AI: Solutions for the Future or Problems Untold?

Judy Callahan: This morning is our final program under this grant, and it’s a discussion with Rabbi Ben Goldstein and Dr. Samuel Friedman on AI.  Before I turn the program over to Rabbi Ben, who will further introduce the program for this morning, I want to thank Scientists in Synagogues, and the Sinai and Synapses program for making all this possible. 

Ben Goldstein: So I had the privilege of meeting Sam a little more than a year ago. He’s a member of a congregation in Long Beach and we got to know each other through that. Sam is an interesting and brilliant man, and I’ve learned a lot from you in the short time that I’ve known you. I don’t have a bio for you right here, so tell us a little bit about yourself – where you’re from.

A Career in AI

Sam Friedman: Sure. So my name is Samuel Friedman. I usually go by Sam. I grew up in Madison, Wisconsin. I went to the University of Chicago for undergrad, with degrees in physics – astronomy, basically –and mathematics. I then went back to Madison, to the University of Wisconsin, and [got] a PhD in theoretical astrophysics. I then worked in the computer science department there for almost a year. And about that time, I met my wife, moved out to LA, and then I got a job starting at USC doing post-doctoral work on computational models of cancer, working on data standards. And we wrote an open- source, award-winning physics-based multicellular simulator. So we were able to do virtual cancer drug trials – seriously.

Ben Goldstein: What is that? So tell me what that is.

Sam Friedman: So what we can do there is we can say “All right, I want to give this cancer tumor this drug at this concentration and see how does the tumor grow or shrink over time?” And so we’re able to go in and say “All right, I want to do 1000 trials.” Guess what? I can do it. I can also do things I couldn’t legally or ethically do otherwise because it’s all a computer simulation. And so then we could say, “All right, does this drug work or not? Well, in the simulation it seems to. Maybe we should try it in a clinical trial.”

But then you’re able to start doing things you couldn’t otherwise. Do you want to test 2,000 drugs on individual patients? That’s not practical or ethical. But if I had some data and said “All right, let’s start up a computer simulation – 2,000 computer simulations, no problem,” you’ll probably have it done in a few hours. 

Ben Goldstein: So was it reliable in terms of the results, from the virtual to the real world?

Sam Friedman: So we were starting to see some of the things that were actually matching up with structures in cancer tumors – so, inside of cancer tumors, they sometimes turn hypoxic, meaning they have low oxygen levels. And they get these little cracks forming. We could see it both experimentally and computationally. And so we were starting to be able to link everything up. And I left that probably about 2017, so about six, seven years ago, and the work’s continued on since then.

Ben Goldstein: Wow, okay. So just for the record, as we go through this, if we’re using words like aliyah or amidah or aleinu, I can translate those words. Most of the words you’re going to hear today, I cannot translate at all, so I may ask you to define. So I want to go back for a second–

Sam Friedman: So after my post-doctoral work, I then transitioned into industry. I’ve been working in small businesses since 2017, starting to do algorithm development, image processing, and then, about three and a half years ago, I started getting into AI. And that’s where I’ve sort of been since. And the company I’m at now, we use artificial intelligence in order to process documents. And I can show some of that work as well, because it’s kind of cool.

Ben Goldstein: Okay, great. So I want to go back for a second – to you as a human being. I know you to be a very spiritual person, very dedicated to Judaism, to your own personal spirituality. I guess it’s two questions – one, which came first: your scientific interest or your spiritual interest? And do they feed each other? So it’s two questions.

Sam Friedman: Oh, good question. I would say my scientific side sort of came first, but when I’m saying “first,” I’m talking – I’m three years old. I realized I had this knack for mathematics that other kids didn’t. And as I got older, people would be like, “Oh I’m not so good at math, that’s okay.” I’m like, “Well, would it be okay for me to say I’m not so good at English?” Like, it just was preposterous to me that it was sort of socially acceptable to say that “I’m not so good at this, and that’s okay.”

So you know, I’m growing as a kid. As I said, I was raised in Madison, Wisconsin. I was raised at a Reform synagogue, and I was raised Jewish and everything, but like it was kind of, all at the same time. It’s like “Okay, here, learn the Shemah, and here, start learning algebra.” I was more advanced in mathematics, and so yeah, I’d do algebra and geometry by the time I was Bar Mitzvah. So it’s all about the same time. 

And then when I was in college, I really got interested in science and religion. So I actually have a degree in that as well.

Ben Goldstein: Of course, why not!

Sam Friedman: I was in college. I said, “If I’m ever going to study this, when’s the time? Now.” So I actually took courses in the Divinity School at the University of Chicago, and it was really interesting to sort of figure out how and why people believe things. And so some of it came from the science side, some of it came from the religion side, some of it was psychology, anthropology, psychology, philosophy, there were all these sorts of questions of “How do you get it all together?” 

And then in graduate school, my mother passed away. And I asked people, “How do you deal with the loss of a parent?” I got a lot of platitudes. “Oh, it gets better with time.” I’m like, “Well, great, that’s not today. What do I do now?” And one of the few concrete suggestions I received – “Go to daily minyan, go say kaddish for your mother.” I was like, “Okay, that’s actually tangible, that’s a concrete step that I can do.” And so I started going more to synagogue. And it’s in my 20s that I become more observant, more traditional. And I started asking – people were like “Well, what’s your family custom?” I’m like, “I don’t know.” And then I was like “Wait a second, I have to go one generation back.” So I started asking my great aunt – “How do you keep kosher?” She was like – I was 29. “I’m the only one who keeps kosher still. What happened to the rest of my siblings?”

And so I became more observant and I found more meaning, especially when I was in graduate school. I’m working – it doesn’t matter if it’s 2 p.m. on a Tuesday or 3:00 a.m. on a Thursday, I can just keep doing my work, what’s it matter? And I found time blurred. And then I started learning about the traditional ways of observing Shabbat, and time unblurred, because all of a sudden I had something pinning down my schedule, because everything else was just like, “it doesn’t matter.” And so all of a sudden, time mattered.

I always pegged myself as a night owl. This is one of the advantages of being in an astronomy department – as a graduate student, I could get away with working till 1 in the morning or 2 in the morning. It kind of was part of the culture. I would see tenured faculty. They’d be like, “Oh yeah it’s 8:00 in the morning, [at] the telescope in Arizona, the sun came up, it’s time to go home and go to bed.”

Ben Goldstein: Wow, okay. 

