Talking to Animals
AI-mediated conversations with non-humans are a long way off. But LLM-assisted projects with dolphins are a breakthrough nonetheless.
Imagine you’re walking down the street when you encounter an extraterrestrial. In one of its weird appendages it holds out a very cool looking gaming console – alien, yes, but comprehensible and deeply fascinating, as if it were the Nintendo Switch 5 or something. You’d love t play with it. At the same time a box on the alien’s thorax emits a word-like sound. “Dwerg!”
You’re a smart primate. After a couple of go-rounds you think to say “Dwerg!” And the alien hands you the toy. Communication established!
This is one of the strategies being used with DolphinGemma, a new Google DeepMind tool that uses AI to try for a shared “vocabulary” that dolphins and humans could communicate with some day.
In this work, the “extraterrestrials” are humans, hanging out in the oceanic world of the dolphins. A human researcher holds up something dolphins like to play with, like seagrass or a scarf. Her onboard AI then emits a sound. It’s one no dolphin ever made but it’s one any dolphin can make, because it has been synthesized from actual vocalizations the animals use.
Over time, this could give the two species a set of words whose meanings are clear to both, and which can be uttered by dolphins and by AI-connected speakers strapped to human chests (the AI itself is adapted from the same engine that runs “AI mode” on Google searches).
Treating Dolphin Clicks and Whistles Like Prompts
Naturally, researchers are also trying to learn natural “dolphin language”: that’s a DolphinGemma project that takes actual dolphin sounds and treats them kind of like prompts: Given a dolphin whistle or click, the AI tries to predict what sound is likeliest to come next.
No doubt this is weird for the animal. It’s the equivalent of walking down the street and starting to say “Trump is,” then hearing a disembodied voice say “driving the country into a ditch.” (Or, “headed for a third term, he’s the best,” I guess depending on where your pod of humans wanders the Earth.) Researchers can judge from your reaction whether the AI’s guess is meaningful to you, which will build their understanding of the meaning of the sounds you emit.
But real language (if dolphins have one), among speakers who share a past, a family or at least a culture, is a slippery thing. What does “Trump” mean, given that some Homo sapiens (like me) go wild with rage at that sound, while others are in rapture? What if the humans involved are playing bridge, and are seeing a whole other meaning in the word?
Decrypting all those complexities is well worth the work, of course, but it may raise difficulties that we can’t resolve for a long time, or possibly ever. Maybe you have to be a dolphin to be able to relate to one of these creatures, and maybe you have to be able to relate to say anything meaningful to another being, as Ludwig Wittgenstein famously asserted (“if a lion could talk, we could not understand him”). An artificial pidgin would be much more limited than real dolphin or real human communication, but at least all parties would know what the hell they were talking about.
None of these nuances stopped the hype machine last week from proclaiming this month that AI is about to let us converse with our dogs as God spoke to Adam in the garden. DeepMind’s CEO Demis Hassabis took a flying leap beyond justifiable pride when he tweeted this about DolphinGemma:
If you think about what is involved in constructing that seagrass-scarf pidgin, and the further complexities of learning dolphin-speak, you can see that AI-powered chats with your dog or cat are a lot further off than that tweet suggests. (Yes, I know we share a lot more references with dogs than we do with dolphins, but your pet wagging his tail when you say “walk!” doesn’t mean you know what “walk” means to him. You just know the signal he sends when he hears you, which is not the same thing.)
But let’s assume the day does arrive. If your dog or cat can talk about their experiences of you and your decisions, will you like what you hear? And what will you say back? What do you do when your neighbor tells your dog to throw off his shackles of oppression? (On X, user PresenceAwareness (@EternalSunrise7) saw the problem. The illustration for this post is his reply to Hassabis.)
Next post (also on language): Is there a distinct AI dialect of English?
Other Things to Read or Notice This Month
Robots Cause Crime
In this paper by economists Yang Liang, Joseph J. Sabia, and Dhaval M. Dave, they estimate that each additional robot added per thousand workers between the early 1990s and 2010 “led to a 4 to 5 percent increase in property crime arrests.” The logic is simple: Robots reduce the need for human workers (either directly swapping in for them or indirectly, by enabling a company to produce more with fewer employees). Those human ex-workers, now unemployed, suffer from what economists call a “need to generate income.” So they steal stuff. (It also makes sense that the authors’ methods did not find any signs that robot adoption spurs increases in violence or person-on-person crimes.) Earlier work has found a correlation between robotics adoption and rising unemployment, so this was a natural question to ask.
Artists Playing with AI
Though the composer Alvin Lucier is dead, his brain, with AI assistance, is continuing to make music. Meanwhile, an art project at the University of Applied Arts in Vienna created an AI applicant for its program. It was admitted.
Rulers Playing with AI
The United Arab Emirates announces it will use AI to write laws and evaluate their effect.
Do Large Language Models Reason?
Melanie Mitchell explains why the jury is still out on that one.



I wonder how this fits in with the pet word-button trend. Here is good example https://www.youtube.com/watch?v=z8k2upr9vCE, lots more people have posted similar videos...