Project CETI - Using AI To Understand What Whales Are Saying To Each Other
Cormorant Brands and Georgetown Bottling are both SPCs, “Special Purpose Corporations”. What that means, in our case, is that we’re permanently and irreversibly committed to giving ten percent of our net income to projects that study or protect whales and dolphins.
We’ve just made our first contribution to a whale-related project. It’s tiny, because we’re still at the very beginning. There isn’t much, yet, in the way of net income, for us to hand out ten percent of. But there will be some day. So we might as well get started now, even if we can’t do much at first.
What I want to do in this first blog post is say a little bit about why we chose this particular thing, this unbelievable, incredible, amazing thing, as the very first thing to give money to.
So what is it?
Its name is Project CETI.
That name is a pun on the name of another project - Project SETI, the organization that Searches for ExtraTerrestrial Intelligence. The CETI in the name of Project CETI is short for “cetacean”.
Project SETI uses radio telescopes to look for intelligent life in outer space.
But of course, we may have intelligent aliens right here on Earth. We know that some species of whales and dolphins have languages, of a kind. Culturally transmitted communication systems. We don’t actually know how complicated they are. They may be rather simple, or they may be densely packed with meaning. We can’t tell.
We know that different communities of orcas and sperm whales have unique dialects. We know a lot about whale communication systems. We just don’t have any idea at all of what they’re actually saying.
A few years ago, some people associated with Project SETI began to wonder whether we might not be able to use machine learning to decipher whale languages. Alien languages, when we find space aliens, will be very difficult to translate. This could be practice for that. We could learn how to talk to the space aliens by starting with the sea aliens.
By now everyone knows that AIs have gotten a lot better than they used to be at translating human languages and generating meaningful text. If the machines could learn to translate from English to French, these people asked themselves, why couldn’t they learn to translate from Whale to English? Could AI be the key to unlocking the mysterious languages of the whales?
It’s easy to come up with reasons why that might be difficult. Or even impossible.
Whale languages, if they even have languages in the human sense, are probably more different from any human language than any human language is from any other human language. Human cultures tend to have names for a lot of the same things. With whales, there won’t be that much overlap in the things we have names for.
Human languages seem very different from each other, but the fundamental way they are organized is actually pretty similar. The way whale languages are organized might be completely different. So finding a meaningful translation will be a lot harder.
We have a lot less data to work with. AIs that work with human language have a huge amount of text to train on. Everything ever written. But for any given species of whale, the number of sounds we’ve recorded so far is comparatively tiny.
We also have a lot less information about what the whales are likely to be talking about. They may be reciting epic poetry. I don’t think that’s what’s happening, but it’s a worst-case scenario we’ll have to think about. In that case, the context in which the sounds were produced would be completely irrelevant to their meaning. .
Okay, so… It’s going to be difficult. It may be impossible. Does that mean we should quit without even trying? No. Because it also might be possible. And we are us, now! The world of 2023 is very powerful. Difficult things, we do at once. This might just take a little longer.
In this first post, I don’t want to get down into the weeds of how machine learning works, or all the different tricky ways you might try to approach this problem with AI. I’ll talk about that later. For now, all I want to say is, for every difficulty we can currently see, it seems possible to imagine a solution. That might work. Maybe.
It’s going to take some effort. It’s going to take some ingenuity. It’s going to take some perseverance. It’s going to take, actually, some money. Kind of a lot of money. Possibly over a long period of time. But we’ve solved difficult problems before. And this is a nice one. Right on the edge of what’s possible.
So… assuming it even makes sense to do this in the first place… exactly what kind of whales are we talking about, here?
That is a very interesting question.
The particular species of whale the folks at Project CETI have chosen to focus on first is the sperm whale.
Sperm whales don’t seem to converse much when they’re down in the depths a mile below the surface of the ocean, hunting squid. But when they’re resting on the surface, or just starting down, they often exchange complicated series of clicks.
Saying “clicks” makes it seem like these sounds are simple. But actually we humans only perceive them in that way because the nuances are hard for us to hear.We’ve already used AI to look at them a little bit, one of the papers describing that is linked below, and apparently they’re a lot more complex than they seem.
Sperm whales are an obviously interesting choice, because they have the largest, most complicated brains of any creature on Earth. Their brains are five or six times the size of a human brain, and seem to have the same kind of complexity.
We’re pretty sure that they have some sort of culture, because different communities of sperm whales speak recognizably different dialects of their click language.
The difference can’t be genetic, because members of the different communities breed with each other, but the dialects remain distinct. This makes cultural transmission - learning from other whales in your group - the only way the different dialects could be transmitted from generation to generation.
Project CETI’s plan, which they describe in quite a lot of detail in one of the papers I’ve linked to below, is basically just to non-intrusively gather a lot more data on sperm whale talk, and what the sperm whales are doing while they’re talking. A lot more. And then just basically AI the heck out of it.
I know “AI the heck out of it” is not a complete and accurate description of the delicate and complex model-building task these boffins will have to carry out, and I promise to talk about all of that in agonizing detail in future posts. But really that’s the general idea. Collect a lot more data than we currently have, and then just throw the whole evolving arsenal of machine learning and artificial intelligence at it, in any way that might conceivably work.
It will be difficult, but we’ve got a better chance of making progress on this problem now than we ever had before. And of course, if we may be able to do this with sperm whales, then eventually we might be able to do it with other kinds of whales as well. And with all the different kinds of dolphin. If we can manage to do this with a single cetacean species, we can eventually do it with all of them.
We may finally be able to know what all of them are saying to each other.
All of that is pretty exciting, and it’s exactly the kind of thing we thought we wanted to fund when we first decided to do this project. But now I’m going to end my post with a cliff-hanger that should make the whole thing seem even more exciting. For people in WA, at least.
If everything I’ve described might be possible, it raises a question about how to interpret the math in Project CETI’s big theory paper. (Linked below.)
A little while ago I said that the question of which kind of whale to AI up first is a very interesting one. So if that’s a genuinely interesting question… are they really just talking about sperm whales in that paper? Or, when I am reading it, should I be thinking about it as a mathematical model of the problem of communicating with whales in general?
You can certainly read the paper as just describing the difficulties involved in an effort to understand the click language of sperm whales. But if you chose, you could also read it as a more conventional kind of scientific paper, a paper that makes a prediction. A paper that predicts which species of whale will have its language translated first.
Will it really be sperm whales? Or will one of the other communications systems be deciphered sooner, before we finally succeed with sperm whales?
In particular… Should we, perhaps, be thinking about orcas? Should we be thinking about the Southern Resident community that lives right around here? According to Project CETI’s very pretty mathematical model, compared to sperm whales, would it be easier, or harder, to understand them?
If that question interests you, please stay tuned for the next installment of this blog.
Daniel Cloud
We strongly encourage everyone to support this amazing project. If you’d like to find out more or get involved, you can go to their new website:
https://www.Projectceti.org/get-involved
Stories (and a video) about Project Ceti from the media:
https://hakaimagazine.com/features/are-we-on-the-verge-of-chatting-with-whales/
https://www.youtube.com/watch?v=hph9OeKjg3w
https://www.nytimes.com/2022/08/30/science/translators-animals-naked-mole-rats.html
Actual science stuff:
https://www.sciencedirect.com/science/article/pii/S2589004222006642
https://arxiv.org/abs/2211.11081