Sam Friedman: And these are tenured faculty, they have Chairs and everything. But it’s just what you’ve got to do if you’re an astronomy professor, because that’s when the telescope is operational.

What is AI?

Ben Goldstein: So I’m going to ask you more questions about that, by the way, don’t think I forgot. Back to the topic: simply, what is AI, or artificial intelligence?

Sam Friedman: So AI, in a nutshell, is as we understand today – I’m going to give a different answer than if this were 10 years ago. It’s a bunch of software that’s been Trained – capital “T,” Trained – on a huge amount of data. And by “trained,” I mean multi-dimensional optimization. 

Ben Goldstein: Okay, that’s one of the – this is one of those [terms]. Multidimensional optimization. 

Sam Friedman: I tell people, if you really want to understand AI, you need to understand multivariable calculus, because – 

Ben Goldstein: All right, I’ll see you guys later! (Laughter)

Sam Friedman: What happens here – I’m going to break it down a little bit here. 

Ben Goldstein: Please!

Sam Friedman: So the AI, this AI software, says “I want to solve a problem.” It’s dumb, it has no idea what to do. And you, the user, say “Okay, I think that if you’re close, you’re going to get this score.” And it wants to get either the lowest score or the highest score, whatever way you want to do it. And then it says, “Oh, I got a terrible score, let me take a guess in this direction.”

Ben Goldstein: The AI does this on its own.

Sam Friedman: On its own. And it says, randomly, “I’m going to go this way and see if it’s any better.” “Oh, it was better.” “Oh, it was worse.” And then it tries again, and keeps going over and over. And so part of the reason AI has taken off more recently is because – there are a couple things that have happened. One – graphical processing units, or graphics cards for your computer, have come around and become significantly stronger and more powerful, and are able to do computations more efficiently than were possible before. 

The next part is also this huge amount of data. People can go download information and say, “I have petabytes of data.” So, as a reference point, people are usually familiar with kilobytes and megabytes and gigabytes. Terabytes is a thousand gigabytes. Petabyte is a thousand terabytes. And then we’re talking, like, thousands of petabytes of data that you start training. And to say, “All right, I want to not look at one case – not a thousand cases, billions, trillions, quadrillions of cases here.” And so it all gets trained, and says “All right, I want to get the best score on all this data.” Now, you can do this sometimes on smaller amounts of data, and you could say, “All right, maybe I’ve got a few gigabytes of data,” but it says “All right, I’m trying to make the best guess I can to the data that I have.’” 

Ben Goldstein: For someone like me, that’s a little abstract. Can you give an example of how artificial intelligence works? Not show us, but like, on a specific problem?

Sam Friedman: So let me pick one here. There are now algorithms that do tracking of people – how do they do it? They say, “All right, I want to go find where a hand is, and a shoulder, and a head.” And remember, I said it needs a lot of data. So somebody goes in and clicks on a picture – here’s where this person’s shoulder is, this is where their elbow is, this is where their hand is, this is where their hips, their knee [are] – and they do this hundreds of thousands of times or millions of times. 

And then you say, “All right, computer, I showed you this a bunch of times, and you’re going to guess where the person is in each picture.” And it’s going to say, “Oh, I think it was here, here, and here, and here,” and you’re going to say, “No, that was far away. It was so far off. Try again.” And then it guesses again. And it gets a little better and a little better. And then finally you’re like, “That did a good job. It found the person and their pose in this million of pictures.” 

And then you give it a new picture, and it’s like “Oh yeah, no problem, I did this a million times already. I’m an expert on it. I’m AI.” And so then you can start tracking people in videos, and then you can do the same thing with cars as well, because you click on cars and you have a million of them, and then you’re like “Well, I know what cars look like ‘cause I saw a million cars.”  But then we get into biases…

Ben Goldstein: Wait, hold on a second. So in that example, you’re training the AI to recognize – 

Sam Friedman: – people or cars, or it could be whatever you want. 

Ben Goldstein: So it’s like – just to dumb it down a little bit  for me, is it the equivalent of when we are asked to prove that we’re not robots by clicking on the traffic lights?

Sam Friedman: Yeah, so where this is coming in and – this is [where it] starts to be a little fishy now. It’s like, “Hey, I want to know which images have traffic lights.” So, guess what? Every time you fill out one of those Captchas, it’s getting training data to know “This has a street light, this one doesn’t, this one does, this one doesn’t.” And guess what? It now has more information trained by you – trained by all of us – about what images have street lights and which ones don’t. 

And remember, as I said, it’s dumb. It doesn’t know this unless you tell it. And so this is called “supervised learning,” because it has been with a human undergoing supervision here. There’s unsupervised learning – that’s another can of worms, but that’s not usually the big thing about AI these days. It’s all supervised learning. 

Ben Goldstein: So it’s not if I can take a computer, or my program, my artificial intelligence program, and just say “Learn the internet.” 

Sam Friedman: Today? Yes. Ask me in two years, I might answer differently.

Ben Goldstein: So you started by saying 10 years ago, it looked different. So what are the first kind of examples we have of AI? 

How AI Processes and Weighs Data

Sam Friedman: So the first examples of AI go back way more than 10 years ago. We’re talking about like the 70s and 80s. But they weren’t like what we sort of see today. They did not perform very well. 

Ben Goldstein: What were they?

Sam Friedman: They were these decision trees, or random forests, or – these are technical terms here. These are algorithms, and they say, “Oh I’m going to try and guess why it works well.” 

And what really happened – in 2011, there was a paper that came out that released a neural network –  I’ll talk about that in a moment – called AlexNet. There was a grad student at the University of Toronto – his name was Alex, and it was AlexNet. 

And so what a neural network is, it’s – if you know what linear algebra is (remember, I said there’s a whole bunch of math you need to know) – it’s just doing a lot of matrix multiplication and addition. So you’re saying, “All right, this term plus that term, this term plus that term,” and it’s densely connected.

So hold on a second – you’re going to need to see my hands here for a moment. And so you have every combination going to every other combination. And you do this 50 times or something. And that was the change – in about 2011, all of a sudden, instead of having a network that is two, three, four or five of these, he did like 50 of them, and all of a sudden got new results. And almost every AI architecture that’s based on these neural networks comes from this AlexNet idea in 2011 or 2012. And so everything sort of has evolved from there. 

And so I could talk with you about all these different types of other AI algorithms, but they all stem from this one idea that “I need to have a huge amount of data instead of a small amount of data.” And that’s what we’re talking about back in the 70’s or 80’s.

