Click, Click, Whistle!

We have officially reached the point where large language models (LLMs) have either attained unbelievable new heights of awesomeness, or have totally jumped the shark. These generative artificial intelligence (AI) algorithms that have risen to prominence in recent years because of their knack for carrying on in human-like chit-chat are now being used to shoot the breeze with dolphins. Yes, dolphins.

Model-specific architectural differences aside, the main function of an LLM is to analyze a sequence of words (or tokens) then determine what token should come next. Wash, rinse, repeat. While this basic approach is fairly simple, after being trained on enough data it is enough for the model to be able to answer all of our questions about our homework that we don’t feel like doing, help us with coding, or act as a research assistant.

After initially falling behind in the present AI gold rush in a very public way, Google is now applying it to almost every conceivable use case. Their latest foray into the field resulted in the development of an LLM called DolphinGemma, and it is exactly what it sounds like — a Gemma LLM for dolphins. And no, it is not exactly trying to be a translator, either. It works like traditional LLMs, but instead of using words, or parts of words, as the tokens, DolphinGemma uses the clicks, whistles, and squeaks of dolphins. After that? You guessed it — it attempts to predict the next most likely clicks, whistles, and squeaks.

Yet again truth has proven itself to be stranger than fiction. I suppose the modern dolphin will at least be able to get some tips on where to find the best squid this time of year. But all joking aside, the developers hope that if DolphinGemma can capture the structure of dolphin vocalizations, that those patterns might then be used to decode their language.

That is, of course, if dolphins have a language. Despite decades of research on the topic, no one really knows if they are talking to one another, or if the clicks, whistles, and squeaks are just clicks, whistles, and squeaks. And that is where the pattern-matching capabilities of the DolphinGemma LLM could turn out to be really useful.

If you have been dying to find out what Flipper was talking about since you first saw the TV show, your chance might finally be coming. Google is planning to release DolphinGemma as an open model sometime this coming summer. Keep your eyes open so you can download it and test out your best dolphin impression. I have been practicing myself, but everything is still coming out sounding more like a Wookiee. Hey, I think I've got a great idea for a new LLM...