Click on the links in the side bar to read the parts of the transcript that interest you the most.
Click here to go back to podcast page.
Below is the full transcript.
Setting the stage
A thinking partner
A thinking partner who will make it so easy to think together that enables you to explore fields you never thought you would understand. It is astounding at that. We’ve never had that before. And I think it changes everything. And the good thing about it is curiosity. Go as deep as you want, go as far afield as you want.
JANE
Today I’m with David Weinberger to talk about his new book, Beautiful Particulars. The subtitle is How AI’s Attention to the Smallest of Differences is Reshaping Our Biggest Ideas. The success of AI reveals the truth and power of the small differences, the particulars that hold us apart so we can hold ourselves together and that make our world possible. Listen to this thought-provoking educational conversation.
David, the last time we talked, I don’t know if you remember, I looked it up. You actually talked about at the end of the conversation, you talked about the fact that you were working on a new book. And I think that’s what we’re talking about here. One thing you said in that conversation was that one thing we had to think about was how AI machine learning, and I’m reading now from the script, how AI machine learning in particular will change how we think about ourselves and our world. And you said it’s too early, isn’t it, to think about that?
And we need to wait and see what will happen and then you continue saying, But I don’t feel like waiting and I’m writing a speculative book and its intention is to be speculative. I think that’s a great introduction to what we’re gonna talk about today.
DAVID
It’s s yeah no, it’s still absolutely true today. It’s still way too early to actually be able to predict what’s gonna happen in almost any phase of AI, from regulation to where it will be technically or what effects it’s gonna have on anything. We’re like two days into the AI era at this point.
Too early to predict
Yeah. Well, when you look at a lot of the press and a lot of the stuff that people are writing, there’s just so much crazy stuff about AI. It’s I think everyone’s trying to get on the AI wagon, a lot of consultants and big agencies and all that and trying to sell AI to their customers. And I think what we’re gonna talk about is going to really help people understand more about machine learning. That your book is fabulous for that. Really fabulous.
Yeah, well I think you’re gonna hear compliments for a long time since I know it’s coming out in a few days. And I think you’re gonna be hearing a lot of and getting a lot of compliments about it. In your notes to me, you said that one thing that we needed to do, we had to start with how AI works. We needed some basics about that. And actually before we do that, DAVID, I wanted to ask you the title of your book is amazing. I know people probably will be asking you this all the time, what are you talking about?
And what I think is really extraordinary is the subtitle because you make a beautiful contrast between the title and the subtitle. Can you talk about your choices there in words?
Beautiful Particulars. Why the title and sub title?
So glad you like that because an enormous amount of work went into settling that. But that that’s been true of most of the books I’ve written. The titles go through twenty or twenty-five different stages sequentially, where I think, no, no, here’s a great title for it, because my idea of what the book is about changes as I’m writing it. and I hope eventually I figure out what the book is about by writing it. And then so the title is Beautiful Particulars, because I wanted a title that did not sound like an AI title. And it’s a it’s actually pretty descriptive once you read it, which is not what a title’s supposed to do. But the subtitle is intended then to be much more prosaic in order to let you know that what to expect. For one thing, it is a book about AI. The cover is very non-AI-ish looking too. So it’s all up to that subtitle.
JANE
It’s amazing the cover is extraordinary.
DAVID
good, I really, really like it. The people at MIT Press, the publishers, they did a a really great job. Especially since most of their books are technical and it’s proper that they look technical. But anyway, so the title is I have to read it because I never get it right. Okay. How AI’s Attention to the Smallest of Differences is Reshaping Our Biggest Ideas. And that is actually what the book is about and the use of the word particulars in the title and differences in the subtitle is also actually thematically very important.
The bigness and the smallness
The thing that’s extraordinary about the title and the subtitle is the idea of bigness and then smallness and how smallness takes us to bigness. I mean that’s intriguing. Someone just sees your cover, say walking by a bookshop and you see it in the window, and you wonder how can such a beautiful book, it must be dedicated to nature, must have to do with nature. And in a funny way it does. And looking at the comparison between the small and the big
I think that’s a lot of what you’re talking about in more detail in the book.
We in the west – at the peak of technology
It absolutely is. So I’ll tell you how I view that. and and then I think a lot of it does depend on having a framing of how AI works. Because the premise of the book, explicitly, is that we in the West, which is, you know, what this book pertains to, we in the West for centuries and possibly millennia have understood ourselves in terms of our latest peak technology.
The easiest example is in the age of enlightenment when watches and clocks were the pinnacle of technology and they are still amazing. If you can never go see one in a new in a museum or any other place, they’re astounding of inventions that seem impossible that human hands could craft it. A really, really technical. So then the universe started to look like it was running according to a few simple rules, a few simple gears, but you put together you get enormously complex mechanisms.
DAVID
And I’ll jump ahead to computers, but there are other ages in between. There’s some before, like it doesn’t matter. So in the age of computers, really briefly, we began to understand ourselves and our world in terms of programs that are processing bits. Information in, information out. Even the idea of information was basically new with computing. Like in nineteen forty eight, it got its current meaning. and that took over how we think about everything, from how to organize a business to how our brains work.
I sometimes find myself saying, I can’t process it. Slow down. It’s too much information, which is all computer talk. But it’s if I feel this processing, as I think many of us do, because we’re now thinking about our our experience in terms of that technology. And the same thing with the web, the internet, when everything started to look like messy webs, which I’m in favor of as a worldview, but still it’s a very big change from the tightly controlled computer view. And now we’re in the age of AI, so is it too early to ask how it may affect us? Yes.