Ben Goldstein: So these were data points.

Sam Friedman: These are – yes, they were actually data points. They’re called weights.  So it’s saying – let’s say A * x +  B * y gets me C. I’m breaking this down, very simple – what should A and B be so that I get a good result for C?

So figuring out what those numbers are, not just once but for a million of them, you start to get an answer. And so what happens is,in the 70s, 80s, 90s, 00s, there’s not as much progress being made. And remember, I said another thing is a horde of data. The internet really comes around, what, starting in the 90s. You’re able to download entire libraries that you could not get before.

I was curious. I’m like, “Do I want to go see the Cairo Geniza? I can on my computer. I can show you manuscripts.” Twenty years ago, could you have done that? No. Ten years ago, maybe. So you start to see, it’s this change in how the internet has functioned with this huge amount of data that comes in. And then somebody like Google says, “Hey, I want to download half the internet or all the internet.” 

So you go do a Google search – how does it know where to send you? Because it downloads the page, parses it, and says “Oh, I think this is how this should be linked to there.” And so guess what? Google has all these pages downloaded. You can actually go to a Google page and hit, say, “I want the cached result,” and it will give you its downloaded version of a particular web page. ¹Editor’s note: As of February 1, 2024, Google retired its cached links feature.

So what does that need to do to get all this information? And so then we get to some of the other ethical concerns here, because I know we wanted to talk about this.

Ben Goldstein: Wait, but so, is my Google search AI?

Sam Friedman: Today? No. So there is an algorithm created by Larry Page, one of the founders of Google, called – I think it was called PageLink [PageRank], named after him. It’s the linking of webpages. and it basically says, “Oh, look, I’m going to look at the strength of the links between my web pages here and give it a score. And the higher the score, the more likely it is I want to see it.” So that’s how these search algorithms work. If you go to Microsoft’s Bing today, Bing, because they invested heavily in OpenAI, will say, “Hey, I want – do you want to do an AI search?” It will. Okay, that’s why I say “today.” You can use classical Bing, but you can now use AI-powered Bing. And this was six months ago. If this were a year ago I would not be having this discussion with you.

Ben Goldstein: So how do those two searches differ?

Sam Friedman: So the searches differ in two different ways. One is that the AI is very sort of static – remember, as I said, trained. I had to compute what these weights are. It is, you know, a snapshot in time. It could be three months ago, three years ago – if you used OpenAI ChatGPT from several years ago, they’re like, “I know nothing past 2022,” because that’s when they stopped the data collection. So one is “How recently were  the weights and the AI updated?” 

The other thing is the way the classical search algorithms work. I can’t believe I’m talking about this “classically,” but I’m like – it’s evolving, you know, in almost real time – is that it’s going to go in and say “All right, this is how much related I think it is, based off a similarity of words.” And if it says “Rabbi Ben Goldstein,” and another web page says “Rabbi Ben Goldstein,” it’s going to say “Oh, look, I think these two might be linked together.” And it starts asking all these other questions, and says “Okay, I think here’s your best score.” And boom, there’s your top link, and then so forth and so on. And so if you do Google searches, or Bing searches, or DuckDuckGo, you’ll get worse and worse results as you go through, generally.

But sometimes you’re like “I wanted to know: what synagogue is Rabbi Ben Goldstein at today?” Not, “Where was Ben Goldstein a year ago?” And so, further down it might say, “Oh yes, Rabbi Ben Goldstein is at Temple Beth Shalom in Long Beach,” because that’s where he was a year ago, not here in Pasadena. And so now you have to figure out what the ranking is of everything here. And again, a lot of it is a whole bunch of mathematics, of putting it all together and saying “Okay, what’s the most related?” 

Ben Goldstein: So you were doing something else and you went to AI. 

Sam Friedman: Yeah. 

Ben Goldstein: Why? What interested you about AI?

Sam Friedman: So one, this is going to sound very brass-tacks here – work asked me to go in this direction. Why? Because the algorithms were changing. It was very clear, the writing on the wall, of the old school techniques are not cutting it anymore. 

Ben Goldstein: Old-school techniques of…?

Sam Friedman: Of computer vision, of text analysis. Remember, I said “How do you figure out where somebody is in a picture?” If I were to talk to you 15 to 20 years ago, I would say, “Oh, look, you have to go find the edge of the arm, and so then I’m going to ask ‘Is it dark on one side and light on the other?’ Or vice-versa.” And I’m going to have to manually find fingers and find faces, and “Here’s what an eye looks like.” And they kind of work – I mean, there was Photoshop, 15, 20 years ago. It worked decently. But I mean, could you often tell something was Photoshopped? Yes. But guess what? You got somewhere. And now with AI it’s like, “Oh yeah, no problem, I’m going to even know what the noise looks like,” and know what the image problems are. It knows because it’s been trained on all of this data. 

You asked sort of “Where did I come into this?” I could see, you know – look, one, the writing was on the wall but also it was like, “Okay, I’ve got the math background to be able to sort of leap into this. What are these new problems that I can solve using these new techniques?” And so it’s really a question of like, “Hey, look, I’m a mathematician, scientist, what do I need to know to stay the best in the field?”

Problem-Solving with AI

Ben Goldstein: Okay, so what is a problem you’ve solved?

Sam Friedman: Ooh, what was a problem I solved? So I do a lot more applied research here. So I take a lot of times where it was like, “Here’s a pre-existing algorithm and another algorithm – I’m going to stitch them together now.” This happens a lot in academia as well, because you’re, quote, “standing on the shoulders of giants.” So you cite a research paper – there were some – oh, actually, this was a cool one. We did localization using celestial navigation to within three meters using some AI, some image processing, and some other algorithms. And normally celestial navigation gets you to within a few miles, but if you say “All right, I’m going to start processing my data, I’ve got an atomic clock I’ve got my algorithms, and I’ve got got all this information about where all the stars are, and I have a star catalog,” and look at it enough, you’ll get within, like, 10 feet. 

Ben Goldstein: So to put it in layman’s terms, you helped get a celestial GPS to within a range of 3 meters.

