Reasoning from Fictional Evidence

Byron Reese and I get into the uncomfortable gap between what intelligent systems know and what they can explain. Machine learning can expose the biases buried in our data, but it can also leave us staring at an answer without a satisfying reason.

Then we turn to the stories that teach us what futures feel possible: Hollywood’s dystopian AI, Star Trek’s optimism, the real Tricorder XPRIZE, and Byron’s idea of Agora. If we want people to build a better future, we need better stories about technology worth building.

Transcript

Pablos: Hyperspacing back in the conversation: Helen Keller was a human who lived a life for quite a while, longer than most, before getting the language. Because the language isn't something you're born with, you have to learn from the world around you. Tarzan doesn't get it. Helen Keller didn't get it. They had a unique experience in the world of being a human in a modern world but without that language. There's a huge amount of thought, especially for me and you—and I'll say mostly males—that is very much in language. I don't have any thoughts that are not in the language, more or less. Everything I think, I use the language to do it.

But for some people, they rely on a kind of intuition for a lot of their decision-making. They're making a decision informed by their experience, by their knowledge, by what they have in their mental model. It's not embodied in language; it's something else. It sometimes helps to make good decisions, better than I can make with my language and with my logical thought processes.

And not to be too stereotypical about it, but you see it a lot more in women who often know exactly what they think is going on in a situation, but they don't know why. They can't explain it. I've been thinking about this because when you look at what we can do in machine learning, it's creating models that have more data behind the answers they give you, more complex algorithms than we know how to parse or explain. They give us better answers, but they don't tell us why. We can't understand why.

And it's more like intuition, right? Which is much different than an Excel spreadsheet, where if you stare at it, you can understand. You can read the formulas because they fit in a cell. You can understand why those changes—why those variables—have the effect that they do. That's logical progression. And I think that's a little bit more like how I think of it as mapping pretty nicely to how stereotypical males think versus females think, because they have this intuition and they are confident in their intuition.

It's often difficult for me to accept when my girlfriend says she saw a creepy guy. I'm like, “What did he do?” “Nothing. He was just creepy.” That's an indictment. He hasn't done anything. I'm more innocent until proven guilty, but she might be more guilty until proven innocent.

It's because of that intuition, and it makes me question whether there's maybe some epigenetic knowledge that she has access to. There's some process in her mind collecting knowledge over time, because she's probably right. Maybe a creepy guy is dangerous, and I don't know. None of them seem dangerous to me. I wonder why. There might be something beyond the language, another class of tool that doesn't fit—not just language helping them think. I'm wondering if maybe machine learning is an analogy for some of that. We're getting to the point where computers can have that ability too. It's an incomplete concept, but that's what I'm imagining.

Byron: The problem with it is: why didn't I get this loan?

Pablos: Because nobody else who was Black ever got that loan. The AI would be like, “Hey, I kind of think you're creepy.” Exactly. I think this is one of the major classes of contention in machine learning right now: so-called bias in machine learning. My view is that you have to discover those things, because it means fundamentally we failed to communicate our values to the model. Once you learn, “We didn't actually mean don't give loans to Black people,” you have to express that to the model, and it can start to adjust for that. We didn't do it. It's exposing biases that humans had because it learned all that data from the human and what we chose before.

Byron: An AI that is trained to spot a cat doesn't really know what a cat looks like or what a cat is. It doesn't say, “Look at the pointy ears, the whiskers. I think it's a cat.” It is taking all these cats that it has; it chops them up into little—I oversimplified it—two- and three-pixel pairs, and says each of those is a cat. And every time it is fed an image, it takes those same pixels and compares them. And it's only right 50.0001% of the time.

Pablos: But if you do that a billion times, then all of a sudden, that's your neural nets.

Byron: Exactly, and that's how you can be confident. That's a hard thing to convey to people: why did you think it was a cat? A billion decisions told me that. The more interesting question, of course, is how we do it. How do I train you with one drawing of a cat, and you can pick a photograph? Or my favorite is a little kid will say something like, “Look, there's a kitty without a tail,” when they see those Manx cats, even though nobody ever told them there was a cat without a tail. But there's some catness about it.

Pablos: I was really dense and took a long time to accept how important the stories are, because people are driven and operate off of the stories in their heads. I suppose a lot of them are from childhood and whatever they developed over the course of their life, but the stories in people's heads drive their behavior so much. It feels like there are a lot of failures in our ability to get people to appreciate and accept the impact that technology can have on the things that they care about. Nerds, engineers, and technologists have neglected to put enough effort into telling the stories that would help people get that appreciation.

I think people are very suspicious of technology. They have a dystopian view of it. Their idea of technology is largely just their iPhone now. They don't necessarily understand and appreciate the technology making it possible for them to exist in the first place. And so I look at how Hollywood, in some sense, was built as a propaganda machine for the Jewish diaspora in America.

I'm not saying anything derogatory about that. It's impressive and amazing. I don't have any Jewish in me, I don't think, and I know a lot of Jewish history—not just from Schindler's List, but all the stories that I was told over my lifetime about the history and the plight of these folks. And I don't have that same level of understanding or knowledge or history or appreciation for the plight of Cambodians, which might be equivalent in a way.

