I can’t help but wonder if we were able to feed a lot of whale communications into LLMs with this new found layer of information if we would be able to translate someday.
That’s the kind of task I really want to see LLMs used for - not just to replicate things humans can already do, but to explore things that humans haven’t figured out yet. Maybe we don’t know enough to pick up on the rhyme and reason of whale communication yet, but brute-forcing it with an AI seems like a great way to get a fresh perspective.
https://www.earthspecies.org/The motivating intuition was that modern machine learning can build powerful semantic representations of language which we can use to unlock communication with other species.
Well, lets get an LLM to translate one of the not understood human languages first.
Could we build a chat GPT for whales to use, with this approach, sure. Does that get us any closer to translating it. NO, because without shared context to start building connections and relationships its highly unlikely that were going to find commonality.
>Well, lets get an LLM to translate one of not understood human languages first.
You can do both. Models trained to receive and predict audio tokens alongside text are coming. You could just add such data as part of training.
>Does that get us any closer to translating it. NO, because without shared context to start building connections and relationships its highly unlikely that were going to find commonality.
If you trained it alongside human languages then shared context(if available) will be figured out by the model. They'll coalesce in the same shared space the way they do for human languages. Enough to translate without examples instead of just speak? Perhaps - it's not like there are examples in the training set for every lang to lang combination modern models are capable of translating.
the sounds for man, and the word man can be mapped. The concept of man can be mapped from English to Chinese. Given enough of these clues, and a large enough data set it's not shocking that models get decent at inference between languages.
But whales? What we are (probably) proposing is putting two languages side by side with NO context and saying "figure it out".
couldn’t we approach it like (my understanding of) code breaking worked back in ww2? meta observations of recurrent symbols, like time, start/end message, etc reveal certain consistencies. then we fill in gaps through possible coherences.
while ww2 code breaking contained other human contexts, perhaps like-context observations could allow for certain categories that might reveal linguistic insights of whales. like whale sounds observed consistently near boats, divers, danger, eating, play, etc.
no clue about how LLM/AI systems process meta contextual training data inputs (is this even a valid phrase about AI?! haha) but i hope whale linguists have considered such an approach.
>the sounds for man, and the word man can be mapped. The concept of man can be mapped from English to Chinese. Given enough of these clues, and a large enough data set it's not shocking that models get decent at inference between languages.
1. Many such concepts cannot be directly mapped between different languages especially with distant language pairs.
2. Language Models and Image models even if both are only trained on their respective modality learn structural representations so similar, you can connect them with a simple linear layer. That's it. Entirely different modalitieshttps://arxiv.org/abs/2209.15162
I think you are severely underestimating the extent to which representations can group in neural networks.
They're mammals, they're social animals, they see. You're assuming a level of alienness we don't have the knowledge or understanding to truly ascertain.
given unrestricted compute, could an AI attempt to decipher a completely unknown language through interpreting O(n) brute iterations and reporting versions that are coherent to human heuristics?
There are uncountably many “valid” interpretations by this approach. With scant source material there is no way to verify the guessed translation as correct or not.
It’s like “guessing” the encryption key of a one-time pad.
I was under the impression that the reason why we don't understand the languages that we don't understand is because we lack sufficient records of them.
The technique in that paper, is likely not generalizable.
It works for EN/FR, less well for EN/DE... because English has a huge overlap with these two (more French than German) for its word by word translation. With something like Chinese, and enough time it might be able to guess its way in due to overlaps in the training data.
If you took two completely random corpus's from two languages, and jammed them into this model and handed that system to someone who only knew one language, the approach would fall flat on its face.
I vaguely remember reading about a research effort into this earlier this year, and I'm positive it wasn't what sp332 linked to. Would be glad if I can find the link.
I mean since the glorified autocomplete bots we call LLMs can't actually understand language, all we have to do is get enough samples of whale language and they can easily communicate with them.
It's not that simple. LLMs are great for predicting next word, but this needs a more unsupervised approach.
Ultimately, you need to know the environment of whales. Sure, you know the next probable token, and even relationships between tokens, but how do they relate to what whales are doing?, in other words what does it mean?.
It reminds me of the great quote by Feynman: "look at the bird"[1]. You can know the name of a bird in many different languages, but that tells you nothing about it. If you want to understand birds, you need to look at it and see what it's doing.
