Dolphins & DeepMind: Cracking the Code of Animal Language with Dr. Denise Herzing | [Sidecar Sync Episode 107]

Summary:

In this episode of Sidecar Sync, Mallory Mejias is joined by marine biologist and behavioral researcher Dr. Denise Herzing for a one-of-a-kind conversation about dolphins, data, and deep learning. Dr. Herzing shares insights from her 40-year study of Atlantic spotted dolphins and how that lifetime of underwater research is now powering DolphinGemma—an open-source large language model trained on dolphin vocalizations. The two discuss what it means to label meaning in animal communication, how AI is finally catching up to the natural world, and why collaboration across disciplines is essential to understanding both language and intelligence—human or otherwise.

Dr. Denise Herzing is the Founder and Research Director of the Wild Dolphin Project, leading nearly four decades of groundbreaking research on Atlantic spotted dolphins in the Bahamas. She holds degrees in Marine Zoology and Behavioral Biology (B.S., M.A., Ph.D.) and serves as an Affiliate Assistant Professor at Florida Atlantic University. A Guggenheim and Explorers Club Fellow, Dr. Herzing has advised the Lifeboat Foundation and American Cetacean Society and sits on the board of Schoolyard Films. Her work has been featured in National Geographic, BBC, PBS, Discovery, and her TED2013 talk. She is the author of Dolphin Diaries and co-editor of Dolphin Communication and Cognition.

Timestamps:

👥Provide comprehensive AI education for your team

https://learn.sidecar.ai/teams

📅 Find out more digitalNow 2026:

https://digitalnow.sidecar.ai/digitalnow2026

🤖 Join the AI Mastermind:

https://sidecar.ai/association-ai-mas...

🔎 Check out Sidecar's AI Learning Hub and get your Association AI Professional (AAiP) certification:

https://learn.sidecar.ai/

📕 Download ‘Ascend 3rd Edition: Unlocking the Power of AI for Associations’ for FREE

https://sidecar.ai/ai

🛠 AI Tools and Resources Mentioned in This Episode:

Wild Dolphin Project ➔ https://www.wilddolphinproject.org

DolphinGemma ➔ https://deepmind.google/models/gemma/dolphingemma/

TIME AI 100 ➔ https://shorturl.at/OUQ2h

Google DeepMind ➔ https://deepmind.google

👍 Please Like & Subscribe!

https://www.linkedin.com/company/sidecar-global

https://twitter.com/sidecarglobal

https://www.youtube.com/@SidecarSync

Follow Sidecar on LinkedIn

⚙️ Other Resources from Sidecar: 

• Sidecar Blog
• Sidecar Community
• digitalNow Conference
• Upcoming Webinars and Events
• Association AI Mastermind Group

More about Your Hosts:

Amith Nagarajan is the Chairman of Blue Cypress 🔗 https://BlueCypress.io, a family of purpose-driven companies and proud practitioners of Conscious Capitalism. The Blue Cypress companies focus on helping associations, non-profits, and other purpose-driven organizations achieve long-term success. Amith is also an active early-stage investor in B2B SaaS companies. He’s had the good fortune of nearly three decades of success as an entrepreneur and enjoys helping others in their journey.

📣 Follow Amith on LinkedIn:
https://linkedin.com/amithnagarajan

Mallory Mejias is passionate about creating opportunities for association professionals to learn, grow, and better serve their members using artificial intelligence. She enjoys blending creativity and innovation to produce fresh, meaningful content for the association space.

📣 Follow Mallory on Linkedin:
https://linkedin.com/mallorymejias

Read the Transcript

🤖 Please note this transcript was generated using (you guessed it) AI, so please excuse any errors 🤖

[00:00:00] Intro: Welcome to the Sidecar Sync Podcast. Your home for all things innovation, artificial intelligence and associations.

[00:00:14] Mallory: The ocean has a voice, and for 40 years, Dr. Denise Herzing has been listening. Hello everybody. Welcome to the Sidecar Sync Podcast. My name is Mallory Mejias, and I'm one of your hosts along with Amith Nagarajan. And today we are joined by the very special Dr. Denise Herzing. As a marine biologist and behavioral researcher, she spent decades swimming with Atlantic spotted dolphins in The Bahamas, recording their clicks, whistles, and bursts in an effort to understand how they communicate.

[00:00:47] Dr. Herzing is the founder and research director of the Wild Dolphin Project and was recently named to the Time AI 100 list for her pioneering work using artificial intelligence to decode dolphin communication. [00:01:00] Her research now powers Dolphin Gemma, a collaboration with Georgia Tech and Google DeepMind, and the first large language model trained on dolphin vocalizations.

[00:01:11] In our conversation, Dr. Herzing shares how AI is revealing hidden structure and dolphin sounds and what that means for the study of language itself. What it takes to make 40 years of field data usable for modern AI tools. Why her team believes open source collaboration is the future of animal communication research, and how projects like Dolphin Gemma remind us that technology when used wisely can help us listen more closely to the natural world, everybody.