I'm not trying to say anything about either of these people here. I'm just saying my understanding of lots of other peoples who've been oppressed. Let's say North Korea or Uyghurs in China. I know a lot less about everybody in the world than I know about Jews, because they told their story at a large scale over and over again. That's very impressive.

In my mind, I think part of the job here is to figure out: how do we take on telling the stories about technology successes at a larger scale? Otherwise people's attention goes too much, I think, to recycling and to things that aren't really going to work. Their idea of technology is just Snapchat.

I guess what I'm curious about is: you're four books in—and not just books. You were involved in media that was about technology, other projects like that. What do you think we've got to do? I don't know how successful your books are, but I'm presuming that they're not as successful as I think they should be. What do we got to do to close this gap and get Spielberg to make positive stories about the future of technology?

Byron: That's the thing. I'll pay 10 bucks to watch Will Smith battle the rogue AI robot. What happens is, if you see the other movie enough times, there's a name for what it does to you: it's called reasoning from fictional evidence. You see that and you're like, “That could happen. Seen that before. I knew it. I've seen that before. That seems plausible.” That was the plot of Her. That was the plot of Ex Machina. And so, of course, you've seen it before. It was just in other movies.

Pablos: It's amazing. We have this whole generation. Basically, everybody a day older than me grew up with Star Trek, and they have all these positive visions of how the future's going to be awesome. We're going to go exploring, and we're going to build amazing things and meet other civilizations and become better. Nothing since Star Trek has done that.

Byron: And the thing is: did it just live in a world where there's no competition, and so those stories were good enough? We kind of made that. They had an XPRIZE to make a tricorder; they called it that. The capacity of it to inspire people, I think, is huge.

Pablos: But Star Trek—that was so long ago.

Byron: I agree. I don't know the answer to your question.

Pablos: I found that when I was speaking and I would show off our inventions, which largely were exciting possibilities, at least if I bothered to show them to people. If I show somebody a machine, they get excited about that. They get inspired by that. It doesn't necessarily make a good movie. It's not worth ten bucks, but it was easy to get people excited in that context because nobody else is doing it.

Byron: Utopias: when I talk about utopia, I say the earliest utopias imagined these impossible things, like legal equality between men and women, democracy where you choose your ruler as opposed to hereditary monarchy, places where you could choose the religion you wanted to be in, places where everybody had certain rights no matter who they were. Those were utopian ideals that we are on track to achieve.

A lot of it, I think, somehow being positive has become equated to being naive. That's the challenge, I think. But when you say, “Look, we used to dream these things and we achieved them, or we're achieving them. And now we have different ones. Wouldn't it be something if there were no hungry people? Or if no children died of disease?” That sort of stuff.

But I don't have any particularly good answer, or I would have done it. One thing I tell in the book, though, is Agora: our collective stories that we all tell each other. Our conversations are Agora's thoughts. If speech is us thinking, our conversations are Agora thinking, and the stories we tell ourselves and we tell each other, those are Agora's imagining. And what's fascinating is you can see them change over time.

The Big Dipper is associated with a bear almost everywhere. For some reason in Europe, the four stars are the bear and the three stars are its long tail. The reason that's weird is bears don't have long tails. In northeastern Asia, in Russia, the myth is slightly different: it's a bear being chased by three hunters, and there's a really dim star next to that second hunter. That's a bird that's showing them the way.

What's fascinating is the Navajos in the western United States, who speak a language very much like that one, Ket, had the same story. We know it's not European, or the bear would have had a long tail. They've got three hunters with a helper bird chasing it. What happened is 15,000 years ago, the people that went across that land bridge brought that story.

That was Agora in the very earliest days, looking up at the star and wondering. And then you get these fables. Aesop's fables only come about when we start living in cities. My book, I go through time and show how the common stories of the day are Agora working stuff out.

Pablos: If Agora's thought processes is the neurons—which is us having conversations—then the medium matters. This is one of the things you can observe now: our ability to have productive conversations is stunted, because so much of it is happening in this Twittersphere context, where a conversation gets assassinated in a drive-by shooting.

If you and I tried to have this conversation on Twitter, then a whole bunch with a different agenda can interject and derail it. There have been different reactions, but by and large, a lot of it's self-censorship. I don't want to have a conversation. I don't want to say anything that opens up to that. I don't think the conversations that matter happen in public.

I think this is better. This is a conversation between two people. Record it; we can share it. Anybody who wants to come along can, but they can't ruin it. There's a missing piece to the internet right now, I think, which is that we need a way to do this online.

Byron: When we step back and say, social media is, what, 15 years old?

Pablos: It's totally new. That's what I think too. These are developmental phases, and we're still in an adolescent phase with social media.

Byron: What I take comfort in is that, broadly speaking, there are more people in the world who want to build than who want to destroy. Right now, unfortunately, the technology enables the destroyers to have an outsized influence, it seems.

Transcript

Links

Thanks, we need good co-conspirators.