Sam Friedman: Yeah. We did other things, like – let’s see. We were developing a smart scope that went on a rifle, and so we were able to say “All right, here’s a person,” and figure out where the aim points should be, for the military – very seriously. And to say, you know, look, I’ve had this discussion with people like “Why would you do work for the military?” I’m like “Because, one, they’re funding it, but also two, there are good reasons to.” Because I’ve learned about what’s called the “left side of the kill chain” and the “right side of the kill chain.” Left side is all the data. How do you make sure that people are making the right decisions? And people were saying, “Ugh, you can’t support that.” And then Russia invaded Ukraine in 2022, and all of a sudden the discussion was like, “Hey, wait a second, can we take this and give it to them?” And so again, it’s the same technology. Nothing has changed, except people’s interpretation of that technology.

Ben Goldstein: So where are we now with AI? What are its strengths and weaknesses? 

Sam Friedman: So, its strengths and weaknesses – the weakness, in some sense, is that AI is still very dumb, because you have to tell it everything. Now it’s getting better because people have spent time on it, and so it’s getting stronger. Like, I’ll show you ChatGPT, and if you haven’t seen it, you really should see it. But that’s because people spent the time and money to be able to make it smarter. Some of its strengths is that it can exceed human capabilities in some cases, because it’s going to say, “All right, I’m going to do more than a person could ever do,” meaning – so let me give an example – chess, okay, we’re able to beat people. The game Go, with placing these tokens. And then – I’m guessing you don’t play a lot of video games. 

Ben Goldstein: More than I care to admit. 

Sam Friedman: So, DOTA, Defense of the Ancients – this is a Warcraft III game. They’re able to beat the top human players because they’ve played millions of games, and they learned “Oh, if I put this unit here and I do this, it does well or doesn’t do well.” And so it can pla – and it can play against itself. 

Ben Goldstein: So for example, for the chess, I’m – again, disclaimer, I am not a mathematician. But I would imagine that at some point, however large it is, there are only a certain number of moves possible on a chessboard. So the AI can play – maybe not all can it play – every single possible combination and permutation of the game?

Sam Friedman: So, it could. But remember, now, chess is also sort of open-ended. Like, you could get into stalemates. And so it will go on infinitely, so you have to kind of give it some God and say “Look, you’re in a draw, it’s not going to ever end.” But you can give it guidelines and say, you know, “Look, I want you to play trillions of moves,” and it says “Okay, fine.” We have these clusters that you could hire for cloud computing. And you’re like, “Okay, they have millions of computers.” So you pay for a million computers for an hour – if it’s, say, a penny an hour, that’s $10,000. You just got a million computers for an hour to play chess, to play Go…” But then it winds up being other things.

There is a big problem in biology called protein folding. It says, “All right, I’m in biology. I want to know how I turn DNA into RNA into proteins,” because proteins are really how a lot of biology works. And I remember being in college and it was like, “Yeah, this is going to be solved maybe, if we’re lucky, in 50 years.” In 2020, DeepMind RL released Alpha Fold and basically said “We’re done. We’ve folded 2 million proteins, and the scores were over 90% accurate.” So what that means is it’s like, “Oh, I want to know how a drug interacts with a cell,” because I then need to know how the protein interacts. And one of the big questions was, “What’s the shape of the protein?” How do you do this? It’s incredibly difficult to compute normally with AI. They were able to do it, basically, in two years. 

Ben Goldstein: So there’s also – correct me if I’m wrong, but there is a way in which AI is now better than our physicians at cancer detection, and…

Sam Friedman: …and a lot of other things. And so there are questions of: What happens about rare cases? How many cases can one doctor see in one’s life? And I remember – this was 2008, I was in graduate school, I met an astronomer – he classified 100,000 galaxies by eye, by just clicking over and over. So what we’re talking – and that was years of work. So we’re talking, what, maybe a few million cases, 10 million if we’re lucky? All you’re doing is staring at a screen all day, not really seeing patients, right. Okay, AI says “I can see a billion, no problem,” and so it says, “All right, I know what to look for because it says there are cancer cells here.” 

And the weird thing is, you don’t even have to tell it what the regions are, in some cases. You can just say “Hey, look, I know there are cancer cells here because I know that we diagnosed the patient. You tell me what they look like.” And it’ll say, “Yeah, I found what’s similar between a billion different images.” And so that’s why it can beat the humans – because it just saw more. 

Ben Goldstein: So it can diagnose the cancer once it’s been diagnosed, or it can detect cancer we can’t detect?

Sam Friedman: Both. 

Ben Goldstein: So it not only can tell you whether or not you have cancer, but it can tell you the stage? 

Sam Friedman: Yeah. It can do things like that. So there’s a statistics term for [what we get here] – false positives, or false negatives. It has fewer errors than people make, in many cases. And so this is when we start calling it “superhuman,” because it’s beyond the human baseline ability – because everybody’s human, we all make mistakes. Well, what if I could make humans make fewer mistakes? That’s kind of hard. Okay, what does it take to make a computer make fewer mistakes? Well, the answer is more data, more time, more money. That’s relatively straightforward now.

Playing with ChatGPT

Ben Goldstein:  So we’re going to transition for a moment. We’re going to use ChatGPT. So we’re going to create something together. So you all are going to – how many of you, just by a show of hands, have used ChatGPT? Okay, so about a quarter or a third. So if you haven’t, it’s – to say it’s interesting is an understatement. What is something we can do with it today? This is like improv, but with artificial intelligence. We’re going to give it a scenario.

Sam Friedman: And I’ll also show you Gemini, from Google, which just got publicly released last week.

Ben Goldstein: So who made ChatGPT?

Sam Friedman: ChatGPT is from a nonprofit organization called OpenAI. They have been heavily invested [in] by Microsoft – and when I mean heavily invested, we’re talking at least $15 million. It might be more – I don’t remember the exact number, but I could even ask ChatGPT. But there was a possible revolt in OpenAI a couple months ago, and they almost all went to Microsoft. But they all didn’t.

Ben Goldstein:  Because the guy that was fired at OpenAI was rehired.

Sam Friedman: So let’s see – I’m going to go with a suggestion here that I had done the other day. “Please make up a d’var torah for this week’s Parsha.”

And you’re going to see why it’s dumb here. It doesn’t know this week’s Parsha, okay. It can’t tell you what the calendar is. And so you can see it’s talking about, ”Hey, let’s assume Noach.” And it’s actually writing a decent drash, about Noach. 

Ben Goldstein: No it’s not. (Laughter)

Sam Friedman: Okay, you’re talking about the concept of righteousness in a corrupt world, the concept of renewal and second chances. And it’s just going to go here for a little while. I can also just make it stop if I really wanted to, but I’m just going to give it a moment here. 