[00:01:40] Please enjoy this interview with Dr. Denise Herzing. Dr. Denise Herzing. Thank you so much for joining us on the Sidecar Sync Podcast. What an honor to have you here. How are you doing today?

[00:01:53] Denise: Good. Glad to be here. It's exciting time in ai. So yeah, we're talking dolphins in AI these days.

[00:01:58] Mallory: I know. I have to say [00:02:00] this is a first for us on the podcast.

[00:02:01] I don't think we have ever discussed dolphins, but I'm excited to be doing that today with you.

[00:02:06] Denise: Okay, let's do it.

[00:02:08] Mallory: I always like to start off these, uh, interview episodes with a little bit about your background from your perspective. So can you take us back to the beginning of your career, what made you start studying dolphins and kind of tell us where you are at this moment in time.

[00:02:25] Denise: I was really inspired as a 12-year-old, believe it or not. Uh, just from looking through the encyclopedia when, when we had books and, uh, I, I would stop at the whale and Dolphin page and I'd be like, what is going on in their minds? Do we know? And kind of simultaneously, you know, like Jane Goodall was out doing chimpanzees.

[00:02:47] Diane Fosse was doing gorillas. Cynthia Moss was doing elephants and Jacque Cousteau had opened up. It's the ocean world. And so I was like, that's it. That's what I'm gonna do. So I studied marine biology, you know, did some [00:03:00] graduate work with dolphins and whales and seals and sea lions, but kept focused on, I was really interested in dolphin minds, dolphin communication, and when I got to the point.

[00:03:11] Uh, graduate school. I moved from San Francisco to Florida, uh, to work in The Bahamas, and I found a place, I was kind of surprised. No one was really out there working, but, uh, I said, well, I'll plant myself here for 20 years now it's 40 years, but who's counting? And, uh, really from the beginning. I knew I wanted to record sound and behavior.

[00:03:34] I wanted to understand the individual roles, the society, their behavior, and just see what we could learn because. I chose The Bahamas. 'cause you can see underwater and most places in the world, it's not so easy to research, uh, dolphins or whales underwater 'cause it's murky. So that was my goal, was to take a look underwater, which is where they really behave.

[00:03:58] So I was lucky to find a great [00:04:00] site. I was lucky to find a species that was relatively friendly to humans over time at least, and had a very, uh, full visible life at sea. So. Here I am 40 plus years later and my dream has come true 'cause AI now exists. So I, I knew it would come, I actually knew it would someday there'd be a computer I could throw my sounds in and make sense of, I mean, meanwhile we're doing all this manual work, you know, pulling sounds out with behavior.

[00:04:34] But, um, it's here now. So we have a great data set and we're using it for that reason.

[00:04:40] Mallory: I, I love that phrasing AI was a dream come true and that you knew it was coming. I feel like we've heard that a lot on the podcast. Amit talks about that as well, of kind of a lot of converging innovations that had to happen in order to make AI kind of explode in the way that we know it.

[00:04:54] But the, the building blocks were there all along. Right,

[00:04:57] Denise: right. Yeah. No, it's exciting.

[00:04:59] Mallory: So you [00:05:00] founded the Wild Dolphin Project in, in 1985.

[00:05:04] Denise: Yeah, I set up, uh, the Wild Oven Project as a nonprofit. Mm-hmm. So I could run my grand money through it. And I had individuals who wanted to help with the work over time.

[00:05:12] So yeah, we've built it over the years and that's how we run our work in our organization. And we have a 60 foot power catamaran that was actually donated to us Woo.

[00:05:23] Mallory: Wow.

[00:05:23] Denise: In 1992. So that became the permanent platform for our work. It's a nice, stable, live aboard platform. So that lets us go where the Dolphins live, which is.

[00:05:33] Kind of the first step.

[00:05:35] Mallory: Wow. Uh, what did technology look like, uh, back in 1985 compared to what you can do now with ai? Uh, what did that look like when you were collecting research out in the field?

[00:05:45] Denise: Well, I always had video and sound through Okay. You know, commercial, uh, cameras. So the sound range was human hearing only.

[00:05:53] And we know dolphin Spink sounds way up. Ultrasonically. Right. Um, and there were some ways to record high [00:06:00] frequency. They were pretty, um, challenging for our situation 'cause we're in the water swimming with the dolphins away. We're not like hooked up to the boat with a hydrophone hanging 'cause they aren't hanging around the boat.

[00:06:10] They're on the move. Um, so yeah, it was really basic. I think we had one of the original soundboards that we put in our Mac computers. That was like $5,000 or something. It was crazy. But it did like. We could analyze with Spectrograms and in graduate school, I actually had, you won't remember it, many people won't remember it.

[00:06:32] I had a, a cylinder. Uh, you know, the old fax machines that, that, um, worked with carbon paper that, well, it was a cylinder where you put sounds in and it rotated, sketched out the sounds Wow. On carbon paper. It was crazy. But that's how we got our images of dolphin sounds. So, and then once it became digital, of course.

