Kathryn Hume: Chapter 2: How Math Gives Us Meaning

M-A-C-H-I-N-E-Y-E-S-T-I-X-E.

It's Machine Yearning from Assist. It's a podcast where we think and dream about the future of AI, the talking internet, and how we are reshaping our culture. This chapter is an excerpt from our longer conversation with Kathryn Hume, Katherine Hume came to us via Robin Sloan, a great friend of the pod. In this chapter, Katherine does a fantastic job of walking us through the ways language gives us meaning and the challenges facing AI as we wade into the trickiest thicket of all, how all humans form relationships. Katherine works at Integrate AI now and has done product, marketing, investment, and mentored startups. We loved our time with her and we think you will too. Enjoy.

What is the difference between AI thinking and natural thinking? Are they fundamentally different?

I think there's certain aspects of natural thinking and AI thinking that are the same and others that are really different. So one of the big capabilities and breakthroughs in AI over the past 5 years has been in the field of what they call computer vision. So this is basically you show a picture— excuse me, you show a computer a picture of a cat. Or a dog or a glass of wine or like a baby. And basically it could come out and it can give you a label and say, this is a cat. So then you say, all right, so is that like the same way that we as humans recognize a cat? Yes and no. So from the yes perspective, basically these systems are trained where they make a mathematical model. So when I say mathematical model, I literally mean like if you remember back to when you learned algebra in high school and you had like y f, like a function. So some of those functions can be pretty simple, like, y 3x. And unfortunately, that's not complex enough to do something like recognize a cat in a picture. But if you take it and you say y equals like— imagine projecting this out into 50 dimensions, which is a cool imaginative exercise, right? Like we can imagine 3 because we live in them. We can maybe add on like time as a dimension or maybe color and get us ourselves up to like 6. But imagine now having to go up to like 50 million You know, and it's like, my mind gets blown. There's no way for us to even have an intuition. Like, we can't, like, imagine that in our mind. But that's how many dimensions these functions are working with when they are presented with some super complex data. And then they are asked to do something like say what's in that data. So then we say, all right, so is that like or dislike human thinking? Like, it is because our brains are super complex computational machines that can take in all sorts of information and do stuff with it. What we can do that computers can't, though, is if we're shown 3 pictures of cats, we kind of start to see the patterns and similarities between them. And if we see a 4th picture, we can get a sense of what that is. If you're a machine today, you need to see like 50,000 in order to start to get a sense of what catness is. So they can't construct a schematic representation of catness as easy as we can. So we come hardwired with some pretty cool processing tools.

And is so what changed in the world that there's just enough machine-readable data to start training now? And all of our language is online. And so we're putting all this machine-readable text all day into an area where it can be read? Is that what's changed in the world of like— or is it the models have gotten better or the data's gotten better? Like what's changed?

It's a combination of the three, but I think it's fundamentally the first two that you mentioned. So computers are a lot faster. And also the shape of the underlying hardware has shifted. So like the— if you think about like Silicon Valley and what they call central processing units, basically chips that were like sort of running the technological revolutions of the '80s, '90s, early 2000s, they took electrons and they crosses them in a line. And the math that's powering AI does things in sort of crisscrosses. Like, it's like, so it's beneficial to have these lattice structures where electrons can cross paths and move in and out of one another. And those chips actually came from video games. So they're graphical processing units and they were developed for purpose A. And then somebody was like, hey, what if we try this with this AI stuff? And they were like, oh shit, that just went, like, we went from having to take weeks to train an algorithm to being able to do it in like hours and if not minutes.

Once it's trained, is it trained or does it always keep learning?

Depends on the system. So for some of them, like, you know, you sort of do this huge training process and then let it loose in the wild and it's good enough for jazz. Most of the time, the phenomenon that you're trying to predict or model changes. So it's actually really great to start off with a model that, makes a lot of mistakes, but then learn and get like real-time feedback from the world in time.

Can we focus back to language? 'Cause you said what changed about photos breakthrough in the last 5 years, but are there breakthroughs in language?

So there was this big shift over the past 5 years in terms of how statistical systems, machine learning systems are programmed that are enabling them to capture more of the nuance in language. So, Old school natural language processing, which is the subfield in like the overall umbrella term of AI that focuses on language, basically was built on the— so the task at hand is language is a mess. It has so much nuance. There's so many like nitty-gritty details that we process when we extract meaning from some utterance in like sort of sequence of sounds. And in the past, what a natural language processing a practitioner would do would be to take a messy sentence and sort of break it down into parts. So it would be like, "This is the subject. This is the verb. This is, like, the object," right? So things that you learn when you learn grammar when you're in grammar school. And then they would use those parts and sort of put them into a database and start to find regularities on them. So there was a huge step that was the human mind coming in and taking something complex and making it more simple so that it could become tractable using computers.