Even two and a half years later it is still way too early. But the book goes ahead anyway, in part just to give people a sense of what it can be like to think about how technology affects our really up close and personal ideas about our own experience, understanding our own experience. So that’s roughly what the book is about. But in order to understand how AI might affect our thinking, you have to understand something about how AI or machine learning type of AI, which is a dominant one now, how that stuff works. Because the premise of the book is it’s that sort of understanding, like understanding how a clock works, how computer works, that affects our self understanding.
How does AI work?
Can you say a few words about how AI does work? I mean that’s what people want to know.
DAVID
A few words is really hard. I will do my best, Jane. And you can interrupt me at any point and say That’s too many words.
JANE
I’ll put the clock up and I’ll tell you.
DAVID
Yes, please. Yeah, so AI’s gone through many evolutions since say the the mid nineteen fifties when the term was invented, going down different paths. And around mid two thousands, two thousand and seven or or so, a computer researcher, a young computer researcher named Fei-Fei Li had had an idea because progress had slowed. It was as they say an AI winter. It was it was a small winter, but the times it just it sort of stalls some in its brief history. and so she was thinking and she said, you know what maybe the problem is? The computer scientists generally now are looking at the algorithms, trying to figure out what the right algorithms are, for example, to be able to have a program that can identify objects in a photo photo in a or an image, which is a standard and difficult problem for computers in general and even for AI.
Ì thought, well, maybe the problem is that they’re looking at they’re trying to construct algorithms, and maybe we should be paying more attention to the data. Maybe we should start from the data, which is a real revolution in how we have treated computing, how we’ve done computing in the past. Where if you want to something as simple as a spreadsheet, which is a type of program, you figure out what the relationships are among all the different factors, you know, if it’s a big business plan or something. How much how much are we spending on marketing and the sales and all the rest of it and and the relationships among them and how you’ve built your spreadsheet, extremely useful.
Based around rules and principles and generalizations
But it’s based around rules and principles and generalizations, which is fine. That is what do we expect the formula to be for increase in marketing spend on our sales? And you come up one way or another, you come up with a principle. So this whole thing is a set of of rules and principles. And and that’s in general how it’s been done. So if you’re trying to identify a simple type of image, like the ten numerals, if that’s your task and it’s before this change, then you say, okay, so what is what’s the rule for a one? Well a one is a vertical line. okay, great. How about for that one’s easy? How about four and eight? two circles on top of each other, one smaller than the other and they’re centered. It gets a little harder with the other ones, but still in other words, you are repeating properly so, what your first grade, second grade teacher taught you is how to make a one and a two, and that’s perfectly fine. But if you look at actual examples of the way that people write numerals, you’ll you can look at hundreds of them or thousands of them, and you will find very few of them where an eight is two circles, one smaller than the other and on top of it and then centered. People for all the numerals, even ones, it’s very hard to find ones that are simply upright, most of them lean and some of them etcetera, etcetera.
The ideal of an eight shows up way less than every possible variation of an eight. So if you’re trying to figure out what the number written numbers on a check are, which is part of the impetus for this back when we used checks, you want to make sure you’re getting it right. Money’s involved. But the range of eights is really wide. It’s really hard to specify it. So these programs, old style programs that based were based on rules, they worked okay, but not not great. However, so this is the AI thing.
I knew I’d get to it eventually. So with machine learning, you would instead gather up lots of examples of of numerals, written numerals. That’s the data that that Fei-Fei Li was talking about. And instead of saying, okay, let’s impose the second grade version, the ideal instead you say, okay, let’s let’s have the computer figure out which pixels, because these are digital images, so it’s based in pixels, right? let’s have the computer figure out how each of the pixels and patterns of pixels for each of these thousands of examples we’re going to give it, how well each of them nudge its decision towards a number, a numeral, particular numeral.
Patterns not knowledge
So you look at it and and you really you’ve done the analysis and it sees that, for whatever reason, numeral eights, as written by real human beings, tend often tend to have a little bit of a sprinkle of pixels in the upper left that aren’t part of the classic definition, but some of them have that.
And many of them have two spaces, you know, well, lines and closing spaces. Not all of them do. Many of them do. The shapes of those are highly irregular, but that’s that’s sort of an indication, if you have that, depending on how well defined, then that also help nudges it into the eight box. The computer is just looking for patterns, it has doesn’t know anything at all about our second grade teacher, it just is looking at dots.
And if a pattern helps move it one way or another, then that’s going to be part of the algorithm, so to speak. And so and and some of the patterns you may not be able to figure out. Okay, three dots here, one there, another what we might call a curly line here, and that makes an eight and makes no sense, but it makes it two percent more likely that it’s an eight than it’s a two or a three or whatever. It turns out and some of the I keep
Stopping short of saying this, some that we don’t understand at all why it matters. It’s completely counterintuitive, but the computer doesn’t have intuition. So it it just compiles all this mass, and now when you feed in a numeral, it looks at it looks for patterns in it. And if it has patterns that nudge it one way or nudge it the other, then it says, okay, this is an eight, no, this one’s a three, no, I can’t tell you why. Now with numerals, it’s sometimes i it can, but when you’re talking about identifying animals in the wild.