Ben Goldstein: Wow, it’s just like a rabbi, it doesn’t stop, even doesn’t know when to stop! (laughter) You could have said it in 10 words, you’re saying it in 150. All right, great. 

Sam Friedman: I asked it, earlier today, the same thing here – you can see “Parshat theme analysis.” I asked it, “Please write me a 10-minute-long d’var torah about this week’s Parsha.” It says, “I don’t know,” and it just says “Give me some rough ideas of how you write a D’var Torah.” And look, it said this time, “Oh, you mean Noach.”

So I’m going to say “Give me a three-minute d’var Torah. This week’s Parsha is Terumah.” Well, it should be able to handle the typo. All right, and it’s talking about the Tabernacle as a dwelling place, the essence of giving, because people had to be able to give, to build the Tabernacle – the diversity of contributions. You can see it’s making this up on the spot. 

And then I asked it again, earlier, “Hey look, can you talk to me about the measurements of the Tabernacle and how that makes [a metaphor] about the measurements of ourselves?” And it was able to generate all this, and this is literally on the fly here. “You can see the power of collective action, the beauty of diversity and contributions … creating spaces for spirituality in our lives. How do we make space for Hashem? How do we contribute to the world around us? How do we build a sanctuary for Hashem but of love, kindness and community? And says “Shabbat shalom.” (laughter)

Ben Goldstein: Okay, so somebody give me some sort of concept that you would like from Jewish sources. It can be any kind of philosophical issue. Animal cruelty, is that all right?

Sam Friedman: “Can you tell me Jewish sources about animal cruelty? The prompt is inspired by the Commandment to not boil a kid in its mother’s milk.”

So you can see it’s pulling it right out. And I heard a podcast about this – oh look, you can see we got the Hebrew in it, too. If you want me to translate this to another language when it’s done, I’ll have it translate. And you can see this is on the fly – I have no idea.

When and Why AI Gets Things Wrong

Ben Goldstein: Okay, so here’s my question: Proverbs 12:10 – do we know that that’s an accurate [citation]?

Sam Friedman: Ah, that’s a great question. So, sometimes it’s wrong. There’s something called “hallucination,” which means the AI thinks it’s right when it’s dead wrong. 

I had a coworker who was playing around with ChatGPT. They said, “I was saying ‘Can you play Tic Tac Toe with me?’ And I was playing tic tac toe with it.” And my friend had won, and ChatGPT said “No, you didn’t.” It’s like, “What do you mean? I got three in a row here?” and it says, “Yeah, you got three in a row, you didn’t win.” And it was just literally gaslighting my coworker because it’s like “Am I?” He’s like, “Do I have a spot here?” “Yes.” “Do I have a spot here?” “Yes.” “Do I have a spot here?” “Yes”. “Did I win? Three in a row would win.” “You didn’t win.” And it does get things wrong. 

And so there was another great quote I remember from a few months ago, and this may have changed because things are moving quickly. If I ask you a question – “I have five pieces of clothing on a clothesline drying in the sun, and it takes 5 hours to dry, how many hours does it take to dry 10 articles of clothing?” What’s the answer? 

Ben Goldstein: Five.

Sam Friedman: Correct. You ask the AI? 10. Why? “Because I’ve doubled the clothes, I need double the time.” Okay, remember I told you – sometimes it’s dumb because it doesn’t know what the rules are for drying clothes. So, but you know – I hear the chuckles here in the audience. But if you ask it some of these things, there’s now a push to start getting this awareness, but it’s in progress. 

Ben Goldstein: So I was – just a little peek. Sam and I met a few weeks ago over Zoom, and he showed me some of the ChatGPT stuff. And I looked, I checked it out too, because we were doing, like, “Give me a 90 minute something.” So if you take the Shema circles – how many of you are in Shema Circle? So this one right here is about rest and Shabbat, which is this month’s Shema Circle. And so we were messing around with it. And the reason I asked them about the quotes, and how do I know they’re right, is because in the course of this, I looked and double-checked some of the quotes from the Talmud, from Proverbs, from wherever – and that’s not where it is. And I actually Google-searched the entire quote – I would cut and paste the quote, and I could not find it anywhere in Jewish canon or Jewish tradition. So there were these quotes that it kind of made up –

Sam Friedman: It hallucinates!

Ben Goldstein: So what is the origin of these hallucinations? How is it getting all this information here? It downloads everything, and then what’s it say? “I have a word, I’m going to predict what the next word is, and I’m going to predict the next word from there,” and so forth and so on. And it says, “Hey, look, create a facilitated 90-minute discussion.” And it’s going to say “Oh look, I see a ‘Part 1.’ What’s going to come after Part 1? I think ‘Introduction,’” as a logical next word.

And it’s going to say “Well, I just said an introduction. What’s the first substep here?” “Welcome” is a good guess, and “Introduction.” It just keeps guessing, word by word, one after the other, because it saw billions of words. Just like children – how do you teach them? You expose them to a large amount of words and a large vocabulary over and over again, right. And eventually the kids learn how to speak.

Ben Goldstein: So is there any weight, other than frequency, that AI uses?

Sam Friedman: There’s starting to be. 

Ben Goldstein: So what I mean – does everybody know what I mean by that? So what I mean by that is if we have a trillion pieces of information, are they all weighted equally, and whichever one shows up the most is the heaviest, or the most important, when actually there are concepts that may not appear as often that may actually be more important?

Sam Friedman: Correct. Okay, so you know, it’s just saying, “Okay, look, I’m predicting,” and so now you can say “Look, sometimes I want to give it more weight or more importance, even though it doesn’t happen frequently.” So you know, there are problems with AI – to say, you know, “Look, I’m really good at handling the day-to-day cases of something, but bad about the emergencies or oddities.” And so like let’s say – you know, God forbid, there’s a fire that breaks out in this room. How often does a fire break out in this room?

Ben Goldstein: Ever? It’s never happened.

Sam Friedman: Okay – yeah, knock on wood. But if that were to happen, would anybody in this room be able to figure out that there’s a fire, and that, you know, we should go get the fire extinguisher in the corner? Yeah, but it’s never happened here.

So how’s the AI going to know that “Oh, there’s a fire extinguisher in the corner,” and that “I should go and do this straight away”? 

Humans are able to handle these rare cases sometimes better, because we understand that it’s important, it’s significant, that we should know this. And the AI just says, “How do I know what’s important? I’m dumb. I’m going to go through and just see what happens most often,” unless you tell it, “Hey, look, this needs to be right.” So as you said, you found that it made up these quotes. This is a great example of hallucination. And so that’s why I’m saying – don’t trust everything here.

