Engineering your own voynich

[nablator]

My aim was to make a text that is similar to voynichese by statistical properties. 

For that, you would need to reduce the conditional entropy h2.

[Mark Knowles]

Hi David!
I need to examine your results. They are very interesting, especially as they are not the original text.

I suspected that ChatGPT had not found the correct solution, however despite this I would not be shocked if one of its descendents decipher the Voynich manuscript in the next 10 years.

[Searcher]

My aim was to make a text that is similar to voynichese by statistical properties. 

For that, you would need to reduce the conditional entropy h2.

Hi nablator! Could I ask you to calculate that for my enciphered text? 
I could try to calculate it myself, but I'm not sure which formula must be used.

[ReneZ]

Some of them are known Latin aphorisms, others may be translated, like #5:
Aller Anfang ist schwer.

ChatGPT often generates bullshit when it doesn't have a clue.

That could be interpreted as a case of 'gnothi seauton' by ChatGPT.

[nablator]

Hi nablator! Could I ask you to calculate that for my enciphered text?

Hi Searcher,
h2 = 2.435

Byatan's ciphertext (in this thread) has nearly the same value of h2, not close to the anomalous values of Voynichese: I just checked two samples of lowercase EVA transliteration of the same size, got 1.3 and 1.9.

[Mark Knowles]

ChatGPT often generates bullshit when it doesn't have a clue.

Despite getting the decipherment completely wrong, it appears, I am actually more impressed at ChatGPT's attempt at decipherment here than I have been at any attempt to decipher the Voynich of a human so far.

In fact nobody seems to have tried to crack Voynichese as if it were a cipher. The focus seem to be on compiling statistics, which whilst certainly useful is dancing around the problem. It sometimes feels like many researchers have never tried to break a cipher and so would have no idea where to start with trying to decipher Voynichese. Sometimes one learns by trying and failing and then trying again.

I think with a few more iterations of these Large Language Models we might well end up with something that could take on the Voynich and win. I hope this is not the case and it will be disappointing if the victory is stolen from humans by a machine.

[Searcher]

Hi Searcher,
h2 = 2.435
Byatan's ciphertext (in this thread) has nearly the same value of h2, not close to the anomalous values of Voynichese: I just checked two samples of lowercase EVA transliteration of the same size, got 1.3 and 1.9.

But it is not so bad, right? It gives h3 (3.94-2.44) = 1.5. I need to think a little how to reconstruct the cipher to "voynichesly" improve it more.

[Searcher]

"est res iudicata saepe sequitur quod initium coeptum est."
"in rebus difficilibus opportunitas est saepe nascentium."
"adversa fortuna est opportunitas maxime cognoscendae virtutis."
"qui amat veritatem, amat etiam iustitiam."
"omne initium difficile est."
"etiam victis victoria praesto est."
"qui sine amore vivit, morbus sibi est."
"atque in eo est quod amamus."
"quod bene dictum est ab uno, non reprobo a plures."
"amor omnia vincit et nos cedamus amori."
"fortuna caeca est, sapientia vero numquam."

Unfortunately or not, I've find no coincidences.

[Mark Knowles]

Unfortunately or not, I've find no coincidences.

What I would be interested in is whether ChatGPT's solution was logically consistent i.e. does it work? Or is it fundamentally flawed? I haven't checked this.

[mja0075]

What I would be interested in is whether ChatGPT's solution was logically consistent i.e. does it work? Or is it fundamentally flawed? I haven't checked this.

It doesn't work, the mappings it gives as examples aren't used when it's "deciphiring". 
If you follow up with it and ask it to give a full algorithm or key, it will apologize and say it made a mistake.