[00:06:53] Everything's digital, right? But, um, we've gone through a lot of technology. You know, the cameras got smaller, the [00:07:00] sound recording then became high frequency. You know, we worked with different colleagues to sort of evolve that over time. So we kept improving the technology as it turned up. Uh, now it's all GoPros and high frequency recorders and 360 cameras.

[00:07:16] Wow. Also a great thing, um, for behavior.

[00:07:20] Mallory: Mm-hmm.

[00:07:21] Denise: So, yeah, so technology is our friend.

[00:07:24] Mallory: We love to hear it on the side Cursing podcast. Now, Dr. Denise, we found you, well, you've been, you're a very well known figure, but we found you when you were selected to the time, uh, AI 100 list. So I've gotta say congrats to you on that.

[00:07:38] What was that recognition like for you?

[00:07:40] Denise: Thank you. Well, it was. Kind of a surprise because I'm on the user end, right? I work with this great team at Georgia Tech and Atlanta, that Starners Computer group and the Google folks, DeepMind, they're the ones doing the AI and working with our data. So it was a real honor to also just showcase there [00:08:00] are other purposes for AI that maybe we haven't thought of.

[00:08:03] And of course, science of course is huge and medical is, is. Probably, I think the best example of a great use of ai, but scientists like myself, we have smaller data sets. We want quicker ways to look at it, and our human minds are great, but AI can really help with pattern recognition that we can't get at.

[00:08:25] So I think in some ways it's letting go of like. Humans, we can do everything. We're so smart. It's like,

[00:08:32] Mallory: oh,

[00:08:33] Denise: ooh. We've created a tool that probably is smarter or quicker in ways that our brains. Kind of can do, but we're not there yet, you know, whatever. So it helps us mine our data. So, yeah, so it was a cool honor, like I said, a surprise to be with those amazing people, the other 99 awardees.

[00:08:53] And, um, they're the ones that have done the work, right? It's like I wanted to thank them for creating. AI tools more than [00:09:00] anything because I'm just using it for my science, which is great.

[00:09:03] Mallory: Well, but I've gotta shout you out too. You're taking that amazing technology and, and utilizing it in a very unique and interesting way.

[00:09:09] So shout out to you for that.

[00:09:11] Denise: And what's interesting is that the, uh. Current AI system with Dolphin Gemma that we're doing with Google. Mm-hmm. It's really designed, it's, it's modeled after with our acoustic data. It's been trained with our acoustic data, but the whole point of it also is to create an architecture where other researchers can put in their own data and use it.

[00:09:32] So it's not just like, okay, we're looking at dolphins grade. It's meant to scale up to other researchers. That's a really important point. To do science is that you wanna create tools that you can all use and then compare potentially. Right. So

[00:09:46] Mallory: we're big proponents of open source technology here on the podcast.

[00:09:49] Was that always the plan? I know you said you collaborated with Georgia Tech. I'm based in Atlanta actually, and with Google DeepMind. But was open source always the plan when you were working on uh, dolphin Gem? [00:10:00]

[00:10:00] Denise: Uh, DAFO, Gemma. Yeah. It's always been a plan to be open source. Mm-hmm. It's not yet. They're still working on updating some architecture with, uh, some Gemini stuff, but, um, yeah, no, that was the plan, certainly with Dalton, Gemma to share it.

[00:10:14] Mm-hmm.

[00:10:15] Mallory: Yep. After you said 40 years, you kind of whispered it at the beginning of the podcast after 40 years. Before we get into more of the, the technical breakthroughs and, and what you've learned, if you could sum up what. You've learned about dolphins as it might be a very challenging question for you in 40 years.

[00:10:33] What would you tell our audience? Like what do you want our listeners to know that most people probably get wrong when it comes to dolphins?

[00:10:41] Denise: So dolphins are very long lived, uh, complex social mammal. They have friends, they have families, they have fights, they have a lot to communicate about. Um, they can live until their fifties.

[00:10:55] So there's a lot of knowledge in the elders in a dolphin community. [00:11:00] Um, they have cultures. They're unique. So when we, uh. See oceans crashing and food sources drying up. We're losing cultures of dolphins and whales like we would consider cultures of humans, you know, losing them for not saving the environment.

[00:11:17] Um, they're smart. The species I work with had, well had past tense particularly. I won't say easy life, but they had a a year-round food source, so they had a lot of leisure time, I guess I would say. Like again, like human cultures might

[00:11:34] Mallory: mm-hmm.

[00:11:34] Denise: In productive areas where they live. So we've been able to spend time with them both observing them and interacting with them in various ways.

[00:11:42] They're smart. They interact with other dolphin species, they're bottomless dolphins, uh, along with the spot of dolphins, if you see above me here. Uh, so they interact complicated lives with their neighbors. They cooperate with 'em. They also fight with 'em, so they kind of work through it. Um, [00:12:00] I really do think we're gonna find some complicated parts that, uh, might.

[00:12:06] Show that they have some kind of language. Uh, it's a tough definition for people to swallow, right? Um, but social mammals with long lives and complicated sensory systems have the potential for that. Sperm, whales, dolphin species, um, orcas. We don't know, and I'm just excited because we

[00:12:29] Mallory: mm-hmm.