DAVID
Often you simply cannot, because there’s so such variety in photos of an elephant in terms of the background foliage, its position, how much of it is being shown. there’s so much information in a real image, a real photo, as opposed to a single numeral. Thing that for me is really interesting about this, and I think in the book’s bet such as it is, it’s not really predicting, but you know, but the speculation is based on the idea that from this instant the more that we understand as a culture how the basics of how it works, the more likely we are, I think, to start thinking about our experience in terms of not rules and laws that explain it. Because the rule is two circles on top of each other, but rather on the patterns that we find in vast amounts of particulars, where particular is a thing, I mean it in the normal sense, right? Particular is a thing in its differences from other things.
Particular is a thing in its differences
DAVID
Yes.
So if you say, no, no, I want that particular donut, and you point at it, what you’re saying is, no, that one’s different from the other ones in whatever whatever way. It’s got chocolate on it or whatever. I want that one. So particular means it’s understanding things and their differences. The deep thing that’s happening in this sort of machine analysis, the way that it works, is that when you when you’re doing a rule based program, which most programs are computer programs, and they work very well, of course, when you’re doing that, you are
When you put an input in, you are reducing it to what the computer is looking for. Right? So in the case of numeral analysis, you’d be looking for the dots that compose two circles. With AI, what matters more than anything is not what two things have in common, but what differentiates them. When it’s doing the analysis, it’s trying to find the pixels in an eight that help differentiate it from all the other numerals. And that’s finding
Yes.
DAVID
Ultimately I think finding meaning in differences and telling us something about what we have lost by focusing so hard in our knowledge and understanding on the rules and laws and principles and second grade teachers Lord love them for sure, who are properly but still reducing ones to r reducing eights to perfect circles. Differences become really, really important in the age of AI. That’s my contention.
JANE
That’s a point you make strongly in the book, at different points in time with from different angles. it’s like the differences are more important than the commonalities.
The voice of particulars at least is being heard now
Yes, or at least there’s a balance and we have so relied upon and valued the commonalities that the voice of particulars at least is being heard now. We also need the rules and the universals, you know, Newton’s laws are still right. I am reluctant to go up against Newton in this very short any any one. but differences in particulars gain a voice.
And that I think has huge I think generally positive effects on how we think about things.
JANE
In the book you go into different aspects of particulars and different aspects of AI, and as you said to me in the email, we can’t cover everything. I’ll confess to my listeners that I sent poor David Weinberger an email that would have kept us online for five or six hours minimum.
DAVID
I they were fantastic questions. It was a really, really good list.
JANE
And you wisely cut them down to what you consider to be the essential of what we could cover in a first conversation. I played my role and you played your role. Perfectly. So we’re completely in sync.
DAVID
It’s always good working with you, Jane.
JANE
One thing you talked about that I found fascinating in your book is fairness, because fairness has always seemed sort of obvious as a term. Something is fair or it’s not fair. But that’s not the point you make very clearly. Can you can you talk about that a little bit, David?
Yeah, no, I think that’s a really that’s a really good first example. because exactly because, as you say, fairness seems so simple. In fact, it’s so simple you can teach a four-year-old granddaughter, let’s say, what fairness means. and you do it without even realizing you’re doing it. But the made up example in the book is it’s the four year old’s older sister’s birthday party and the four year old complains that the older sister, let’s just call her Ellie for no particular reason.
Fairness seems so simple
got a larger piece of cake than the four year old Poppy got. And Poppy, who is an adorable and very sweet child, by the way, is just she’s on the verge of crying. She doesn’t she feels hurt. She got a small why did Ellie get a big piece? And so you say to her, something like, you know why, El Poppy? It’s because first of all, Ellie is bigger than you. She’s older and bigger than you, and you know the bigger you get the more you eat. And second of all, it’s her birthday, it’s her special day, and when it’s your special day, and when you’re bigger you’ll get bigger slices as well. And so you just explain fairness to her and and sh she’s still gonna cry until you give her like a tiny little sliver more to say that now you the same amount and she’s happy. But what you’ve told her is that fairness means treating people equally, which you have not done in terms of cake size, unless there is a relevant difference. And you gave her two relevant differences size and birthday party.
And so she is not saying she’ll be able to tell you what fairness is, what the principle of fairness is, but she knows, you know, you you’ve instructed her. Yeah. and so fairness is so simple that in the Western philosophical tradition, it is not it’s often ignored when talking about morality, because it’s too simple. I mean you got nothing to say about it, it seems. Although some philosophers have talked about it. Philosophers are much more interested in in justice, which is sort of a much bigger term and more complicated and involves this the state, et cetera, authorities.
In any case, so fairness turns out to be somewhat harder than we thought, a little more complex than we thought. But AI is is often criticized for being unfair, quite properly, because i it doesn’t have a sense of fairness, right? It doesn’t have any of that stuff. it’s working off the data.
AI can be very biased
So first of all, AI can be very biased, especially the machine learning to learning from human examples. It will if the data, that is if the culture is biased in some way, and you can see this in in the large language models, the chat bots, as well frequently, if the data what that we have produced in terms of writing for the chatbots is biased in that a reference to a nurse in much of what it’s been trained on is likely to be followed by the pronoun she. She. Yeah. And likewise, you know, doctor. And that’s in the past twenty years that’s gotten much better, but it’s trained on a very broad set of stuff. And so it’s very likely to pick up on the statistical relationship between she and nurse. And we’ll mirror this unless it’s either corrected or trained on, you know. And it gets much worse than that. It can be much, much subtler than that as well. I mean, and there’s a lot of work going into trying to detect this and to write it, to to you know well, to correct it. But it’s really hard.
