Language is not essential for the cognitive processes that underlie thought

One reason might that LLMs are successful because of the architecture, but also, just as importantly because they can be trained over a volume and diversity of human thought that’s encapsulated in language (that is on the internet). Where are we going to find the equivalent data set that will train this other kind of thinking?

Open AI O1 seems to be trained on mostly synthetic data, but it makes intuitive sense that LLMs work so well because we had the data lying around already.

I think the data is way more important for the success of LLMs than the architecture although I do think there's something important in the GPT architecture in particular. See this talk for why: [1]

Warning, watch out for waving hands: The way I see it is that cognition involves forming an abstract representation of the world and then reasoning about that representation. It seems obvious that non-human animals do this without language. So it seems likely that humans do too and then language is layered on top as a turbo boost. However, it also seems plausible that you could build an abstract representation of the world through studying a vast amount of human language and that'll be a good approximation of the real-world too and furthermore it seems possible that reasoning about that abstract representation can take place in the depths of the layers of a large transformer. So it's not clear to me that we're limited by the data we have or necessarily need a different type of data to build a general AI although that'll likely help build a better world model. It's also not clear that an LLM is incapable of the type of reasoning that animals apply to their abstract world representations.

[1] https://youtu.be/yBL7J0kgldU?si=38Jjw_dgxCxhiu7R

Videos are a rich set of non verbal data that could be used to train AIs.

Feed it all the video ever recorded, hook it up to web cams, telescopes, etc. This says a lot about how the universe works, without using a single word.

> One reason might that LLMs are successful because of the architecture, but also, just as importantly because they can be trained over a volume and diversity of human thought that’s encapsulated in language (that is on the internet). Where are we going to find the equivalent data set that will train this other kind of thinking?

Probably by putting simulated animals into simulated environments where they have to survive and thrive.

Working at animal level is uncool, but necessary for progress. I had this argument with Rod Brooks a few decades back. He had some good artificial insects, and wanted to immediately jump to human level, with a project called Cog.[1] I asked him why he didn't go for mouse level AI next. He said "Because I don't want to go down in history as the inventor of the world's greatest artificial mouse."

Cog was a dud, and Brooks goes down in history as the inventor of the world's first good robotic vacuum cleaner.

[1] https://en.wikipedia.org/wiki/Cog_(project)

I always start with God’s design thinking it is best. That’s our diverse, mixed-signal, brain architecture followed by a good upbringing. That means we need to train brain-like architectures in the same way we train children. So, we’ll need whatever data they needed. Multiple streams for different upbringings, too.

The data itself will be most senses collecting raw data about the world most of the day for 18 years. It might require a camera on the kid’s head which I don’t like. I think people letting a team record their life is more likely. Split the project up among many families running in parallel, 1-4 per grade/year. It would probably cost a few million a year.

(Note: Parent changes might require an integration step during AI training or showing different ones in the early years.)

The training system would rapidly scan this information in. It might not be faster than human brains. If it is, we can create them quickly. That’s the passive learning part, though.

Human training involves asking lots of questions based on internal data, random exploration (esp play) with reinforcement, introspection/meditation, and so on. Self-driven, generative activities whose outputs become inputs into the brain system. This training regiment will probably need periodic breaks from passive learning to ask questions or play which requires human supervision.

Enough of this will probably produce… disobedient, unpredictable children. ;) Eventually, we’ll learn how to do AI parenting where the offspring are well-behaved, effective servants. Those will be fine-tuned for practical applications. Later, many more will come online which are trained by different streams of life experience, schooling methods, etc.

That was my theory. I still don’t like recording people’s lives to train AI’s. I just thought it was the only way to build brain-like AI’s and likely to happen (see Twitch).

My LLM concept was to do the same thing with K-12 education resources, stories, kids games, etc. Parents already could tell us exactly what to use to gradually build them up since they did that for their kids year by year. Then, several career tracts layering different college books and skill areas. I think it would be cheaper than GPT-4 with good performance.