Do they call that entities?

Yep, there's entity resolution and extraction and stuff like that.

Got it.

And the shift in the new deep learning paradigm is we say, let's go bottoms up as opposed to top down. So let's stop trying to impose meaning onto things and sort of tease it out using structure and just feed in millions of examples of real live sentences and allow like the complexity to stay and not reduce it down into these simplifying assumptions and just allow the computers to sort of pick up those patterns and be able to execute tasks like translation or, you know, sentiment analysis, whatever that may be. And therefore we're enabling, like we're keeping some of the complexity in the language in the model. So it's able to pick up a lot more interesting and nuanced stuff. The trade-off there is that sometimes it's harder for us to understand exactly what the model's picking up on why it's found nuance. So we sort of lose control over the system because we have to in order, you know, we're not using our mind to try to sort of bake simplicity into the process. We're sort of allowing there to be more complexity.

So how does this work? I take millions of sentences of any language or one language?

You normally wanna stick with one as opposed to just sort of random.

Does a certain language work better than another?

Depends on how many training examples you have. So what we're doing is you take, a sentence, and then basically the first pass is to represent it as a series of numbers in what we call vector. And the numbers are sort of projected into space. So you turn it into like a geometrical arrow, if you will, okay? And then what— this is what I found, this is why I got into the field in the first place. It was so fascinating to me that you, when you turn language, sort of process it sequentially and turn it into this number, You can then do mathematical operations on it. Like you can add, subtract, multiply, all the things you do when you've got like, you know, a line in space. And then you, so you do all that processing in the world of math and then you output again sort of a language, like a sentence representation. And it's as if it's been doing sort of language the way we think about it, right? It's as if it's been finding grammatical relationships or semantic relationships, but all that it's doing is it's using math as a proxy for meaning. And miraculously it works and it doesn't work perfectly and it makes all sorts of mistakes, but like we can use math as this approximation space for meaning, which is like, I find mind-blowing. I mean, really so exciting.

What does that mean?

It means that there's something going on in how our minds, if we were able to, like cut open our brains and look in and sort of observe the processes of how neurotransmitters are flitting around from place to place and all these little connections that are made between synapses to make meaning, like we're doing right now, having this conversation, we'd see something that was a much more complex, fancy version than these artificial intelligent neural networks are doing, right? So I don't wanna give the impression that it's like, there's a direct analog between how brains work and how these machines work, 'cause they're not, it's more like a metaphor than it is like an exact replica. But it does mean that there's some sort of mathematical foundation for the way in which we communicate with one another. You know, if we, again, if we could peer into our brains with some special magnifying glass, we'd be able to see.

Wow. And I would wonder like, what is meaning then?

Yeah, for sure. Well, does it matter? Does it matter as much as we think? I mean, in terms of linguistic meaning.

Does it or not?

I mean, I take that from a perspective of it certainly has transactional value. So we're here today because we were able to exchange an email and agree upon a time and a place. And that time and place had a referent. So indeed, 65 Langton Street. Is a place that I can look up on a map and I can come there. And when you asked to do a podcast, that had meaning for me. Like, it wasn't, you know, and it was different than coming over for pizza with somebody I've never met before. So, you know, that's something different.

Meaning has so many variables.

Has so many variables. And so it has a, it's a token of exchange for us to be able to, like, communication is just a form of, you know, being able to do stuff. Together, as opposed to sort of just existing in our own little world where we're unable to actually find a space for commonality.

Is Ms. Sun— oh, go ahead.

But I was going to say, but that's a really different sense of meaning than the emotional meaning of falling in love, of watching a movie that makes you cry, and then you say, like, that's— you know what I mean? So something that is meaningful versus something that has meaning, right, and has sort of pragmatic transactional value.

All right, thanks for listening. Get in touch on Twitter @assist. DMs are open. We're super interested to hear who you think should appear on the podcast. Please subscribe so you never miss an episode and share this with someone who cares about how we make sense of these changing times. Machine Yearning is made by Paul Chufo and Michael Elsesser for Limina House. Have a great day.