Worth a very interesting field, well worth doing. So it’s it’s good that AI is being critiqued about that sort of fairness. But there are other issues as well, not just the bias one. It’s let’s say you are using AI to evaluate applications to your college. You’re on the team doing that. You just want to take a first cut. Maybe you want to decide entirely, which would be really not right, but at least a first cut, you got, you high volume of Trim it. You tell the AI. You’re in charge of the AI, by the way, Jane.
So but being the conscientious person that you are, you notice that the AI is still favoring it seems to be favoring the male applicants, because sixty percent of the males who apply get into the next pool you know for consideration, and only thirty percent of the women do. And so you’re concerned about this, as you should be. and you start looking into it. This is because when AI is being trained on data it does not care. It only cares whether the things that it’s noticing, the differences that it’s noting, help it make more accurate decisions. It doesn’t care whether we understand them or not.
It’s very likely that it’s hard to interpret what’s going on. But you’re gonna be thoughtful about it and you look at it and you go through a set of proposals to the board or or whatever, which is also at least nominally concerned about this. And you say we can manually adjust the cutoff points and we can make it harder for a man, for you know, a male to get in.
And I can show you how that would work. You can basically just turn a dial. It’s way more complicated than that. But you know, you can do that and say if we did that, just adjust the difficulty, we can get the whatever you want. We can get the 50-50, we can get the, you know, we could make it all women if that’s what you want to do. You just raise the the peak so high. And certainly you will get pushback on this idea because somebody, maybe many people will say,…
well, no, now you’re biased against men. You’re purposefully applying different standards to men and women. I think there’s some very good answers to that. That’s not the point of what I’m saying. And you will try other there are other ways of college application app fairer. And these are ideas of fairness that we didn’t have or at least hadn’t formalized. And so a AI is leading us to an understanding of just how complex fairness is. Yes.
How are we training for success?
And that’s a really good thing. I’ll get one more little step in this because at some point somebody is going to ask but one of the very right questions I think, which is, How are we training this? And you’ll say, well the way that you wanted me to, which is you gave me a set of student applications, you know, thousands of them. You noted whether they got not only did they get these are all people who were admitted. I’m not doing a good job designing the AI, but this is I’m gonna continue anyway.
And you’re gonna say, Yeah, but you also told me what constitutes being a successful student. Because if you only look at who got in, a lot of them flumped out and the rest. So so I absorb that data and you or somebody should ask, maybe the problem is our idea of what co constitutes success. So we said compare these applications to successful students and we define s because you have now dis define success, a the person graduated and grades are indicator of success.
And we checked up a couple of years later, are they employed? Did they go to graduate school? how good a job are they in? Fine, but you know, dropping out of college when you should drop out of college for good reasons, that’s not a sign of failure. If you’re going to support your family in need, if you discover that you’re not this college isn’t serving your needs and you want to you now know what you want to do, that’s a huge success. I I would think. Somebody said, No, I wanna actually want to be a musician or I wanna go to a science school or whatever.
How do you decide what constitutes success? I don’t know in college. And there isn’t an objective standard. So AI fairness is bringing us to face issues we never really had to face head on before and make concrete decisions that a machine could understand. And I think that’s a really, really healthy thing to know, to see how complex and arguable, even relatively complex ideas about fairness.
Success and relationships with others
That’s really, really a good example because one thing leads to another thing. Like fairness leads to what is success. And we could probably we won’t do it now, but could take what is success will lead to your relationship with other people, for example. I mean you could go on and on. I don’t know how we as people can manage to do that, let alone machines do that.
DAVID
We already do, but we tend when we’re thinking about that experience to underplay just how complex these situations are. Because we’re not that great. Humans are not that great at complexity and especially when things are incredibly complex we can’t hold it all in our own head. Fortunately AI doesn’t have a head to hold things in, so it can just you add more processors or and memory things. So and I think success is actually a good example of this. I I think that’s an example it can be an example of fairness for sure.
Dimensions for AI
DAVID
But it is, I think, always an example of our tendency, human tendency to be low dimensional in our thinking. By which I mean and this is another term or concept that AI forces on us if you want to understand how it works. And I and I hope that it will become part of more of a part of our jargon. So for AI, a dimension is basically and the AI people won’t agree with this entirely, but that’s okay.
Is basically a factor that’s being considered and that can be measured. So if you’re trying to figure out how much you should be setting your sale price for your house, because I’m coming from a generation that still was buying and selling houses, you might say, well, one of the factors, what are the dimensions, is how many bedrooms. So I’m gonna do some research in for my area. I’m gonna look up get data about three bedroom, how prices correlate with sale price.
So on one axis, I’m gonna have sale price, the other one I’m gonna have the number of bedrooms as the first factor. Yes. And I get I get some information that way, but people don’t buy just on the basis of bedrooms, number of bedrooms. So you start adding you add bathrooms. So now you have a 3D chart, which is really hard to read, but and then you add all of the other things, and there is no list, complete list of things. Some, you know, lawn size, how good the school system is, depending on your circumstance, you know, if you have kids or not, and so on.
DAVID
When people look at your house, they they’re looking at things that you would never put on a chart because they’re dimensions to them. it might be how steep the stairs are. Well, that’s relatively measurable. It may be which way the sun is coming up. The sun strikes they’re visiting and they see the sun coming up through one of the windows. that’s just so beautiful. That is not going to be on your list. Neither is a million other things. there’s a coffee shop down the street. it’s near a big city they like.