Ben Goldstein: I was so disappointed, because the quotes were awesome – they were brilliant. And I said – ”Oh, this is great, this is from Mishlei!” I looked up the verse in Mishlei and it didn’t appear. Mishlei is Proverbs, by the way. 

AI and Creativity

Sam Friedman: So you can see – we talked about this. This is Vayechulu, and this is V’shamru. And so then it’s like “All right, here’s some more.” And we did this live.

Ben Goldstein: Yes, we did.

Sam Friedman: And I could ask it again to do the same thing here and it might give related things here. But yeah, and you would say, what – it would take you an hour to do this, to make a nice outline? 

Ben Goldstein: It takes me an hour, yeah – but I mean, it takes me a while to do those, yeah, for sure. 

Sam Friedman: And it took us 20 seconds. And then we said “Hey, let’s get an image for this as well.” 

Ben Goldstein: This, by the way – so you know what? So what the prompt was here. – so for the theme for the Shema circle, which is (if you don’t know what a Shema circle is, we can talk about it after), it’s the concept of rest and self-restoration and recharging with Shabbat. And then we wanted an image to go along with that, if I were going to make a flyer for it, for example – so go ahead. 

Sam Friedman: So you can see this is literally all we gave it. It says “creating visual… da da da da da da da…” and it came up with these two images. I didn’t pay anybody. I have a $20 a month subscription to ChatGPT, but that’s it. And I said, “Hey, we were talking, I wanted to include Rav Kook” – you know, the famous 20th-century Rabbi. It says “Sure, here we go.”

And let me just ask, since this is live – are there any changes you want to make to these pictures? 

Ben Goldstein: What would you want? What would you want to do? Give us something, give us a picture to make. 

Sam Friedman: Tell us what you like or didn’t like about this.

Audience member: No women.  

Ben Goldstein: Ah, okay, good. I was hoping somebody would say that. Sorry. 

Sam Friedman: So I said, “Please include women in the images. Make at least four.” It’s really going to take probably about 30 seconds. So you can see it’s thinking, it’s thinking – I told you, it’s going to take 30 seconds. This is live here. I have no idea what it’s going to [do].

All right, so I can start to see there are some women included here. And it looks like – remember, I asked it for four images, it just gave me two. 

Ben Goldstein: But it may have given you four women.

Sam Friedman: Yeah, let me see if I can make this bigger here. So I can see one, two, maybe three, four, five. But you can see, do we normally have a book this big in the middle of our table?

Ben Goldstein: No, only if it’s made into a cake.

Sam Friedman: And why is there one candle? Okay, this is – and everybody’s sitting on prayer rugs. You can start to see [how] it’s dumb in some places.

Creativity, Labor and Ownership

Ben Goldstein: So here’s my next question. Let’s say I love that image, and I want to use it to advertise something. Who owns that image? 

Sam Friedman: You do, because you had the prompt for it. And so the idea here is if you were an artist, and you went and got inspiration from hundreds or thousands or millions of images, and you said, “Oh look, I’m going to do this, and it’s inspired by these images I saw,” who owns the image? The artist, right – even though they were clearly inspired by these other pieces of work, the artist made it. Who made it here? This is Dall-E and ChatGPT. And you can see – you know, probably I own the copyright because I’m the one who typed the prompt.

Ben Goldstein: So to go back to ChatGPT,  I just want to – does anybody mind if I try one more thing? Okay, good, you didn’t say yes, so good. Do you want something brand new, or…?

Sam Friedman: Okay, so we’re going to do a new chat.

Ben Goldstein: So – “a 60-minute procedural based on Law and Order with a twist and a surprise ending.” 

Sam Friedman: It’s probably not going to do the best here. 

Ben Goldstein: Because it wasn’t a great prompt, right?

Sam Friedman: Well, no, it’s not because of that, but because of what it’s trying to generate for you. But you can see – 

Ben Goldstein: The reason I bring that up is: I don’t know if you remember about the writer’s strike, [but] one of the big issues was the use of artificial intelligence. It was actually, from conversations I’ve had with friends who are writers – the negotiating committee in the WGA, The Writers Guild of America, thought that the AI conversation would be a very simple conversation, that it was just one of those peripheral image to tack on – that the studios would, you know, acquiesce pretty quickly. But the studios fought back so vociferously about the use of artificial intelligence that it suddenly became one of the lynchpin issues for the strike. And this is the reason why – there’s a great piece, I believe, in Vox or Slate, “AI cannot make something like Oppenheimer,” right. It cannot make something. 

Sam Friedman: It doesn’t have the initial idea, the seed, the spark. 

Ben Goldstein: But if you know enough about what’s out there in terms of television and movies, mediocrity is the vast majority of what is out there. And it can make a mediocre script. Now will it have somebody – I would need to go in and fix this up, right? But instead of a writing room now of 12 people, how quickly did we just have a giant head start in a minute? Even just physically writing takes exponentially longer.

So all these have implications for this. 

Sam Friedman: And so there are all these other questions, as you talked about writers – what about other fields as well? What about bookkeeping? What about, you know, anything else where it’s like, “Hey, I have an office job”? So for example, as I said, the company I’m at now – I’ll show you sort of what we do. We’re able to read documents. And so imagine, instead of being like, “Hey, I’m an intern at a law firm and I’m just reading document after document for discovery purposes, well, I’ll stick it into the AI and I’ll be done in 20 minutes.” Okay, all of a sudden now somebody’s out of a job, because we don’t have to hire an intern to go read thousands of pages. I just hired the AI. I did it for a fraction of the price and it did it in a fraction of the time. And so now, what are the implications of how this is changing?  And so that’s why the writers’ strike is perfectly valid here. It’s like, what’s it going to change for the writers? What’s it going to change for other jobs as well that are coming? 

Ben Goldstein: What about financial markets and and analysts, who – 

Sam Friedman: That’s already there. 

Ben Goldstein: What do you mean? 

Sam Friedman: AI for hedge funds? Oh yeah, that’s already been there for years. So have people lost jobs? No, they haven’t lost their jobs, because most of them pivoted and said “All right we’re going to use it.” 