[00:12:30] Denise: Have a tool to answer this question potentially.

[00:12:33] And there's other land mammals too. People are looking at all sorts of species, penguins and, and, and elephants. And so it applies to all of nature, even ecosystems potentially. And it's just a question if we're willing to listen and explore nature relative to our, our place in it.

[00:12:53] Mallory: Beautifully said. When it comes to Dolphin, Gemma, you said, so this is a large language model that has been trained [00:13:00] on how many years of, of sounds would you say?

[00:13:03] I

[00:13:03] Denise: think they have about eight or nine, maybe 10 years of sound. Um, and then, you know, again, it's various sizes of sound depending on the, the years. Um, we have a lot more to throw into the model Right. As soon as they kinda get it updated and, um, yeah, it's generative ai, right? So it allow, allows us to. A sound and see what comes out, which is pretty cool.

[00:13:25] Right.

[00:13:25] Mallory: Okay.

[00:13:26] Denise: So it should show us patterns that are likely to follow, like a whistle or a sequence of squawks or something, and we're trying to integrate it and then go back in our video and say, okay, what is going on when this pattern is. In the sound right, and are they doing the same thing? Can we encode it like we might encode words or phoning like speech and then take a look at it and see what do we see?

[00:13:53] You know, we have the underwater video. We can go back and look at what matches, and we know the individuals so we can say. [00:14:00] That's a pattern from a mom and a calf. And it's a little different, but it's a different mom and calf here, right? Maybe with a different name or signature whistle, which is their name, basically.

[00:14:08] So it has to be tested and the most challenging part of it, even after we get through looking at our video with these sound patterns, we'll be to correctly interpreted,

[00:14:20] Mallory: right?

[00:14:20] Denise: And we should be able to do a lot of that with our video. But it's still the question of are we. Really getting it right. So we might have to actually play back sounds to the dolphins and see what they do.

[00:14:32] That's probably the biggest challenge. Um, I dunno how realistic it is, but it's possible. Uh, so we have, you know, a long way to go, but we have the data and we're gonna work through it and try to take a look at it.

[00:14:45] Mallory: Wow. Okay, so you provide sounds to the LLM that's been trained on sounds and then it can generate new sounds because it has all that training data and it's recognized the patterns, recognizing the patterns.

[00:14:56] Is it, I don't know how to ask this, but is it a kind of a [00:15:00] two-way interface where it's producing. Dolphin sounds and then attempting to put those into words that we would understand.

[00:15:07] Denise: No, no. It's just producing the sound patterns and we would take the sound patterns. We're actually right now trying to integrate it with another program we have that will label those sounds, for example, like A, B, BC.

[00:15:20] Mallory: Okay.

[00:15:20] Denise: And we're also working on interface, uh, with an MIT colleague, um, Mark Hamilton. Uh, he's actually taking the video. A sound and creating interface. So it makes us EAs, it's easier to track. Okay, let's go back to the video and find this sound pattern. So yeah, we will probably just symbolically label it arbitrarily, but it still will tell us if there are patterns.

[00:15:42] I mean whether, I don't know if we'll ever be able to call them words, but I guess if we had a sound sequence and they were chasing a very specific type of fish, maybe eventually we'd find a label for, you know, the macro versus the flounder. You know, that kind of thing. Uhhuh, if they have. We dunno if they have words, [00:16:00] I suspect could be pretty handy to have words.

[00:16:02] Mm-hmm.

[00:16:02] Intro: Um,

[00:16:03] Denise: like you can imagine if you're out in the deep water with your friends and family and a shark shows up, it'd probably be pretty handy to be able to label that shark as a dangerous tiger shark or a harmless nurse shark, for example. So, and you know, terrestrial animals, vervet monkeys, prairie dogs.

[00:16:22] We know they already have alarm calls that label specific predators. So we, we see that in evolution already with community animals is that they have labels for specific predators. So you can react specifically. Right? So that's, and it makes sense that those words would evolve for survival, right? Mm-hmm.

[00:16:42] Be like if I yelled, look out. And you'd be like, look where? Look where. And I versus I said Asteroid. And you'd look up, right

[00:16:49] Mallory: and Yeah, exactly. Yeah.

[00:16:51] Denise: So it helps you maybe, well, I don't know if we'd survive an asteroid, but anyway, you'd have a few

[00:16:56] Mallory: minutes. That's a question for another pod, but I'm curious because, [00:17:00] and some of our listeners might be interested in the nuance as well.

[00:17:02] I certainly am. So. An alarm call that we might see with monkeys is different from language. Can you talk about kind of like the nuance there, because to me that sounds like a sound that's communicating something. So how is that different from a, a language?

[00:17:16] Denise: It's because it's labeling a specific object. So like they would've an alarm call for a leopard versus an eagle versus this.

[00:17:24] Okay. So it's, it's a, it's a, it's a label. Okay. So it's called referential communication. So it refers to something versus if they're just like, say the Vervet monkeys are having a fight and they're screaming and scratching each other, whatever. That's just emotional. Kind of expression. And that's called a graded system of communication, right?