There’s no end to the number of dimensions. And if you could get all those dimensions together, you’d have a a chart that had so many axes, so many dimensions, you could not make sense of it, but machine learning could. Yeah. And so we can’t handle a lot of dimensions. AI can handle as many dimensions, thousands of them, in some cases millions of them, when it’s considering all the factors that are involved. And this makes a single dimensional value like success, as we often use it, seem really just it’s not a rich enough idea.
So if you’re saying somebody’s a s a success because they made a lot of money, which is a perfectly reasonable thing to say, okay, but what does that tell us? How about the other things that matter in human life? There’s no shortage to them, but maybe we should or if you’re running a company and you’re only seeing the bottom line, you are gonna i i th life is gonna be hell for just about everybody. You’re not valuing, not seeing success in anything else that matters, including human values like we know are your employees leading reasonably good lives being employed by you?
What is morality? Human caring.
You know, it’s it’s almost an endless conversation, isn’t it? I’d like to actually turn it in the direction of something which was you talked about morality when you said that one part of moral life in AI is its embedded particularity and on the other hand it does not have what you call embodied care. AI can have plenty of particulars, but it cannot care, not have care embodied in it.
And you go into a lot of examples and reflections about that in the book. I don’t really know where to start. I mean…
DAVID
I think that’s a really excellent start. So moral morality is another a good example of universal principles being applied when in fact the situations are way more complex and you in applying them we often have to reduce the situation too far.
When you’re teaching history of morality in a philosophy class, generally you’ll do a history of it and you’ll point out some of the historic ways of doing this, and one will be because it’s still present with us, a principle based morality, in which you have a set of rules, whether, you know, don’t lie or ten commandments or way more specific ones. And that’s how you evaluate what to do. You but you’re then going to move on to a different philosophy of morality, and you’ll get there by showing all examples of the problems with principle-based morality. It le can lead to cruel decisions, depending on how strictly you apply them. Because just life is is more complex than that. I’m I I’ll I can give you a r really a classic and simple example from a couple hundred years ago of a principle like do not tell a lie. Good principle. It’s it’s a commandment. and the the the standard response, which actually comes most seamlessly from Kant, but not exactly these terms, is okay, a murderer comes to your door and says, Have you seen a a tall man wearing a yellow hat running down the street because I want to kill him and the murderer is holding like, you know, a knife in his hand. So you can make it as extreme as you want. Do not tell a lie. you really should tell a lie. Or this says is do you have this person in your house? And you do. Principle based would say, No, gotta you gotta
Telling a lie or maintaining a principle?
Tell the truth. Any normal person would see this as you know, the issues get much more complex as they get more realistic. Yes. Even all the way down to how you decide to divide up the a a a shared meal meal bill, which is not nearly as, you know, consequential. So the problem with principles in general, and also there’s they work because they apply to so many cases. Don’t tell a lie, okay, throughout your life, every situation, which is a strength of it. The weakness of it is the particulars of the case get dropped out. So there’s a type of philosophy, it was a feminist philosophy, goes back usually dated to mid eighties, early eighties, and it’s become more and more influential. It opened up moral philosophy to me in a way that all of a sudden actually made sense.
JANE So what’s that?
DAVID I don’t have time to tell you that. I will tell you that so there there is a standard problem in philosophy but also in life, which is okay, we have moral we use the word ought to say that we mean something as a moral statement. You ought not lie, or you ought to give to charity. And if I say, No, that’s right, I people ought to give to charity and you say, Well, how much do you give? and I say none. I don’t give the charity, then you your likely response might be well, you just said ought. You don’t know what ought means then.
Because if you know you ought to, then you ought to. and where that word, it’s just a word, where that word gets any power at all, has been a philosophical issue for a long time. So interesting thing to think about. So the ethics of care solves that problem instantly. So one of the founders of it, a woman named Nel Noddings in nineteen eighty five or so, said, When my child is crying, I go to her, not because I ought to, but because I must, I have to, I must go. And so that just snapped ought off the table for me, which made me very, very happy. And it explained why this phenomenon of feeling you have to exist. It’s because you care about that person.
Ethics of caring, particulars and the importance of relationships
That’s an outcome not of morality but of human caring. Which is for me was I mean it’s so obvious, but it didn’t make its way into Western philosophy to any significant degree. Ethics of care has a lot more to say about the nature of the importance of relationships to in understanding anything but in particular understanding our moral senses. Leaves lots of problems open. But it is such a to me it was such a clarifying truth. I I don’t feel proud of saying this. Something that I thought I should have known forever. And i it’s not surprising me to me that this came from women philosophers, not from men. One of the important messages, things we learn from from how AI works is that principles and other generalizations hide particulars. Real morality sh real morality shows us this because it is all about the particulars of the person, their situation, and our relationship to them. That’s the connection that I hope that AI helps move move us along.
JANE
It’s the particulars that move us towards a certain action. And we could talk about this more, but in fact I wanted to talk a little bit about knowledge. You talked about knowledge a lot. Different kinds of knowledge which I’d never thought about before and I found that fascinating. The different kinds of knowledge that you talked about. Sort of I think the basic ones, the what did you call them? The principal ones, and then the ones with practice, and then you had a third one.
DAVID
Tacit knowledge, just sort of bodily what things your body knows how to do.
