.One thing that I assume is that these AIs have no access to the contents of the Voynich Ninja forum, so whilst they can draw from the contents of Nick Pelling's blog the information here is hidden from them.
Actually they do. And will use that contents in their hallucinations about the VMS, or any topic that is relevant to the VMS.
All the best, --stolfi
(01-05-2026, 04:54 PM)Mark Knowles Wrote: You are not allowed to view links. Register or Login to view.I wonder if there is some way of making the different AIs collaborate to produce a better result than they would individually. So, getting GPT, Gemini, Claude etc. to work together. Or maybe you can feed the thoughts of one of them into another of them and see what it comes up with and repeating that for the different AIs.
I have used ChatGPT to check the accuracy of English translations of Chinese produced by Google AI, and vice-versa. (Google Translate often produces total garbage.) I usually takes a couple of iterations for them to agree that the translation is correct.
Which of course does not mean it is
really correct, sigh...
All the best, --stolfi
(02-05-2026, 02:55 AM)Jorge_Stolfi Wrote: You are not allowed to view links. Register or Login to view. (01-05-2026, 11:43 AM)Mark Knowles Wrote: You are not allowed to view links. Register or Login to view.One thing that I assume is that these AIs have no access to the contents of the Voynich Ninja forum, so whilst they can draw from the contents of Nick Pelling's blog the information here is hidden from them.
Actually they do. And will use that contents in their hallucinations about the VMS, or any topic that is relevant to the VMS.
All the best, --stolfi
I said this as in the initial searches they didn't seem to have picked up on what had been written by myself and others on early 15th century ciphers in the Voynich Ninja forum, but they had picked up on what had been written in the cipher mysteries blog. I wonder if they rely on the current state of the website or on some previously stored version. I mean I wonder if they are up to date on their information.
(02-05-2026, 10:32 AM)Mark Knowles Wrote: You are not allowed to view links. Register or Login to view.I wonder if they rely on the current state of the website or on some previously stored version. I mean I wonder if they are up to date on their information.
Priming it to base its answers off of here, I asked it my opinion on the Chinese Theory, knowing it would have a pretty good base to work from. It did a serviceable job answering the question and included points I've made in the last week or two. However, it also ascribed to me points Jorge made in the last week or two and garbled a few things.
On a lark I asked it what my affirmative beliefs on the manuscript were, knowing I play these closer to my chest, and found out that I believe its a functional system produced with volvelles. This comes as a surprise to me!
(02-05-2026, 10:32 AM)Mark Knowles Wrote: You are not allowed to view links. Register or Login to view.I said this as in the initial searches they didn't seem to have picked up on what had been written by myself and others on early 15th century ciphers in the Voynich Ninja forum, but they had picked up on what had been written in the cipher mysteries blog. I wonder if they rely on the current state of the website or on some previously stored version. I mean I wonder if they are up to date on their information.
I believe that unlike the very first versions that had an original learning dataset, current LLMs simply conduct many genuine real-time google searches at once. So anything that can be found through google can be found by the LLM. Whether it does find it, or decides to include it in its answer, is down to the prompt, subject or simply what mood the AI is in.
Generally LLMs/google show voynich ninja results based on their titles as opposed to individual comments within threads, so the exact title of a thread influences it too, regardless of the substance of the discussion in it.
It would be nice if the AIs could do some kind of real-time OCR so that typed(or even handwritten) documents that haven't read digitised could be read and analysed.
(02-05-2026, 02:16 PM)Mark Knowles Wrote: You are not allowed to view links. Register or Login to view.It would be nice if the AIs could do some kind of real-time OCR so that typed(or even handwritten) documents that haven't read digitised could be read and analysed.
It can do it fairly well now.. Gemini is the best model, but really one page at a time.
(02-05-2026, 02:42 PM)DG97EEB Wrote: You are not allowed to view links. Register or Login to view. (02-05-2026, 02:16 PM)Mark Knowles Wrote: You are not allowed to view links. Register or Login to view.It would be nice if the AIs could do some kind of real-time OCR so that typed(or even handwritten) documents that haven't read digitised could be read and analysed.
It can do it fairly well now.. Gemini is the best model, but really one page at a time.
One page at a time is really a problem.
I didn't explain clearly. I mean as part of its search or maybe Google could automatically OCR all documents found. I don't know how far we are now from that being computationally feasible.
There are quite a lot of typed and scanned, but not digitised inventories which are a lot of effort to read manually and for which being able to search inside would be really helpful. (Obviously, being able to read handwritten documents as well would be amazing).
(02-05-2026, 03:16 PM)Mark Knowles Wrote: You are not allowed to view links. Register or Login to view. (02-05-2026, 02:42 PM)DG97EEB Wrote: You are not allowed to view links. Register or Login to view. (02-05-2026, 02:16 PM)Mark Knowles Wrote: You are not allowed to view links. Register or Login to view.It would be nice if the AIs could do some kind of real-time OCR so that typed(or even handwritten) documents that haven't read digitised could be read and analysed.
It can do it fairly well now.. Gemini is the best model, but really one page at a time.
One page at a time is really a problem.
I didn't explain clearly. I mean as part of its search or maybe Google could automatically OCR all documents found. I don't know how far we are now from that being computationally feasible.
There are quite a lot of typed and scanned, but not digitised inventories which are a lot of effort to read manually and for which being able to search inside would be really helpful. (Obviously, being able to read handwritten documents as well would be amazing).
My view is the first person to build an algorithm and robot to scan manuscripts at volume will make some serious money... There's OCR for typed documents and HTR (Handwriting text Recognition) for written. Transkribus is the gold standard, but the frontier models are catching up fast... But there are 10s of thousands of Manuscripts that no one has even looked at...