[00:17:45] So it's referential versus graded. So humans have both, I'm using words which are referential and you understand them, of course. Um, but I'm also getting excited and I'm waving my hands and my voice is getting louder. So animals do that too, [00:18:00] right? They express present current activities, right? But if you have labels and words.

[00:18:07] That allows us to talk about the past and the future. Right. What'd you do yesterday? You know, versus what'd you learn at school today? You're at the dinner table, you're not talking about the food necessarily. You're talking to each other because you have words.

[00:18:21] Mallory: Mm-hmm.

[00:18:21] Denise: So, yeah, that's the big question. And you know, maybe a lot of animals don't have something like that, but just evolutionarily it makes a lot of sense.

[00:18:31] Or at least a long-lived mammal to be able to talk about things, plan things, maybe. It's just we have not had the tools to look at it. You know, some scientists have done experiments and it suggests a lot for sure. So language is kind of the last bastion of intelligence, right? Like animals use tools they can understand and abstract things.

[00:18:54] Um, so they have a lot of the qualities of intelligence.

[00:18:58] Mallory: Mm-hmm.

[00:18:58] Denise: But language still has [00:19:00] yet to be really looked at, I think.

[00:19:02] Mallory: I'm curious in your 40 years of data collection, I'm trying to think here of things that might resonate with our association listeners at a glance. Right? Dolphins and associations seem quite different, but I feel like something that might be relatable is the fact that you had 40 years of data collection.

[00:19:18] You've ridden out a lot of technological change in your career and getting to this point where we have access to things like large language models. Did you ever feel like. Your data wasn't ready for ai. That's something we hear a lot for associations. They have tons of unstructured data. They've got conference proceedings, course videos, transcripts, mm-hmm.

[00:19:40] Publications, all this stuff, and they think, well, sure, we would love to maybe train a model, but everything's so messy, it's so all over the place. You as a researcher, maybe that wasn't as much of an issue, but can you talk about that?

[00:19:52] Denise: Sure. So I would say our. Our data sets are pretty clean as far as our sound [00:20:00] sequences and the underwater video.

[00:20:02] Um, we have what I call metadata, right? So for each like video segment that has the sound, we would know there's four mother and cap pairs. We know who they are. We know when the calves are born. We know their age, we know their sex. Um, so that's the metadata we would lay on top of it to interpret it.

[00:20:21] Mm-hmm. So I, I guess in the sense of like a conference or meetings, if you've got a soundtrack, you've got the individuals, you've got the contextual cues, maybe in, in that case, key words that come up, uh, in that, um. But right now, literally that's still kind of a manual process for us.

[00:20:43] Mallory: Okay.

[00:20:43] Denise: You know, with the video and the sound, we can go back to our metadata.

[00:20:47] It's in another data set. It's all okay. It's all, um, connected to one label, basically, you know, the date, the time, the encounter number with the dolphins and everything else. Is related to that. You know, we know what the water temperature was, [00:21:00] blah, blah, blah, blah, blah. But what's critical to look at is not necessarily a lot of that data, but it's about who's there, what they were doing, and the behavioral context.

[00:21:10] So, I mean, I think you could kind of do that for meetings and conferences. You could say there's a board meeting versus a, um, sponsor discussion versus, uh, staff. Meeting. So those are the keywords in the context where I would say that's where you have to separate your searching. Mm-hmm. Is look at the behavioral context.

[00:21:30] Like we might look at, you know, here's an aggression video and here are the sequences sound that fall out of that. 'cause we know these animals were fighting, they're males and they're trying to mate with a female, et cetera, et cetera. So I, I could see where a large language model could easily do that. I don't know about.

[00:21:48] The searching a of an AI program would have to let you do that, right? Like

[00:21:53] Mallory: right.

[00:21:53] Denise: Search by keywords according to that segment, but I don't think it's impossible. It seems like that's fairly easy [00:22:00]

[00:22:00] Mallory: For sure.

[00:22:01] Denise: So, yeah, I don't know enough about the programs they're using, to be honest.

[00:22:05] Mallory: Right,

[00:22:05] Denise: right. But just thinking in terms of, yeah, tagging things, and I guess you could even have voice recognition, right?

[00:22:11] You've got humans, right? You can do voice recognition. So you could actually pull out who's there saying what. I know the conference, uh, soundscape is busy and noisy, but I know people have been working on that for sure, because we have the same thing, right? Dolphins are chatting. Who's chatting? What are they talking

[00:22:26] Mallory: about?

[00:22:26] Who's chatting? Who's saying what? Yeah, I guess maybe dolphin problems are not so dissimilar from human problems. Yeah. No, they're not. Um, so you said metadata, uh, for you all is still fairly manual, but I'm thinking as you were saying, AI could do a really good job potentially of watching the video of, um, a dolphin interaction and then saying, this appears aggressive.

[00:22:46] Now, I don't know if it has enough training data to be able to do that as accurately as you could, but it's something to think about. Yeah,

[00:22:53] Denise: yeah. We're working on it.