JANE
Yes, yes. Can can you talk a little bit about knowledge? In particular, one thing that you talked about is knowledge through conversation, which is interrelationships with people, interactions, I think that’s what you’re you mean, versus or in complementary to written knowledge, reading.
Machines know nothing about the world.
One of one of the issues, very well known issues with machine learning, AI, chat GPT, chat bots and the like, is that they don’t – they hallucinate. And they do. It’s gotten much better actually. But that is, they do much less making stuff up out of nothing, or out of s statistical correlations that actually don’t represent any represent anything in the world because they don’t know anything about the world. All they know are what we have written about the world, because they’re trained on what what we’ve written, and they don’t understand words at all. They have no concept of words, they don’t have concept of anything, because they’re not conscious, but it’s all just numbers to them. Which means that especially the conversational AI, the chatbots, are prone to just getting things wildly wrong because the stats they’re on the other hand, we can think of it as book learning. You know, we learn only from what’s been written. and that’s not always such a terrible way of learning. I mean, much of our learning is book learning.
Nevertheless, the problem is really difficult and very dangerous, and they’re trying to address it as well as they can. But even so, even if chatbots get fantastic fantastically reliable, it still has really disrupted our traditional little over two thousand years idea of knowledge. One is that knowledge is certain, that is the definition of the classical definition of knowledge. It’s statements that we know to be true and we can explain why they’re true. With AI, we don’t know why they’re true because we can’t understand its processes in many, many cases. And it would be exceptionally difficult in in many, many cases to try to figure out whether but we accept it. And so this is truth that is based upon probability. Well, that’s exactly what the initial originally the idea of truth twenty five hundred years ago was intended and has been the classical one. It was intended to get us away from doubt and probability.
Now we’re relying on it every time we use AI and I think we’re gonna continue using it. And it’s inexplicable where part of it knowledge have always been in has been that you can justify your your belief. With conversational AI, which I love and am very not literally, by the way, and am very fascinated by and hopeful for. But one of the huge problems is that let’s say you have a really great conversation with AI about something be very technical, it could be whatever it is.
You couldn’t have had that conversation with any human being because you’re pulling together information, it’s pulling together information from multiple fields and you know it these things are unbelievable. so you get a little bit smarter, you learn something, that’s great. But where where does it go? What gets built on that? And the main advantage of writing, well there’s more than one, but a really important advantage of writing things down, which we’ve been doing for about twenty three hundred years.
Writing? A permanent record to build on.
Is that it makes a permanent record and we can build on it. That which is what we do. you know, we we can dig back if you’re doing research on something, you’re gonna go to the current current writings about it. And if you wanted to, you could trace those back through layers and layers and layers of time and see how we got to the point where this paper could have been written. we generally don’t do that because we’re confident in the system of experts and authorities that what we’re reading in you know a good journal or nonfiction book or whatever, we’re confident that it went through a process by which that made it reliable. Sometimes they’re wrong, but we have complex processes for authenticating and and verifying knowledge.
But we don’t have that. That’s gone. We still have it. People are still writing, right? But insofar as we come to rely upon chatbots, that’s gone. That’s not the source of their information. The source is the statistical relations among words, some of which are going to come very likely from sources that either disagree or disagree because they’re wrong. On the positive side, and I think it’s a tremendous positive, that’s why I’m so excited about conversational AI, is that you can ask it dumb questions and it doesn’t think less of you because it doesn’t think and it doesn’t even make snarky comments. You can turn off its built-in calculated support of you as if you you were the smartest person they ever it ever met.
Conversational AI
You are so what you just said was so brilliant. my gosh. I mean they will literally do that for stuff that you know no. No, I don’t even think it was a good idea or is commonplace.
JANE
Why does why does that happen, DAVID?
DAVID
Because the makers of them want you to like them and stick with them because they want to make money. That is the sum total of it you can turn them off. You can turn this off. At least some of them. And with Claude which is a popular favorite these days. and you do this by saying, stop praising me, or only do it when you think it’s actually deserved. I don’t like being praised by a robot or whatever you want to just instruct it. And if it forgets, you can go back. You can say, Remember, don’t praise me unfairly. And it might respond by saying, No, I don’t think I’m I know that. I’m not being unfair in this case. Here’s why. That would be very nice. That would be a really great compliment to get. But the BS ones that they do in order to make you feel good about the product is just disgusting. It’s and a real danger.
I mean you don’t want a truth teller. You’re relying on something, including a book or a teacher for truth. You don’t want to have the sense that they will overly that they’re not giving you an honest reaction to your responses. It’s a big problem. However, having said that, you can they won’t think you’re stupid, they will say you’re too smart unless you tell to stop. They will spend all the time in the world that you want, then you’ll be paying for it more and more over time as they ramp up the profit making with their trillion dollar stock offerings. They’ll take all the time that you want. You can go away for a while, come back, pick up the conversation, you can branch it in any direction, any field. If it if you want to, you can go from driving directions to quantum physics. And it will be it will give you generally quite good answers.
A thinking partner to explore new fields, an open invitation to curiosity
DAVID
And if another really important thing that does is if you want to read something in a field and but you don’t know enough about the field, which happens all the time with me in every direction. And I so I have some article, looks like it’s gonna be interesting. Let’s say it’s a scholarly article or a technical article or whatever. I can just give it to one of the AIs and say, What is this about? And will give me to a fantastic summary. If I don’t understand the summary or a line of it, I can say, I don’t get that, well tell me more. If I say, this sounds like this reminds doesn’t remind this reminds me of some other topic in some other field in quantum mechanics, it will help you along that path. It is an open-ended machine, not for answers, which it can do, and it it’s dangerous if that’s all that we use it for, especially for trying for a student.
