Thank you, Marco and Rene, for the feedback. I still feel it's about time I started thinking about plaintext languages.
Let me try to explain.
I know that the probability of myself finding the solution to the Voynich cipher (if it is a cipher) is extremely small, since the road is well trodden and I'm not a codebreaker or a historian. However, the probability is not exactly zero, as long as I'm trying.
I've been doing some hobby research of the Voynich Manuscript for a little over than 3 years now, in my free time just for fun. Last year I started the Voynich marathon to speed up my investigation, mainly because I thought it would be a shame if the Voynich Manuscript is first decoded by an AI. There is no deep logic in this, this is just the way I feel about it, and this is why I don't use AI (any black box solution, the workings of which I don't fully understand) in my attempts of cracking the manuscript.
I strongly believe, that AIs with very high probability will surpass the average human on most mental tasks in very short future, and it is likely they will surpass the best of humankind on almost any task not very much longer afterwards.
However, there is a very obvious problem with assessing the actual capabilities of modern AIs: you never know what they have seen during the training. Modern AIs can easily process texts in hundreds of languages and their training database consists of billions of pages of text, collected from all over the Internet. The AIs also are not bad at bluntly transferring some ideas between different domains (say, applying chemistry to cooking, AIs are not consistent, but sometimes feel as if they were creative).
Suppose, an AI comes up with an interesting novel result in some branch of maths, you can never be sure whether its discovery is original or a retelling of some post on a related topic from an obscure math forum of a local university in some province of China. It's very hard to find a task for which we can be sure that the AI has never seen it being solved by a human before. AIs cannot reliably produce references for the data they used, even when they pretend they do. This is simply technically impossible for a modern mainstream AI to identify the sources of its reasoning, given the way the training is performed.
And then it occurred to me that the Voynich Manuscript itself is a very good task to test AIs on, provided there is a known solution and parts of the solution are not in the training data. The task is certainly not simple, the manuscript itself is freely available as the full set of images, so it's a self-contained hard task. Also, the Voynich manuscript looks like a very well isolated task, basically you either solve the Voynich manuscript or do something else entirely, so it's unlikely there are ideas that can be copy-pasted from some other field. So I thought about maybe turning my irrational dislike of the idea of AI solving the manuscript into something potentially useful.
It was about the same time (December 2023), that when preparing my fourth article in the Voynich marathon I stumbled upon an idea that I hadn't seen before and that seemed to explain properties of the manuscript quite nicely and looked like something that could lead to a solution. I wanted to share it, but then I thought maybe it would indeed be useful if I tracked this idea to the very end without leaving any breadcrumbs on the internet. In an unlikely case that this turns out to be the solution, not only will have it, but we also would have a very clear cut off date for when the specific information about the solution appeared in the training data. This looks like a perfect benchmark for AIs. I'm not sure if someone would actually spend resources on this, but in case the manuscript is actually solved, there would be some free publicity and the idea could get some traction. And in case it's not solved, the way in which it wasn't solved is not that important anyway.
So, I stopped publishing Voynich marathon articles and in my free time continued to pursue the solution. Considering your comparison with Australia to UK journey, I believe that I currently could be somewhere near the coast of Portugal, and it's already appropriate to think whether to turn into the English channel or go up and sail around Ireland. That is, if I'm right. If I'm wrong, then I very well could be somewhere near San Diego, and I have no hard feelings, it has been a pleasant journey anyway.
Just to make it clear, I'm still not sure whether I'm on the right path, it all looks very plausible, but I have no definite answer to any questions yet, including any specific words, images or marginalia of the manuscript.
Let me try to explain.
I know that the probability of myself finding the solution to the Voynich cipher (if it is a cipher) is extremely small, since the road is well trodden and I'm not a codebreaker or a historian. However, the probability is not exactly zero, as long as I'm trying.
I've been doing some hobby research of the Voynich Manuscript for a little over than 3 years now, in my free time just for fun. Last year I started the Voynich marathon to speed up my investigation, mainly because I thought it would be a shame if the Voynich Manuscript is first decoded by an AI. There is no deep logic in this, this is just the way I feel about it, and this is why I don't use AI (any black box solution, the workings of which I don't fully understand) in my attempts of cracking the manuscript.
I strongly believe, that AIs with very high probability will surpass the average human on most mental tasks in very short future, and it is likely they will surpass the best of humankind on almost any task not very much longer afterwards.
However, there is a very obvious problem with assessing the actual capabilities of modern AIs: you never know what they have seen during the training. Modern AIs can easily process texts in hundreds of languages and their training database consists of billions of pages of text, collected from all over the Internet. The AIs also are not bad at bluntly transferring some ideas between different domains (say, applying chemistry to cooking, AIs are not consistent, but sometimes feel as if they were creative).
Suppose, an AI comes up with an interesting novel result in some branch of maths, you can never be sure whether its discovery is original or a retelling of some post on a related topic from an obscure math forum of a local university in some province of China. It's very hard to find a task for which we can be sure that the AI has never seen it being solved by a human before. AIs cannot reliably produce references for the data they used, even when they pretend they do. This is simply technically impossible for a modern mainstream AI to identify the sources of its reasoning, given the way the training is performed.
And then it occurred to me that the Voynich Manuscript itself is a very good task to test AIs on, provided there is a known solution and parts of the solution are not in the training data. The task is certainly not simple, the manuscript itself is freely available as the full set of images, so it's a self-contained hard task. Also, the Voynich manuscript looks like a very well isolated task, basically you either solve the Voynich manuscript or do something else entirely, so it's unlikely there are ideas that can be copy-pasted from some other field. So I thought about maybe turning my irrational dislike of the idea of AI solving the manuscript into something potentially useful.
It was about the same time (December 2023), that when preparing my fourth article in the Voynich marathon I stumbled upon an idea that I hadn't seen before and that seemed to explain properties of the manuscript quite nicely and looked like something that could lead to a solution. I wanted to share it, but then I thought maybe it would indeed be useful if I tracked this idea to the very end without leaving any breadcrumbs on the internet. In an unlikely case that this turns out to be the solution, not only will have it, but we also would have a very clear cut off date for when the specific information about the solution appeared in the training data. This looks like a perfect benchmark for AIs. I'm not sure if someone would actually spend resources on this, but in case the manuscript is actually solved, there would be some free publicity and the idea could get some traction. And in case it's not solved, the way in which it wasn't solved is not that important anyway.
So, I stopped publishing Voynich marathon articles and in my free time continued to pursue the solution. Considering your comparison with Australia to UK journey, I believe that I currently could be somewhere near the coast of Portugal, and it's already appropriate to think whether to turn into the English channel or go up and sail around Ireland. That is, if I'm right. If I'm wrong, then I very well could be somewhere near San Diego, and I have no hard feelings, it has been a pleasant journey anyway.
Just to make it clear, I'm still not sure whether I'm on the right path, it all looks very plausible, but I have no definite answer to any questions yet, including any specific words, images or marginalia of the manuscript.