voynich.science

[Dunsel]

   

A couple months ago, I started a thread about how I was using LLM's to try to perform calculations on the Voynich. One of the problems I ran into was, it became first, a test of my patience every time it "enhanced" the results and second, a it was still a rather slow process. I would have it generate the Python. Then I'd download it, look it over, run it, get more things added, double check that, etc... 

It generated the Python code rather quickly.  And once you accepted the fact that it would sometimes cheat by running it's own regex which invariably screwed up the parsing of the transcription, or it would chop functions out of the code and run them without using the proper preceding function, you could eventually get some code written.  It was usually 5 steps forward, 4 back.  Then, I went a tad overboard one day and went back to the language I know, PHP, and generated a web page to do more than one calculation. Next thing I know, I've created this:

You are not allowed to view links. Register or Login to view.

First thing:  THIS WILL NOT RUN ON MOBILE at the moment.  I have a lot of CSS tweaking to do before it does so for now, Desktop only.
Second thing: While I have visually looked over the code, I'm not positive of the calculations behind them.  I have double checked most, but not all. I'm still finding minor miscalculations so do not treat this as Voynich gospel.

For the moment, consider it a rough draught and you are welcome to be pre-alpha testers.

Let me know if you see any bugs, errors, situations where I had moments of utter stoopidity, things you'd like to see it do, or just general comments.  

Edit: The contact form on the site is not working at the moment. If you use it, I likely didn't get the email.  You can either reply to this forum post or email me at admin@voynich.science.  Sorry for the inconvenience.

   

[DG97EEB]

A couple months ago, I started a thread about how I was using LLM's to try to perform calculations on the Voynich. One of the problems I ran into was, it became first, a test of my patience every time it "enhanced" the results and second, a it was still a rather slow process. I would have it generate the Python. Then I'd download it, look it over, run it, get more things added, double check that, etc... 

It generated the Python code rather quickly.  And once you accepted the fact that it would sometimes cheat by running it's own regex which invariably screwed up the parsing of the transcription, or it would chop functions out of the code and run them without using the proper preceding function, you could eventually get some code written.  It was usually 5 steps forward, 4 back.  Then, I went a tad overboard one day and went back to the language I know, PHP, and generated a web page to do more than one calculation. Next thing I know, I've created this:

You are not allowed to view links. Register or Login to view.

First thing:  THIS WILL NOT RUN ON MOBILE at the moment.  I have a lot of CSS tweaking to do before it does so for now, Desktop only.
Second thing: While I have visually looked over the code, I'm not positive of the calculations behind them.  I have double checked most, but not all. I'm still finding minor miscalculations so do not treat this as Voynich gospel.

For the moment, consider it a rough draught and you are welcome to be pre-alpha testers.

Let me know if you see any bugs, errors, situations where I had moments of utter stoopidity, things you'd like to see it do, or just general comments.  

Super initiative. I'm sure others will want to check the maths, but really helpful framework. Did you build it in AI Studio? Btw, it does run on mobile in "desktop mode"

You might think about adding Bowern-Gaskill's 42 tests as well, some of which you've already captured You are not allowed to view links. Register or Login to view.

[PeteClifford]

I sent you a message via the site, but wasn't sure it had sent properly, so am repeating it here.

This is a fantastic new site. Thank you, and congratulations. I just wanted to mention a couple of things, in case they were helpful.


Firstly, the webpage You are not allowed to view links. Register or Login to view. has an encoding issue:

First, you will never “see all the charts.”

I suspect the text is in a codepage other than UTF8, which is what the browser is expecting.

Secondly, I was wondering why the Zandbergen-Landini ("ZL") transcription was not an option?

Thanks again.

Pete

[Rafal]

How do you build such websites with LLM?

You tell ChatGPT "Make me a website with statistics on word frequencies in Voynich Manuscript"? 

I guess it is more complicated.

[eggyk]

How do you build such websites with LLM?

You tell ChatGPT "Make me a website with statistics on word frequencies in Voynich Manuscript"? 

I guess it is more complicated.

For what its worth, there are AI tools like Claude Code which are very impressive at generating functional code (with some necessary iteration) if the prompter is competent.

When used by a professional, it can generate effective code very quickly (the company where my partner works encourages its use if needed). I also know someone with zero coding experience that managed to get tons of working code for their project as well. 

These types of AI tools are far more useful than the hallucinating+memory loss AI types, like chatGPT.

[RobGea]

Well done Dunsel , that is really very nice.

[Dunsel]

A couple months ago, I started a thread about how I was using LLM's to try to perform calculations on the Voynich. One of the problems I ran into was, it became first, a test of my patience every time it "enhanced" the results and second, a it was still a rather slow process. I would have it generate the Python. Then I'd download it, look it over, run it, get more things added, double check that, etc... 

It generated the Python code rather quickly.  And once you accepted the fact that it would sometimes cheat by running it's own regex which invariably screwed up the parsing of the transcription, or it would chop functions out of the code and run them without using the proper preceding function, you could eventually get some code written.  It was usually 5 steps forward, 4 back.  Then, I went a tad overboard one day and went back to the language I know, PHP, and generated a web page to do more than one calculation. Next thing I know, I've created this:

You are not allowed to view links. Register or Login to view.