It’s an open invitation to curiosity. It is a temptation for curiosity.
JANE
Can become a thinking partner who will make it so easy to think together that you’ll be able to explore and to explain things
that it will be able enables you to explore fields you never thought you would understand. It is astounding at that. We’ve never had that before. We’ve just have never had it. and I think it changes everything. And the good thing about it is curiosity. Go as deep as you want, go as far afield as you want.
Risks? Dangers of a mainstream.
JANE
Do you see dangers in it other than what you just talked about? Dangers or risks? Because it it sounds like magic. It sounds like magic. Are there risks other than the ones of overly praising you? Are there other risks?
DAVID
Yes, for sure. And one is hallucinations. That’s which is why a responsible chatbot at this point will tell you if you ask it medical question, will suggest that you check with the doctor and tell you that it’s information is reliable. I mean that’s that’s good civic behavior. So one of the dangers is an odd one that I haven’t seen talked well, okay. So especially before the internet, there was a thing called the mainstream, mainstream news. it’s well defined, you know, three stations in the US.
Twenty two minutes a day, thirty minutes a day. It was three white guys who were very, very good. I mean but still they, you know, they had products of our culture and situation. And they almost always agreed in what they thought was important story and how to cover it. And that was in many ways, that’s great. I mean it’s wonderful to have a mainstream, but we also know and have been shown the dangers of having a mainstream in that it’s always gonna be by definition centrist.
That’s the measure of being centrist. When you go to an AI and ask it a question like is the Earth flat? All of the ones that tried will say no, it’s not flat, and then give you, you know, seven, eight, ten reasons why. starting with the simple ones, and then going to some science. Which is pretty much exactly what you want. It doesn’t say, No, you idiot, what the heck is wrong with you? What what a stupid question, which would be my what I would be thinking. Instead
Yes.
DAVID
Takes it seriously and gives you reasons to and that’s great. But what do you want it to say when you ask it, should I should I get a COVID shot? do you want it to say, like in the case of Earth is flat, do you want to say, yes, you should get a COVID shot, they’ve been well proven, et cetera? Or do you want it to say, well, there are two sides to this, or to say, get a go COVID shot, but you should be aware that some people disagree with you and it gives you the arguments against COVID.
The vaccinations and how far do you want it to go? Do you want it then argue with those there isn’t a good answer to this, I don’t think, except okay, give me the mainstream answer. That’s the best answer. If you’re gonna give an answer, that’s the best one to give. That’s the best one to give. But I I’ve been totally in favor of vaccinations, COVID vaccines. Don’t get me wrong. But it’s not clear to me how much I but I’m out of the mainstream on a whole bunch of topics. I don’t know what sort of answer I want, I don’t know how it should handle that.
And I certainly don’t want the corporate entities deciding how it should handle it. It is absolutely in their interest to always have it give mainstream answers unless it has reason to think that you are a wacko or whatever. And that is terribly reinforcing for a a relatively conservative path for thinking, which is better than the some of the wacko stuff.
JANE
Who can determine that? It’s not a question of a a person or an entity or a government or a I don’t know what it is. It’s some sort of flow, wave, spirit, I don’t know. I don’t know how to describe it.
Wikipedia and conversational knowledge takes us to the source
DAVID
Me neither. And I don’t think it’s soluble. So I don’t know what’s gonna happen. A and none of the answers are are great. I that I can think of. Wikipedia gives us a model and this actually brings us back to conversation. so one way of doing it is for the machine to learn about us as we’re using it, which it doesn’t actually do very well, even though people are very concerned about privacy for lots of reasons. but have it learn about us and realize that no, we are anti vax folks and it matters a lot to us ’cause from the questions that we’ve had, so it should respond to vax questions with negatives.
I don’t I’m not I I don’t know how to resolve that. I’m not happy about that.
From another point of view. so Wikipedia manages this problem incredibly well, I think. Not perfectly, because there isn’t a perfect.
Yeah. Yeah.
JANE
Wikipedia is an amazing creation, really amazing.
DAVID
I am still flabbergasted by it. Seven million articles or topics now in the
JANE
Yeah. I’ve I’ve had the person Florence Devouard, whom you may know, who replaced Jimmy Wales when he was no longer he stepped back as the head and she was the number two person at the time. She became the number one person. I’ve had her here. In fact I invited her to my group in Paris. I used to run a group of people in Paris on a particular topic, and she came and talked to them and very, very interesting. And so I have really paid a lot of attention to the Wikipedia because I know her
And I’ve heard I’ve just done a lot of research about it and I’m glad to hear that you like it. I think it’s in probably one of the best intellectual. That’s not the right word. Literary or knowledge. What what I don’t know how to describe it. Invention.
DAVID
Yeah. It’s astounding. It’s astounding. And it’s the processes are very complex because th they should be, they have to be. But the so they have a as you well know, and I think everybody knows this, that they will have an article and they will only use the everything is is footnoted, it’s all backed up. That’s right. and they will only use journals that have are counted as expert in their fields.
Yes.