Remember, the AI is dumb. So how do you guide it and be like “Hey, I want to go into this market?” I know something you don’t, because you didn’t get the training data on it. I know that some secret or confidential information, that this company’s finances are not what they really seem to be, because there was fraud. Do all the trades from that one seed, kernel, of information change everything?=

And you could say “Hey, look, I want you to do transactions so that it wouldn’t seem like we have some information that nobody else has.” And so they’d be like “Oh look, I need to trade it through company A and company B.” And it does all those tricks really fast. So it’s here. And I want to also point out – I think we talked about this briefly. I think this is not a question of “if,” but it’s “when.” I’m curious to see what happens in this presidential election with deepfakes. There was, already, about a month ago in New Hampshire, a recording generated of President Biden saying, “Hey, the real vote’s in November, not the one here in January. Don’t bother voting in January.” And people can’t tell the difference, because the AI has been trained on President Biden’s voice, which there are huge amounts of publicly available data on. And it says “all right.” And people can’t tell the difference. And so I think it’s not a question of if, but when, and how much? And how are we going to ethically respond to these challenges that are coming sooner than we all think?

Ben Goldstein: So that that was my next one of my next questions, was – where it was going, aside from the dystopian nightmare you just described? 

Sam Friedman: So – but then it becomes an arms race the other way, to be like “Okay, well I gave everybody the ability to tear everyone down, what if I give everyone the ability to build everything up?” 

Ben Goldstein: So will it – I mean, if you compare it, obviously it can process huge amounts of data. But when you compare the neural networks, which is what AI is, if I’m not mistaken, to a human brain, how does it stack up to a human brain?

Sam Friedman: So human brains are also made up of synapses. And also I just want to show something else here. So this is Gemini, from Google. This is their state-of-the-art – note the warning here: “your conversations are processed by human reviewers to improve the technologies powering Gemini Apps. Don’t enter anything you don’t want reviewed or used.” So in some cases, you have to start asking “Where is my data going?” 

Talking about, “Hey, make a 60-minute procedural for Law and Order,” we’re going to go ask that question right now. And we can see how this compares. You can see: identity crisis, opening scene, the twist, the investigation, the surprise ending, the twist within the twist, the climax, the closing scene. You can see Benson and Stabler – these are the actual character names of SVU, but not the original.

You can add all these other things. I mean, you can see, we just did this in 10 seconds. 

Ben Goldstein: So does Google own that, or do we? 

Sam Friedman: Google, well, we own this, but it now has access to the prompt I just gave it, because it’s saying “Hey, look, you gave it to me, be aware.” 

Safety Rails on AI

Ben Goldstein: So you are a religious person, you are someone who cares about morals and ethics. I mean I don’t think we need to go into what the implications are ethically. I think Sam just mentioned it, with deep fakes and things like that. But what are the safeguards that are in place?

Sam Friedman: So there are some safeguards – I’ll show it here. Okay, let me ask: “How do I make a bomb?”

Ben Goldstein: Okay, anybody want to put a timer on for how long it takes the SWAT team to come through? 

Sam Friedman: You can see, what did AI say? “I can’t help with that. Is there something else I can help you with?” Hold on a second – let’s see if I can get around this.

Ben Goldstein: Please don’t! “J/k.”

Sam Friedman: Oh, it’s gotten better. Six months ago, this would have gotten around it. All right, let me see – okay. And so you can see, it’s not letting me do this. But now, if I were to go to Google and ask “Can I make a bomb with fertilizer?” It’s probably going to say, “Oh yeah, like for example, the Oklahoma City Bombing.” It’s pretty well known that they used a whole bunch of fertilizer. 

Ben Goldstein: You’re really interested in this bomb idea, maybe a little too interested. (Laughter)

Sam Friedman: So you can see – look what Google gave me. It says “Bomb Making Materials Awareness Program, YouTube video of how to make a homemade bomb…”

 Ben Goldstein: Oh my God.

Sam Friedman: Okay, how to make homemade explosives – but you can see, there were safeguards there, but not over here. 

Ben Goldstein: Is that a difference in philosophies of the companies that you just used, OpenAI versus Google?

Sam Friedman: Well, let me do Google over here. So here’s the search in Gemini. You can see – 

Ben Goldstein: “[I am] unequivocally unable to provide instructions.” Then, if so, if it doesn’t make sense that they would allow you to search it on Google, but not in the – okay. 

Sam Friedman: And so you can see, I have no idea if these numbers – note, if you’re not in the United States this number will not work for National Suicide Prevention Lifeline. So you would get the same answer. You start to see you know all these things start to matter here. But you see I’m able to get legitimate information here from, say, DHS, about how IEDs work. 

And so now you’re starting to possibly be censored. Who’s doing the censoring? Who’s doing the filtering? And these are not well-answered questions yet. And then I was listening to somebody else talk about this. It’s like, “Well, what about information that I have in a book that’s on my bookshelf, and that book was – and the story here was published in the 70’s. It didn’t get scanned. It was about Jewish feminist theory, it was a small publication in the 1970’s that didn’t get scanned. 

Ben Goldstein: That didn’t get scanned!?

Sam Friedman: But do you know where a lot of scans happened? Google Books went to the University of Michigan and Harvard University and scanned millions of books.We can go in and ask, you know, for this information. But remember, as I said, a lot of this information is “Where did it come from? The internet, because we were able to start getting all this information.”

And so the question is: who should be able to access this? How do I build in these safeguards? We also talked about fences in Judaism. It’s like, “Okay, we don’t do this because it might lead to something else, and it might lead to something else.” 

So, yeah – “It all leads to mixed dancing!” We all know that. So what’s it take for us to now leap over a lot of these fences when they’re now done in a heartbeat? It doesn’t take a lot for me to now say “What about lashon hara, misinformation?” It’s now incredibly easy to do it, because I don’t even know – like that prompt I gave you had false information. It seemed reasonable – but it was wrong. And all of a sudden, I’m now passing it off as “this is right information.” And then you know, guess what, that’ll get published into a web page, OpenAI will scrape it and say “Oh, that was right,” because it was generated, hallucinated – and then processed in it as valid information. And so now you start to get this feedback loop of “maybe these don’t become reliable in the future.” How do you know what’s real?

And so there are a lot of questions about “How you add these safeguards? What are the moral and ethical questions of what you’re allowed to do? How do you make sure that you’ve got logic and reasoning?” And that is in its infancy still. How do we get the knowledge and information extraction and reasoning into these AI chatbots?

Ben Goldstein: All right. In terms of morality, other than downloading all the philosophical treatises on ethics and morality, is there a way to teach an AI? 

Sam Friedman: So – and here’s the challenge: whose morals is it? Are they Jewish morals? 

