Hi Psillycyber,
these language models are brilliant at handling languages (English in particular, in my experience), this can easily mislead people into thinking they are smart, since, from our point of view, language and intelligence are strictly associated. As you say, this is not the case: these systems have a very limited idea of reality and facts. In particular, they are not aware of how they function and asking them to explain their reasoning can be fun but is a waste of time.
About the "continuation experiment". AI input from the linked web page (the end of f2r):
AI continuation:
This shows that Bing AI has accessed a transcription of the Voynich manuscript. Since the training process of these tools is so opaque, I wonder if it saw Voynichese during language training, or if it just grabbed the transliteration from the web to reply the question.
The continuation is entirely made up of words from Takahashi's transliteration. While about 14% of Voynich tokens are hapax legomena (appear only once in the whole text) here the two rarest words appear twice in TT's translitaration (olsaiin yor). The continuation includes 56 tokens: in actual Voynichese, an average of 7 of these would be unique word-types that don't appear elsewhere.
Maybe in the future these models will be able to extract structured information from the texts they examine. But applying "large" language models to Voynichese has the problem that Voynichese is anything but a "large" corpus: its ~38K tokens are only about twice as large the context managed by GPT-4; and of course this is not even a single uniform language, but the result of a gradual drift from Currier-A to Currier-B; the original order of the pages is lost, so the drifting process cannot be analysed with any certainty about its actual chronology. Lastly, the greatest uncertainty in Voynich transliterations is word spacing: the Zandbergen-Landini transliteration marks as uncertain more than 5% of word separators, since each uncertainty affects two words, about 10% of Voynichese tokens cannot be reliably read. If I understand correctly, these language models strongly depend on words as atomic entities: I guess that the problem of spaces would have a huge impact on their learning.
But, despite all these problems, I am sure that the field will evolve quickly in the next few years and it's quite possible that AI efforts specifically developed for Voynich research will result in actual advances.
these language models are brilliant at handling languages (English in particular, in my experience), this can easily mislead people into thinking they are smart, since, from our point of view, language and intelligence are strictly associated. As you say, this is not the case: these systems have a very limited idea of reality and facts. In particular, they are not aware of how they function and asking them to explain their reasoning can be fun but is a waste of time.
About the "continuation experiment". AI input from the linked web page (the end of f2r):
Quote:qoky.cholaiin.shol.sheky.daiin.cthey.keol.saiin.saiin.ychain.dal.chy.dalor.shan.dan.olsaiin.sheey.ckhor.okol.chy.chor.cthor.yor.an.chan.saiin.chety.chyky.sal.sho.ykeey.chey.daiin.chcthy
AI continuation:
Quote:shy.ckhy.ckhy.ckhy.qokal.qokal.qokal.qokal.qokal.qokal.qokal.qokal.qokal.qokal.qokal.qokal.qokal.qokal.qokal.qokal.oky.cholaiin.shol.sheky.daiin.cthey.keol.saiin.saiin.ychain.dal.chy.dalor.shan.dan.olsaiin.sheey.ckhor.okol.chy.chor.cthor.yor.an.chan.saiin.chety.chyky.sal.sho.ykeey.chey.daiin.chcthy
This shows that Bing AI has accessed a transcription of the Voynich manuscript. Since the training process of these tools is so opaque, I wonder if it saw Voynichese during language training, or if it just grabbed the transliteration from the web to reply the question.
The continuation is entirely made up of words from Takahashi's transliteration. While about 14% of Voynich tokens are hapax legomena (appear only once in the whole text) here the two rarest words appear twice in TT's translitaration (olsaiin yor). The continuation includes 56 tokens: in actual Voynichese, an average of 7 of these would be unique word-types that don't appear elsewhere.
Maybe in the future these models will be able to extract structured information from the texts they examine. But applying "large" language models to Voynichese has the problem that Voynichese is anything but a "large" corpus: its ~38K tokens are only about twice as large the context managed by GPT-4; and of course this is not even a single uniform language, but the result of a gradual drift from Currier-A to Currier-B; the original order of the pages is lost, so the drifting process cannot be analysed with any certainty about its actual chronology. Lastly, the greatest uncertainty in Voynich transliterations is word spacing: the Zandbergen-Landini transliteration marks as uncertain more than 5% of word separators, since each uncertainty affects two words, about 10% of Voynichese tokens cannot be reliably read. If I understand correctly, these language models strongly depend on words as atomic entities: I guess that the problem of spaces would have a huge impact on their learning.
But, despite all these problems, I am sure that the field will evolve quickly in the next few years and it's quite possible that AI efforts specifically developed for Voynich research will result in actual advances.