Right, so they will not accept even if you’ve discovered an important new fact or something, they won’t put it in Wikipedia until an authenticating body okay’s it. Like you write something that gets published in it, right? which diminishes the freshness of the information sometimes, but greatly advances and eliminates arguments. manages arguments by saying, Okay, we can’t decide on whether this particular protein mixed with that is gonna do this.
We’re just Wikipedians. We’re not the world’s experts. So we have to wait. Once the world’s experts agree with you that it’s doing this, then it goes right in. but so it it doesn’t it doesn’t give flat earth any credence. It does explain why people think that and why they shouldn’t. But they also show their work. And their work is first of all the references, but second of all the conversations that shaped it. Yes.
Yeah. And most people don’t read them and they can be hard to read because it’s very inside baseball among Wikipedians. AI could actually help make those conversations readable. Something I’m gonna gonna suggest to them soon. I’m talking about I’m sure they thought about
Yeah.
JANE
That would be good.
Yeah. Well they know that AI takes a lot of their scrapes their content.
DAVID
Yes, which I’m very happy about, right? Because it it is one of the most reliable sources and it’s all footnoted. So that’s
Yeah.
JANE
I think that the problem with the the Wikipedia is still geographical balance. Yes. There are still areas of the world that are not represented. But we don’t want to get into Wikipedia today. We could talk about that for a long time. Anyway.
AI conversations is where the future is going for knowing
Anyway, Wikipedia has the conversations. Yes. if they’re more legible to us, more understandable by outsiders. And if Wikipedia could draw people into those to see how knowledge, at least one way, knowledge is made, and it’s it’s made in part by people conversing with one another that actually benefit the conversation. The conversation is the thing that counts. AI is teaching us that as well, but is not as good or as dedicated as a discussion between people who have a different set of values, so to speak, that are than what’s built into AI. And so trying to get AI conversations, which is where I think the future is going for knowing, helping them to get better at the human values of conversation, and then finding ways in which the outcome of those conversations or the conversations themselves can be preserved, the way that we preserve written things.
So that we can go back in time and say, how did that how did that crazy idea come up? And then be able to address it more directly. That would be that would be huge. Without that, moving so strongly to conversational knowledge development, which I again I think is wonderful and deeply human, is also dangerous.
How David Weinberger and the Future: Ethical dilemmas to come?
JANE
Yes. I’ve got a few I’ve got a few questions I want to ask you that are not directly related to your work, but I wanted to get your view on certain things.
What ethical dilemmas do you think we will have to face as we move forward in our AI era?
Many of them are known already, incl including the misalignment of corporate interests with human and cultural and civilizational interests. Bias as well, the reliance of the all of these things are the ones that are well known are pretty well known, and there are bunches that aren’t yet. I can’t predict. I cannot predict beyond what is already apparent. I I just don’t have that capacity. I don’t know enough.
Younger generation better than us?
Another question I have that maybe you will have an opinion on since I know you have a grandchild or two, I don’t know how many you have, if you think there’s something that the young generation will be better at as they grow up than we have been.
DAVID
I think and I really, really hope that the way that AI works, looking for the differences that make things what they are, just as your differences from everyone else make sure who you are, that this will become the way they see the world. That is that they can value differences at least as much as the similarities among us and among things. That to me would be an epical change and one that is very, very liberating and hopeful.
Optimistic or pessimistic?
Sure is. Overall, do you feel optimistic in general about the future or more pessimistic?
I think it’s a struggle ahead and I we have shown over the past twenty years at least that we’re not very good at dealing with the important struggles in ways that are designed to alleviate them rather than just sailing through them while we can. It’s what every guilty boomer feels.
JANE
Yeah. Well that’s a good answer. Really, I mean it seriously. I know. That’s a very good answer. The last question I have is you’re probably not going answer this one. I bet you’re not going answer it. I dare you to answer it.
What is going to come after this book?
What what what is your next I’m sure you have a I don’t want to say another step ’cause I don’t like the idea of steps because they’re linear, but what flow are you on? What is the next thing that you’ll be working on once you get beyond the launching of of Beautiful Particulars.
DAVID
I only have a tiny itch of something to write about and I don’t know if we’ll develop into anything. I like writing. I mean that it’s an activity I enjoy. But I’m also right now, say deeply involved in a project I can’t talk about for crummy business reasons. but that’s about trying to address some of the issues with conversational AI. And it’s really fun to be back working with engineers. Software guys and gals on on a project at this stage. It’s really fun.
JANE
Well, I’m eager to see when will we know more about it? Is it like a year or two from now or
DAVID
it’s a year maybe or never. We we don’t know. You know, it’s a very early stage. But it’s a really promising idea. Not my idea. Tech guy I’ve known for a long time
A big disclosure!
Okay. Well, I was right that you’re not gonna answer it, but for different reasons. Those are justifiable reasons. If you can’t talk about it yet, you can’t talk about it.
DAVID
It’s more than I talked about it before. So for me this is like a big disclosure. You’re right there, of course.
JANE
David, do you have any last minute thoughts or anything you’d like to ask me or anything you’d like to talk about?
I’m pretty talked out, which is a sign sign of a good conversation, so thank you.
I really appreciate it. It’s been very interesting. Great talking.
DAVID
Likewise, Jane. Been too long. Thank you so much.
Thank you for your time.
SUBSCRIBE TO IMAGINIZE WORLD ON YOUTUBE
We talk with forward thinkers, scifi visionaries and pioneering organizations about people and society, AI and humans, the earth and survival. Read more Imaginize.World
Subscribe on your favorite Podcast app
