Animals Could Easily Be Talking to Us If We Tried

Some random guy on the Internet's blog post about how he thinks talking animals are nearly a thing, with zero references/evidence or anything, doesn't really seems like HN content?

I chose to read this as a really good satire on the Dunning-Kruger effect.

Our AI systems "work" because we can derive meaning from the words that we feed into it, right? We put words in, train, and words come out. How would that exactly work with animals?

"Woof in, woof out" still means not knowing what the woof's all about.

Don't get me wrong, I have often thought about this exact question: that surely we are close to finding a way to communicate with animals or at the very least study them at greater lengths through the use of LLMs and similar systems. However, I have yet to find the exact way in which we can do this.

I'm sure we can create an LLM that mimics the expressions and behavior of animals (much like we have created LLMs that "mimic" us). But that will still give us very limited interpretability. It will definitely allow us to tinker with the inputs without needing a real animal, but that still gives us a very limited understanding of what exactly is going on.

I would definitely pour my heart and soul into such a project :)

This is an extraordinarily hand wavy blog post.

Fusion is really simple, too, you just hook up the things and there's power!

My opinion is that we have little to no interest in what animals, plants, or even other people are thinking. The vast majority of it would be considered crude and offensive at best.

There seems to be rather little evidence to back up this claim…

Is humanity prepared to find out that our loving dogs walk around with the thought “I can’t wait to kill this guy and take over the pack”.

There's DolphinGemma; no microchips needed -

https://blog.google/technology/ai/dolphingemma/

Putting aside quibbling over what constitutes language, talking, etc, animals do clearly communicate to us and understand us (to some limit). They read our facial expressions, hear our tones, can distinguish words and names. Similarly: any pet owner who pays attention can learn to read the body language of their animal companions, their tone of voice, and sometimes even distinguish what the pet wants or how they feel from individual vocal articulations. We've managed to teach great apes to use signs to communicate to us.

All this is to say: is there value in pretending like we can "translate" to english with complex grammar? Maybe not. But it might be interesting to learn and track, say, which sort of meow is "play with me", which is "feed me", which is "I'm stressed", which is "I want another toy", which is "I'm worried about you", etc.

There have been claims of teaching dogs to use buttons to communicate complex things; some of it is easy to believe (eg I have taught my own dogs to press a button when they want to go out—relatively straightforward conditioning), but some of it might be a performance for social media. I understand the skepticism, but it's surely worth researching to what extent the dogs actually are "communicating" versus seeking specific things, or even indicating concerns or emotions to us.

This gets even more interesting with animals with complex socialization of their own: whales, dolphins, birds, etc. Domesticated animals and our close relatives already have a genetic edge in communicating with us; but intraspecies communication of animals can be opaque or literally outside our ability to hear or differentiate. Surely algorithms and automated recording/correlation could reveal the complexity of these relations.

There is no training dataset to map the sensor data to thoughts. We don't know the meaning of what they're thinking. It's futile without a way to develop training data.

A very limited training dataset can perhaps be created with binary choices that the animal can be trained to communicate via physical actions. With sufficient effort, an embedding model mapping thoughts to concepts can be developed. Still, how to convert the embedding vector to text is unclear.

Here is what my dog would say when I’m going out…

“What?!” “no no no no no” “hey HEY” “come back!”

I could not bear to hear that. Barking is better.

If you could talk to the animals— just imagine it— they would say things that were only boring, distressing, or bizarre. And they would have no comprehension of what we say to them.

Dogs can already talk using buttons and great apes can talk with sign language. This seems feasible, maybe even without a microchip, just with non-invasive reading of brain waves.

My dog already talks. She barks. And her barks mean “hey!”

What exactly is AI going to do to improve on that?