[00:22:55] Mallory: Can you talk a little bit about that first moment? And in my research it looks, sounded like [00:23:00] this was quite exciting. The first moment you heard Dolphin, Gemma actually create the, I think they're called burst, uh, burst pulse sounds.

[00:23:08] Yeah. What was that like?

[00:23:11] Denise: Well, it was like, wow, okay. It sounds like a bunch of dolphins.

[00:23:14] Mallory: What, was it exciting to see that that tech could do that? Or were you kind of like, I know, I know this is possible.

[00:23:21] Denise: No, I knew it was possible. Um, and you know, garbage in, garbage out, right? Yeah. Depends on how it's trained, what data they have, and it's, you know, something we're tweaking and working on as we go.

[00:23:31] Um, and that's why the supervised part of it is still important. I mean, I, I believe that sincerely. I know with our early machine learning work, which was more us creating models and creating. Interfaces to process our data. It was very critical for me to look at the sound clusters and say, okay, these are all similar sounds.

[00:23:50] We'll call that cluster A and cluster B, et cetera. To supervise because computer doesn't really know what you're looking for. It's just looking for patterns and pulling them out. [00:24:00] Right? I mean, the classic story is the polar bear photos, right? When they were looking to identify individual polar bears and the computer decided to look at the types of snow.

[00:24:10] So, you know, it was like, okay, gotta point that out to adjust, right? So I.

[00:24:20] Think we're there that it's looking for, or it's like the conference room, right? Where someone's, um, you know, tapping their pen on the desk and so the computer pulls that out as a cluster of sound. Mm-hmm. Or someone's coughing. Right. Is it critical? Probably not, but maybe this could be if someone's nervous or, you know, who knows, right.

[00:24:39] So

[00:24:40] Mallory: right

[00:24:40] Denise: computer can do things for sure that we don't even think about.

[00:24:44] Mallory: You developed this wearable computer system called Chat, where you could actually potentially have, I guess, two-way conversation with Dolphins. I, I want you to talk a little bit about that and does that use any artificial intelligence?

[00:24:57] Denise: Sure. So the chat system, it's Satan [00:25:00] Hearing and Augmentation Technology. It's basically an underwater computer we wear.

[00:25:04] Mallory: Mm-hmm.

[00:25:05] Denise: Even though our primary work is to decoding their natural sounds, which is our bigger 40 year data set. Um, another way to try to interact and understand dolphins is using a keyboard that's been done historically.

[00:25:18] There's a whole history of dolphins and keyboard interfaces like you would have for primates. They'd be touching a keyboard where, where dolphins have had acoustic interfaces, right? So chat was designed to be a very simple, um, system. That had, uh, synthesized whistles in them that we created that are outside the dolphin's repertoire.

[00:25:40] 'cause we didn't wanna insult them by saying something in dolphin. Right. Like, your mother's my old girlfriend.

[00:25:48] Mallory: I guess I never thought about that, but you're right.

[00:25:51] Denise: Oh, you gotta be careful. So, and it's only got like four words in it that are labels for objects they like to play with sea. Uh, Sargassum.

[00:25:59] [00:26:00] Seagrass.

[00:26:00] Mallory: Okay.

[00:26:00] Denise: Then, um. We take in a scarf 'cause I like to drag things. And we only use it when they're playing with us. We don't interrupt their behavior. That's our, uh, protocol. But, um, it's really designed to maybe create some mutual signals we could use together. It's kind of the second way you could maybe kind of get into the mind of a dolphin.

[00:26:20] And interestingly, um, a lot of the experimental work with other animals have used that kind of model where you create a, a keyboard or interface and so you share words like, like if we were different humans, right? A culture, you know, I might call this phone and then you call it Lala or whatever, right?

[00:26:40] Mm-hmm. So we agree. Let's call it tutu, you know, Uhhuh. So now the word for phone is tutu. And you both understand that even though you have your own languages,

[00:26:49] Mallory: right?

[00:26:49] Denise: So in some ways, um, I still believe this, although it's been a hard project with the dolphins, um, in some ways it's almost easier to create a mutual language [00:27:00] than to understand someone else's language.

[00:27:03] We see that in nature, we see, um, species of dolphins that interact, for example, and they have mutual signals they use, but they don't necessarily know the whole language of the other species. So they have kind of critical signals apparently they need when they're together and they've agreed on that somehow some way.

[00:27:23] And, um, birds, monkeys, uh, other species do that. For simpler interactions, right? I mean, you could probably never have a full conversation. Um, like you couldn't with a human. If you only had six words, you shared, you know, coffee, a coffee and a tea, and Ahuh biscuit or whatever. That'd be a very simple type of interaction.

[00:27:44] But it's good in the sense that it shows the other species you're reaching out. You're trying to empower them to communicate. For example, with the Dolphins, if they mimic one of our synthetic whistles for an object, they get the object and we get to play together. And [00:28:00] that's pretty high motivation for them 'cause they like to play, right?

