"Forget all that. Judged against where AI was 20-25 years ago, when I was a student, a dog is now holding meaningful conversations in English. And people are complaining that the dog isn’t a very eloquent orator, that it often makes grammatical errors and has to start again, that it took heroic effort to train it, and that it’s unclear how much the dog really understands."
To quote the generally incredible Scott Aaronson with a somewhat more optimistic perspective:
Judged against where AI was 20-25 years ago, when I was a student, a dog is now holding meaningful conversations in English. And people are complaining that the dog isn’t a very eloquent orator, that it often makes grammatical errors and has to start again, that it took heroic effort to train it, and that it’s unclear how much the dog really understands.
> Animals with very limited language capabilities have the ability to run their own lives, and to some extent manipulate their environment.
This makes me think of the ongoing effort by some people to train their pets to use talking buttons to communicate. These include the animals occasionally trying to convey relatively complicated ideas with simple sets of words, like "play help" ("I'm bored, come entertain me") [1] and "dog settle" ("make that dog outside stop being annoying") [2].
> Everything I see in dogs suggests that they are sentient, they just don’t seem to need language. They speak when it is useful, like for getting attention from people in different rooms, but they don’t really need it beyond that. So they don’t go any further.
If you haven't check out whataboutbunny [1] on Instagram to get an indication of how far dogs can go when given an opportunity. Bunny is trained to use buttons to "speak". A lot of it is very simple and functional ("outside", "play") that could easily be handled by body language and the odd bark, but occasionally you get fairly complex conversations which seems to indicate introspection and fairly complex reasoning that as you say they "just don't seem to need" to be able to express in language in normal conditions.
This dog and the whole process of teaching is fascinating because it highlights how much we depend on verbal communications and easily discount other types.
> We've been building tech to translate dog barks to something meaningful for us to understand.
Can most dog owners not already understand or differentiate their dogs' barks? I can tell from the quality and context of my dog's barks, whether he
- wants something (food, water, attention)
- is angry
- is scared
- wants to play
- is excited
- is voicing a territorial dispute
and I think most dog people can also be confident in interpreting a fair range of basic emotions and also perhaps unique behaviors they've accidentally trained or organically developed with their dogs.
What additional things can your tech detect or identify beyond stuff like that? Are you using AI with human coders on training data, and if so, do they need special training or is the wisdom of naive dog-adjacent crowds good enough here?
I thought the other info you shared about canine and other animal vocalizations was very cool and interesting. I'd love to learn more about that any time, from anyone!
> For example dogs don’t really seem to recognize single syllable words well if at all.
I'm pretty sure most dogs understand "no" (and the equivalent word in other languages is often 1 syllable). I haven't done anything approaching a study, but I've trained many dogs in my life and my experience disagrees with you. Some other common words they know really well include: treat, sit, off, here - basically anything you teach them.
>I laughed to think that HUMAN'S analysis of whether a dolphin is intelligent is whether WE can understand their language.
>Using that logic, it seems my dog is more intelligent than me: He can understand some of my language as well as his own, whereas I can only understand my own.
That's a false equivalence, as well as untrue.
It takes intentional effort on our part for a dog to learn even a small subset of our vocabulary. When dogs learn our language, they only learn the few words that we make an effort to teach them, and these typically correspond directly to an action we demand of them in exchange for a reward. Dogs cannot pass this knowledge amongst themselves, each dog has to be taught by a human individually.
Dogs have made no specific effort to teach me, at least none that they managed to make recognizable to me, yet I can tell when they are excited, hurt, angry, &c, but mostly because a subset of body language and vocal attitude are instinctive or socially universal to most mammals (shouting and sharp poses convey aggression, for example).
I believe I understand more dog body language than dogs understand human body language. There is no human equivalent of a wagging tail, yet I understand that. There is no dog equivalent of a pointing finger, and they do not understand it: dogs will look at your finger instead of where you point.
> Dogs are not stupid and they are always very eager to learn and decode what you want them to.
Yeah, many are so clever that they very convincingly pretend they don't know what you want. My dog sometimes does this when he doesn't want to go back home. Being smart and obedient are two orthogonal things.
> We're predisposed to seeing meaning, and perhaps "intelligence", everywhere.
I’m guilty of this with my dog. I can’t help it with her head tilts and deep stares! Her inner monologue is probably less sophisticated than I like to think it is.
> I believe that such dogs do exist, but they are rare.
it may as well be that every dog senses it, and it is our ability to communicate with and understand the dog rarely reaches the level necessary for the information to be actually communicated.
> I've clicker trained my cats to sit and also to follow my finger.
Dogs are so invested in human approval that even abuse can be an effective training 'method' much of the time. As you likely know, before scientific training methods started gaining popularity, this was the predominant tradition for training dogs. And in an anthropocentric way and from a position of utter ignorance, that somehow became the cultural standard. If you can't yell at or pinch or shove an animal and still have it be interested in pleasing you, it's 'not intelligent'.
It's absolutely absurd. It's not fair to dogs or to other animals, and erases the massive, fundamental similarity in the way dogs, cats, and other creatures of all kinds actually learn. All because humans struggle to separate the notion of intelligence from being interested in or useful to ourselves.
> In the future, we have talked about analysing a pet’s body in addition to its face in order to improve the accuracy.
I think that's vital. Most of "Dog language" and even more of "Cat language" consists of body language. Things like pose and tension in relation to the rest of the things in the area at the moment. Dogs and cats are gifted with the ability to "speak human" enough to be understood tolerably well, but their language is demonstrably more complicated.
My dog has a better understanding of context than any AI I've ever met. (I mean that sincerely - since becoming a pet owner it's something I've really marveled at).
In the case of the AI for dogs, I wonder if the person who suggested this owns a dog. If so, it seems rather bizarre. Dogs are not particularly cryptic. They're relatively easy to read if you put your phone down and pay attention to them. Humans and dogs have been communicating for thousands of years. The idea that an AI would have greater canine emotional intelligence than a human is dubious.
>Most animals are way more expressive that the average person is able to easily understand
Is the market the average person though or the person willing to spend several hundred dollars on tech for their pet?
My guess is the latter would mostly consist of people who have spent decades in close contact with pets and are able to reliably read the body language of a dog with rather extreme accuracy and nuance.
If you've never owned a dog they're rather difficult to read. If you've been around them as long as you've been around people you can almost immediately discern a lot of information about them based on a glance.
> "...training dogs has a huge wealth of actionable knowledge needed to interact with humans..."
yah, i couch this as 'dogs are first-order humans', feeling many of the same emotions (joy, pleasure, contentment, anger, fear, hurt, etc.) but more plainly and openly, without language (standing in for our more developed frontal lobes) as an indirection layer.
that (indirectly) leads to the notion that in the best of circumstances, you don't train (or own) a dog, you develop trust (and a relationship) with them, for which they're willing to cede their independence to your judgement, much like humans and leadership. the leader has to prove their worth to the group, not the other way around.
https://scottaaronson.blog/?p=6288
reply