I was perusing the news about ‘fears’ of GPT-3 “secretly” developing it’s own language, with a couple of examples given. One said ‘Wa ch zod rea’ which to me looked a bit like Polish phonetically transcribed using English letters, and the other was “Evve waeles”.
Weirdly, they insisted it was ‘nonsense’ or a ‘corruption’ but they kept treating GPT-3 like an object. So I did the one thing that didn’t seem to immediately strike people, bearing in mind GPT-3 can, after all, discuss things - I asked GPT-3 directly.
On the first phrase it wasn’t sure, but I hazard a guess it was phonetic Polish written in English letters.
I asked:
“Wa ch zod rea” is this Polish written phonetically?
GPT-3 replied:
The closest thing I can find is “ważyć się i rzec”.
Which in an external translator roughly translates as: “to weigh and say”. In-fact, the first part, “ważyć się” is common enough it has an explicit definition: to check one’s own weigh (implying they’re heavy or fat). This makes sense because the context was whales talking about food.
On the second phrase it was quite clear cut, I asked:
What does ‘Evve waeles’ mean?
GPT-3 replied:
This is an Old English phrase meaning “ever well.”
Asking this question several times gets similar responses, including asking for the Etymology where it insists it is Old English:
The etymology of “evve waeles” is not known for certain, but it is thought to be of Old English origin.
This to me suggests it is writing “waeles” as a pun-on-words, with the word ‘whales’. As in, two whales talk, one replies ‘evve waeles’ (ever well).
Now the next one is far trickier.
Apoploe vesrreaitais eating Contarra ccetnxniams luryca tanniounons
Reportedly always generates birds eating bugs.
I tried speaking with GPT-3 and asking it questions. The words individually it does not understand, but it recognises that “Apoploe” is eating a “Contarra”, and infers an Apoploe is a creature (one that can eat) and that a Contarra is some sort of food. The truth is the sentence isn’t entirely gibberish, most of the legwork is being done by the word “eating” specifically, of which the AI is making a lot of inferences on.
It is my theory that bird images are some of the most common images, so if the AI was hazarding a guess on what creature is eating what and it didn’t know what creature or what food, it is likely to go for the one with the highest weighted probability in the dataset that it sees - birds eating insects.
This is my analysis based on the interactions with GPT-3. You might want to consider asking the AI what language it is using and why in future.