[00:28:03] We don't feed 'em or touch 'em or anything like that. But, um, so species can be motivat motivated simply by social interaction with each other and that might be enough. So, um. But I think the richness of what we're gonna find is gonna be in their natural sound patterns and using the LLM. But it, so it's been interesting exploration, you know, it's science is so funny.

[00:28:26] Sometimes you find things you weren't looking for when you're looking for other things. And with the chat org, because we looked at how often and how they mimicked our, our, were trying to mimic, which they were in many, many different ways. We found a pretty cool way that there. Try and mimic that we didn't know really potentially existed in their sound repertoire.

[00:28:49] And so now we're going, Ooh, we need to go back in our, our, um, basic sound data and look for those patterns because we weren't looking for those. 'cause it's a [00:29:00] weird thing they're doing. And it's like, it seems to be how they're preferring to mimic. And I'm like, okay, there you go. You've discovered something.

[00:29:07] Maybe not what you wanted with results, but now you've discovered. A way they might be communicating that you hadn't even seen before. Oh,

[00:29:15] Mallory: this is so fascinating. So they're, they're mimicking this mutual, uh, not language, but these mutual, uh, signals or words that you've created. Yeah. But they're not necessarily understanding them, or do you think they are understanding the word?

[00:29:29] Denise: Yeah, no, we haven't gotten to that point. Okay. It doesn't seem like they've been using it to like request the toys from us. Ah,

[00:29:36] Mallory: okay. Gotcha.

[00:29:36] Denise: So much. I think it's because they can kind of get them themselves

[00:29:39] Mallory: too. That's true. Right?

[00:29:40] Denise: And I was like, oh, there's some sargassum, I'll just go get it. Um, but who knows, you know?

[00:29:44] Yeah. Maybe it's hidden in there when they're talking. Maybe they'll say, Hey, did you see that sargassum? We played with the humans yesterday. I dunno, you know. I mean, 'cause that's happened in, uh, captive studies where they've trained them with synthetic whistles and then they go off and they play with a [00:30:00] ball with each other, and then they make that whistle.

[00:30:02] So it's kind of like, you know, maybe they do. So, yeah, I don't know though. I wouldn't say we're there, but we've been trying, we've been trying. Yeah. Who knows? Maybe. And, and the chat system, uh, has been using a bit of dolphin Gemma. Okay. Only in the sense of, um. Taking our synthetic whistles, prompting it in dolphin, Gemma, and then showing us different ways the dolphins might mimic, right?

[00:30:28] Projecting, again, generative ai, projecting how they might mimic and they might give you like a hundred different ways, and then we go through and say, well, we would accept this as a mimic, but maybe wouldn't, we would not accept. Right. So in the past, students would have to generate these sounds and say, okay, let's stretch it out.

[00:30:46] Let's put it up in octave, you know, what do you accept as a mimic? So, yeah, there's definitely a lot of, uh, things that have been involved with chat with, um, machine learning and ai. So, yeah. But it's, it's a tough, I mean, poor [00:31:00] Georgia Tech, you know, Thad Starner and his students, they thought it'd be an easy project.

[00:31:03] And, you know, 15 years later or so, 15

[00:31:06] Mallory: years later, we're just probably scratching the surface.

[00:31:09] Denise: Exactly. But now we have like off the shelf equipment, we're using a pixel phone. I mean, it's all like, we can just take it in the water and it's like a little box this big and we wear it. But, and you know, so it's, that's all changed too.

[00:31:22] So as we go, here we go. Technology.

[00:31:25] Mallory: Oh man. Yeah, I thank you for bringing that up because I did wanna mention that dolphin Gemma, I have the note here. I think it's 400 million parameters, which if you all have been listening to the podcast is pretty tiny. Um, so it can run on device. I mean, some of the smaller frontier large language models that we talk about are like 3 billion parameters.

[00:31:43] 4 million parameters. So 400 million is, is small. Uh, and that's incredible.

[00:31:48] Denise: I'm like, what are those parameters, right? Hitting me. Measuring all those parameters and I'm measuring eight of them, you know, manually.

[00:31:55] Mallory: Yes,

[00:31:55] Denise: I know. It's crazy. But, you know, we have to get more efficient with it [00:32:00] because a lot of researchers don't have large data sets.

[00:32:02] Our lar our data set is not large. It's large for my field, you know, 40 years of data. Um, but animal researchers don't always have volumes of data, so it's nice to get more efficient to really look at it.

[00:32:14] Mallory: Mm-hmm. I wanted to ask you this question because I feel like you have a unique perspective as a researcher, as someone who's obviously passionate about animals and the environment, we tend to be very optimistic on the podcast about artificial intelligence in terms of the good it will do in the world, but we're also realistic, uh, and that it can be used by bad actors.

[00:32:36] It does have an impact, a large impact on the environment. The water needed to cool, the energy required. For you as someone who's just very aware of the earth and and animals, and as I said, like do you find, has it been tough for you to go back and forth with like AI for good, AI for bad, or have, it sounds like you're very clear, uh, on the potential good [00:33:00] that AI can do, but has that been a struggle for you?