If we want to communicate with another species, we need to understand what they're doing, why they're communicating, and from there we can begin to infer what they're communicating. It could be simply a callsign/identification/name for each individual, it could be communicating you've found a place with food, it could be communicating about predators, who knows. There could be all sorts of interesting variation and nuance to their calls. But ultimately it's necessary to look (in the general sense of having information about what it's doing -- could be through all kinds of sensors!) at the ~~bird~~ whale :)
[1] Feynman tells the as a lesson from his father. (that you can know their names and still know nothing about the birds themselves).
This headline feels slightly misleading to me. As far as I can tell, this research shows that sperm whale clicks have spectral properties which are analogous in some ways to human vowels. In particular, they exhibit formants (essentially, amplified frequency bands), and analysing them reveals that these whales produce two different kinds of clicks. While fascinating in their own right, I feel it isn’t entirely fair to call these ‘vowels and diphthongs’ without qualification: these seem to admit a lot less variation than do human vowels.
They gave a list of characteristics that differentiate one human vowel sound from another, and showed that the whale clicks have all of the same types of differentiators.
31 comments
[ 4.2 ms ] story [ 68.6 ms ] threadCould we build a chat GPT for whales to use, with this approach, sure. Does that get us any closer to translating it. NO, because without shared context to start building connections and relationships its highly unlikely that were going to find commonality.
You can do both. Models trained to receive and predict audio tokens alongside text are coming. You could just add such data as part of training.
>Does that get us any closer to translating it. NO, because without shared context to start building connections and relationships its highly unlikely that were going to find commonality.
If you trained it alongside human languages then shared context(if available) will be figured out by the model. They'll coalesce in the same shared space the way they do for human languages. Enough to translate without examples instead of just speak? Perhaps - it's not like there are examples in the training set for every lang to lang combination modern models are capable of translating.
But whales? What we are (probably) proposing is putting two languages side by side with NO context and saying "figure it out".
Here is a great example of people and language and perception: https://news.mit.edu/2023/how-blue-and-green-appeared-langua...
Im fairly confident that unless we give an LLM a LOT of context between data sets to work with that it won't make much progress at all.
while ww2 code breaking contained other human contexts, perhaps like-context observations could allow for certain categories that might reveal linguistic insights of whales. like whale sounds observed consistently near boats, divers, danger, eating, play, etc.
no clue about how LLM/AI systems process meta contextual training data inputs (is this even a valid phrase about AI?! haha) but i hope whale linguists have considered such an approach.
1. Many such concepts cannot be directly mapped between different languages especially with distant language pairs.
2. Language Models and Image models even if both are only trained on their respective modality learn structural representations so similar, you can connect them with a simple linear layer. That's it. Entirely different modalities https://arxiv.org/abs/2209.15162
https://arxiv.org/abs/2304.08485
I think you are severely underestimating the extent to which representations can group in neural networks.
They're mammals, they're social animals, they see. You're assuming a level of alienness we don't have the knowledge or understanding to truly ascertain.
One plank of wood is not enough to determine solely if it was intended to be part of a house or part of a boat
It’s like “guessing” the encryption key of a one-time pad.
https://arxiv.org/abs/1711.00043
It works for EN/FR, less well for EN/DE... because English has a huge overlap with these two (more French than German) for its word by word translation. With something like Chinese, and enough time it might be able to guess its way in due to overlaps in the training data.
If you took two completely random corpus's from two languages, and jammed them into this model and handed that system to someone who only knew one language, the approach would fall flat on its face.
Humans: “We come in peace. We seek to understand and communicate with you.”
No response.
Humans: “We repeat. We come in peace…”
WhaleGPT: “too late”
Ultimately, you need to know the environment of whales. Sure, you know the next probable token, and even relationships between tokens, but how do they relate to what whales are doing?, in other words what does it mean?.
It reminds me of the great quote by Feynman: "look at the bird"[1]. You can know the name of a bird in many different languages, but that tells you nothing about it. If you want to understand birds, you need to look at it and see what it's doing.
If we want to communicate with another species, we need to understand what they're doing, why they're communicating, and from there we can begin to infer what they're communicating. It could be simply a callsign/identification/name for each individual, it could be communicating you've found a place with food, it could be communicating about predators, who knows. There could be all sorts of interesting variation and nuance to their calls. But ultimately it's necessary to look (in the general sense of having information about what it's doing -- could be through all kinds of sensors!) at the ~~bird~~ whale :)
[1] Feynman tells the as a lesson from his father. (that you can know their names and still know nothing about the birds themselves).
https://www.youtube.com/watch?v=ga_7j72CVlc
Bonus related video!: https://www.youtube.com/watch?v=M1TiXLGqlM4