First thing:  THIS WILL NOT RUN ON MOBILE at the moment.  I have a lot of CSS tweaking to do before it does so for now, Desktop only.
Second thing: While I have visually looked over the code, I'm not positive of the calculations behind them.  I have double checked most, but not all. I'm still finding minor miscalculations so do not treat this as Voynich gospel.

For the moment, consider it a rough draught and you are welcome to be pre-alpha testers.

Let me know if you see any bugs, errors, situations where I had moments of utter stoopidity, things you'd like to see it do, or just general comments.  

Super initiative. I'm sure others will want to check the maths, but really helpful framework. Did you build it in AI Studio? Btw, it does run on mobile in "desktop mode"

You might think about adding Bowern-Gaskill's 42 tests as well, some of which you've already captured You are not allowed to view links. Register or Login to view.

Thanks for the link!  I just watched their video on those tests.  And, they make a convincing argument for gibberish but I'm holding out hope.  I'll look a bit deeper later and see how many of their tests I can incorporate.

---

I sent you a message via the site, but wasn't sure it had sent properly, so am repeating it here.

This is a fantastic new site. Thank you, and congratulations. I just wanted to mention a couple of things, in case they were helpful.


Firstly, the webpage You are not allowed to view links. Register or Login to view. has an encoding issue:

First, you will never “see all the charts.”

I suspect the text is in a codepage other than UTF8, which is what the browser is expecting.

Secondly, I was wondering why the Zandbergen-Landini ("ZL") transcription was not an option?

Thanks again.

Pete

Thanks!  Yep, an encoding error. I think I got them all fixed.

I'm using the Interlinear file because I've had a hell of a time finding transcriptions.  I have the original Takahashi from his site, the interlinear and the V101 which I found on his github.  If you have a link to the ZL, please give it to me and I'll see if it can be worked into the database without reloading it all.

[RobGea]

Transliteration files can be found in a table at the bottom of this page
voynich.nu > Text Analysis - Transliteration of the Text
You are not allowed to view links. Register or Login to view.

[DG97EEB]

A couple months ago, I started a thread about how I was using LLM's to try to perform calculations on the Voynich. One of the problems I ran into was, it became first, a test of my patience every time it "enhanced" the results and second, a it was still a rather slow process. I would have it generate the Python. Then I'd download it, look it over, run it, get more things added, double check that, etc... 

It generated the Python code rather quickly.  And once you accepted the fact that it would sometimes cheat by running it's own regex which invariably screwed up the parsing of the transcription, or it would chop functions out of the code and run them without using the proper preceding function, you could eventually get some code written.  It was usually 5 steps forward, 4 back.  Then, I went a tad overboard one day and went back to the language I know, PHP, and generated a web page to do more than one calculation. Next thing I know, I've created this:

You are not allowed to view links. Register or Login to view.

First thing:  THIS WILL NOT RUN ON MOBILE at the moment.  I have a lot of CSS tweaking to do before it does so for now, Desktop only.
Second thing: While I have visually looked over the code, I'm not positive of the calculations behind them.  I have double checked most, but not all. I'm still finding minor miscalculations so do not treat this as Voynich gospel.

For the moment, consider it a rough draught and you are welcome to be pre-alpha testers.

Let me know if you see any bugs, errors, situations where I had moments of utter stoopidity, things you'd like to see it do, or just general comments.  

Super initiative. I'm sure others will want to check the maths, but really helpful framework. Did you build it in AI Studio? Btw, it does run on mobile in "desktop mode"

You might think about adding Bowern-Gaskill's 42 tests as well, some of which you've already captured You are not allowed to view links. Register or Login to view.

Thanks for the link!  I just watched their video on those tests.  And, they make a convincing argument for gibberish but I'm holding out hope.  I'll look a bit deeper later and see how many of their tests I can incorporate.

---

I sent you a message via the site, but wasn't sure it had sent properly, so am repeating it here.

This is a fantastic new site. Thank you, and congratulations. I just wanted to mention a couple of things, in case they were helpful.


Firstly, the webpage You are not allowed to view links. Register or Login to view. has an encoding issue:

First, you will never “see all the charts.”

I suspect the text is in a codepage other than UTF8, which is what the browser is expecting.

Secondly, I was wondering why the Zandbergen-Landini ("ZL") transcription was not an option?

Thanks again.

Pete

Thanks!  Yep, an encoding error. I think I got them all fixed.

I'm using the Interlinear file because I've had a hell of a time finding transcriptions.  I have the original Takahashi from his site, the interlinear and the V101 which I found on his github.  If you have a link to the ZL, please give it to me and I'll see if it can be worked into the database without reloading it all.

I'm 99.9% sure it's gibberish.. the problem is it has grammar and structure and weird properties... I reckon Timm-Schinner probably came closest..

[Jorge_Stolfi]

I'm 99.9% sure it's gibberish.. the problem is it has grammar and structure and weird properties... I reckon Timm-Schinner probably came closest..

But there are many arguments against the "gibberish" theory.  Including all the statistical features that are compatible with natural languages, and unlikely to occur in text generated by any plausible gibberish generation method.  (No, Thorsten and Timm did not come close, sorry...) 

Whereas the only argument for "gibberish" seems to be "we have tried to decipher it, assuming it is an encrypted European language, and we have failed so far".  Which has a logical problem...

All the best, --stolfi