[00:33:03] Denise: It has not been a struggle. I mean, I'm very positive that the complexities of natural systems, ecosystems, climate. Change, you know, hurricanes, whatever. These are all places where data is complicated. Mm-hmm. And I think AI is probably gonna allow us for the first time to really look at these complicated systems and understand them.

[00:33:28] Um, I do not know how the energy issue's gonna be solved with AI needs. I mean, I think. As a country and as a global uh, entity, we all need to be thinking about that. Of course. And maybe AI can tell us how to do it, right? Right. How can you get more efficient with your energy? You know, is it that you know we need more solar and wind, or is it some other way that ai, is it quantum computing and might reduce that kind of thing?

[00:33:59] I dunno enough [00:34:00] about the physics of that, honestly. I've been positive about it because I think it's gonna spur science on, if we allow science in this country still to be looked at in some detail, you know, we're gonna need scientists in some of these agencies that have had a tough time with the guillotine.

[00:34:15] But, um, you know, if we're gonna look at the planet and there's other countries on the globe that are, you know, taking it on too. So US is not necessarily. Um, leading the way, but I think we still have great resources in this country. A lot of great brains and a lot of great agencies and teams and yeah, I would like to see AI continue with analyzing complicated data for the natural world.

[00:34:38] For sure. I

[00:34:39] Mallory: think you're right. We talk about the medical breakthroughs, like you mentioned earlier. We talk about weather, material science. There's, we're, we're on the cusp of what seems like an explosion of innovation and breakthroughs that I hope will outweigh some of the potential negative impacts we see with like such a rapidly growing technology.

[00:34:57] Dr. Denise. What [00:35:00] is next on your horizon? I mean, dolphin Gemma is a huge feat and I know it's not, we haven't reached a finish line. It sounds like it's gonna be ongoing, but what, what are you looking ahead to for AI in general or perhaps, uh, in your field specifically?

[00:35:15] Denise: Well, honestly, once we work through our data sets, which you know, will take some time, I mean, the field is fast, but still going through, looking at the data, trying to make sense of it, it's still gonna take time and human brains.

[00:35:28] I honestly would love to see, and I may try to promote this a bit, is to get other researchers to put in their data sets. You know, maybe we go around the world and collect more dolphin data of, you know, from species that maybe we can't get 40 years, but if dolphin Gemma becomes really efficient. Maybe we can spend a month or two getting data and then throw it in there.

[00:35:51] You know, I think, I think that's realistic. I really do Once, uh, Gemma and Gemma gets really going and tweaked a bit. So I [00:36:00] think again, you know, I feel really lucky where I've worked and what I've seen in our data, but I think the world is a lot larger. Um, there's a lot of critters in the ocean we can get data from and throw it in there and, you know, I just think it would open up.

[00:36:15] Oh, it's not to spotted dolphins that have this complicated, it's every dolphin species or a bunch of them around the world, and we need to protect their habitat so they can thrive, you know, to their right as a, a species to thrive on the earth without us destroying it. And I think maybe that would call attention to it.

[00:36:32] So that would, that's one thing I'm kind of looking ahead as I still try to process our own data, but I, I think it could be worth doing and we'll see about other species I think. We'll have to see that. I think Dolphin Gemma might be able to incorporate like non dolphin species eventually. Okay. Um, I think the architectures may be designed to train, you know, the LLM with other, you know, terrestrial species for example.

[00:36:58] I think that was the ultimate [00:37:00] goal. So we'll see how that evolves. But it might not be limited to ocean creatures, hopefully. Wow.

[00:37:06] Mallory: You're gonna be joining us at Digital Now 2025, which is actually next week. So we're recording this on October 30th. I'll be interviewing you there on the panel again. So for all of you listening that are attending, you'll get to hear a different version of this conversation, a shorter version.

[00:37:21] But Dr. Denise, where can people, uh, keep up with you? Keep up with the work you're doing? Let us know.

[00:37:27] Denise: Sure. Our, our major website is the Wild Dolphin Project, you know, dot org. Um, you know, we're on Facebook, wild Dolphin Project, Instagram, I am on LinkedIn, I'm on Blue Sky. Uh, you know, nowadays people can find you, right?

[00:37:42] So, uh, yeah, we're, we try to do our social media, Facebook, YouTube. We have a YouTube channel for Wild Dolphin Project. So yeah, join us.

[00:37:50] Mallory: Absolutely. We'll include those links in our show notes. Dr. Denise, thank you so much for joining us on the podcast. Again, congrats on your selection to the time, AI 100 [00:38:00] and for everyone joining us at digital now.

[00:38:01] You will get to see her in person next week.

[00:38:04] Denise: All right. Thank you very much. Looking forward to it.

[00:38:07] Mallory: Thank you.

[00:38:10] Intro: Thanks for tuning into the Sidecar Sync Podcast. If you want to dive deeper into anything mentioned in this episode, please check out the links in our show notes. And if you're looking for more in depth AI education for you, your entire team, or your members, head to sidecar.ai.