"Where are we going to find the equivalent data set that will train this other kind of thinking?"

Just a personal opinion, but in my shitty When H.A.R.L.I.E. Was One (and others) unpublished fiction pastiche (ripoff, really), I had the nascent AI stumble upon Cyc as its base for the world and "thinking about how to think."

I never thought that Cyc was enough, but I do think that something Cyc-like is necessary as a component, a seed for growth, until the AI begins to make the transition from the formally defined, vastly interrelated frames and facts in Cyc to being able to growth further and understand the much less formal knowledgebase you might find in, say Wikipedia.

Full agreement with your animal model is only sensible. If you think about macaques, they have a limited range of vocalization once they hit adulthood. Noe that the mothers almost never make a noise at their babies. Lacking language, when a mother wants to train an infant, she hurts it. (Shades of Blindsight there) She picks up the infant, grasps it firmly, and nips at it. The baby tries to get away, but the mother holds it and keeps at it. Their communication is pain. Many animals do this. But they also learn threat displays, the promise of pain, which goes beyond mere carrot and stick.

The more sophisticated multicellular animals (let us say birds, reptiles, mammals) have to learn to model the behavior of other animals in their environment: to prey on them, to avoid being prey. A pond is here. Other animals will also come to drink. I could attack them and eat them. And with the macaques, "I must scare the baby and pain it a bit because I no longer want to breastfeed it."

Somewhere along the line, modeling other animals (in-species or out-species) hits some sort of self-reflection and the recursion begins. That, I think, is a crucial loop to create the kind of intelligence we seek. Here I nod to Egan's Diaspora.

Looping back to your original point about the training data, I don't think that loop is sufficient for an AGI to do anything but think about itself, and that's where something like Cyc would serve as a framework for it to enter into the knowledge that it isn't merely cogito ergo summing in a void, but that it is part of a world with rules stable enough that it might reason, rather than "merely" statistically infer. And as part of the world (or your simulated environment), it can engage in new loops, feedback between its actions and results.

> However, it also seems plausible that you could build an abstract representation of the world through studying a vast amount of human language and that'll be a good approximation of the real-world too and furthermore it seems possible that reasoning about that abstract representation can take place in the depths of the layers of a large transformer.

While I agree this is possible, I don't see why you'd think it's likely. I would instead say that I think it's unlikely.

Human communication relies on many assumptions of a shared model of the world that are rarely if ever discussed explicitly, and without which certain concepts or at least phrases become ambiguous or hard to understand.

I like your premise! And will check out Harlie!

> A pond is here. Other animals will also come to drink. I could attack them and eat them.

Is that the dominant chain, or is the simpler “I’ve seen animals here before that I have eaten” or “I’ve seen animals I have eaten in a place that smelled/looked/sounded/felt like this” sufficient to explain the behavior?

GP argument seems to be about "thinking" when restricted to knowledge through language, and "possible" is not the same as "likely" or "unlikely" — you are not really disagreeing, since either means "possible".

I agree we are not limited with the data set size: all humans learn the language with the much smaller language training set (just look at kids and compare them to LLMs).

OTOH, humans (and animals) do get other data feeds (visual, context, touch/pain, smell, internal balance "sensors"...) that we develop as we grow and tie that to learning about language.

Obviously, LLMs won't replicate that since even adults struggle to describe these verbally.

GP said plausible, which does mean likely. It's possible that there's a teapot in orbit around Jupiter, but it's not plausible. And GP is specifically saying that by studying human language output, you could plausibly learn about the world that have birth to the internal models that language is used to exteriorize.

If we are really nitpicking, they said it's plausible you could build an abstract representation of the world by studying language-based data, but that it's possible it could be made to effectively reason too.

Anyway, it seems to me we are generally all in agreement (in this thread, at least), but are now being really picky about... language :)

Could be! But then there are ambushes, driving prey into the claws of hidden allies, and so forth. Modeling the behavior of other animals will have to occur without place for many instances.