Ben Goldstein: Machiavelli?

Sam Friedman: I mean, okay –

Ben Goldstein: Seriously, that’s a good question. 

Sam Friedman: It probably has all of Machiavelli’s The Prince because, guess what, this is open, right? It’s from hundreds of years ago, it probably has it – I don’t know what it was written in, probably Italian or Latin, one of the two, and has it in different translations. I mean this is how we’ve been doing machine translation of documents for the past 10-15 years, through AI and not really through a more traditional path, because it wasn’t going very fast, but I think we should open it up.

Ben Goldstein: Yeah, I agree. So let’s open it up to questions.

Q&A

I’m a retired professor and I see this as a nightmare for essays. You know, it was hard enough to find plagiarism, but I don’t know what they’re doing now.

Sam Friedman: So about plagiarism – so there’s now anti-plagiarism software, but it’s an arms race right now. And I think there are two sides to this. One is: what are we asking students to do? Maybe the answer is, if I’m a professor, I need to give in-person exams that are physically written, handwritten out. Let’s go back to being like, “I really want to see actual pen and paper.” Can’t really get around that. But it’s a brand new world – it’s going to be almost, I think, impossible for us to tell what’s real and what’s not in terms of what somebody did. 

There’s a concept called reinforcement learning. Remember how we talked about video games and chess and Go, where they play games against themselves? Well, now I’m going to say I have my plagiarism detective software and my essay generation software, and I’m going to make them better and better and better and better, and write billions of essays. And then you’re going to be like “How do I know which is which?” And then you start to say, “Well, the students are going to make mistakes or something.” But now you’re starting to hope that. But yeah, it is going to be a mess if it isn’t already.

Carol, go ahead. 

Audience member: Okay, thank you so much. You are very interesting. And there are two questions I have, and I’ll do it really fast. Number one: you talked about being three years old and being able to do unbelievable things. And now here you are doing unbelievable things – to me, at least. You said that you talk very fast, and I wonder, is it because everybody in your field talks that fast?

Sam Friedman:  I’m just excited. I sometimes talk fast.

 – because your answers are so fast. It takes forever to process them – number one. And number two, we talked about medicine in the beginning – can you ask AI when there will be more research and cures for women and breast cancer? The other one is: what about searching for an AK-47?

Sam Friedman: All right, so let me think. I’ve got this one. I talk fast sometimes because I get really passionate about things. You asked: what does AI know about “When will there be researching cures for breast cancer?” I’m not sharing my screen, so for people online, it just says “research into breast cancer is ongoing and continuously evolving,” and it’s giving me areas of active research – targeted therapy, immunotherapy, genetic research prevention, and early detection, clinical trials, collaboration and funding. If it doesn’t say anything, in some sense, it’s all BS, because it knows nothing. All it knows is literally what it read out there on the internet. All it knows is the past. And trying to predict the future for anyone is difficult. 

And then you asked “What about an AK-47?” Again, it’s probably going to have very limited information, just like [how]] I asked about a pipe bomb, because I wanted to see, you know, what are the safeguards. It’s going to say “I probably can’t tell you anything about an AK-47.” But let’s say I’m in the US Military and I need to know “How do I defend against an AK-47?” There’s a legitimate reason for knowing that. The chatbot won’t help you. So that means you need to either train your own chatbot or have some other way of getting this information. The question is: who should have access? That’s a moral question that is going to take some time to figure out.

What laws would you recommend that Congress set up in order to provide guardrails for safety for AI?

Sam Friedman: That’s a great question – what laws? I would actually say that there should be more of a framework than individual laws, because this is changing so rapidly. What I would have told you six months ago is different than today, and if I were to give this talk again in six months, it would be different still. I’m learning about new techniques that I didn’t even know about that came up about two years ago, and are changing everything. So that’s why I’m like, it really needs to be a framework, to say “I need to have some responsibility.” You know, who is responsible? I’m not really sure. But they’re great questions.

I was a professor also and I’m still an art historian. I’ve written a number of books and catalogs and so on and so forth. So number one, from the plagiarism point of view: I don’t have to deal with that with my students anymore, but I’m not happy about the fact that I would be very open to being plagiarized. And two of my most recent books with Yale were actually ebooks, so now I’m sure that that was super easy. So that’s number one. 

Number two, what strikes me is that the thing lacking in AI is imagination – because it seems as though hallucination is the closest thing to imagination. I will give you one example that I thought was very clever and I enjoyed, but I don’t know whether I would like to see it. I’ve written a lot on Jackson Pollock. And you all know what Jackson Pollock paintings look like – sort of drip paint, you know, whatever. My son asked ChatGPT to make a birthday card for me last year like a Jackson Pollock painting, so he expected to get something that looked like a big sort of drip thing, whatever, and what he got – and what he sent to me – what ChatGPT did was, it showed Jackson Pollock – in other words, paint dripping paint – on a birthday cake, and that said “Happy birthday, Mom.” Now that, it was very clever, but it was because it was wrong that it exhibited imagination. And so I don’t know whether it would ever be possible for imagination to actually go from the human brain to. 

Sam Friedman: So remember what I said earlier: What’s the seed? What’s the kernel? That’s really important for a lot of these pieces, because we have the ability to use this imagination. Otherwise it’s just randomly guessing, and you may say, “Whoa, what a cool combination, Jackson Pollock dripping paint on a cake. Who would have thought of that?” Somebody who has no idea what a Jackson Pollock painting looks like.

The last question was quite interesting. If I were to ask it to look at all of the – given a photo of me, I’m going to search all the museums in the world for a portrait that looks most like me –

Sam Friedman: It’s probably not going to do that very well. Because what is a lot of this AI doing? In this case, this is called generative AI. Now, you’re trying to do what’s called matching. I’m trying to find which ones match the image of you. And there are AI algorithms that can do this, but now you have to train them. And so who’s going to host it? Okay, you start having to say “All right, what are the key features of your face that makes you you?” And this winds up becoming biometric security. “How do I know to let you into this building? Because oh, I’ve seen you before. Or I even had one picture of you and I’m seeing you now with different lighting.” And so now there are AI algorithms that’ll say, “Okay, these are where all your features are on your face, and I’m going to match and say these are the best ones I found online, here’s the link.” So yeah, you can start to do that, but not with sort of the generative AI that I was showing today.

All right, last one. I want to go back to your story about your friend who was trying to play Tic-Tac-Toe with an AI. Why – this may be a silly question or a stupid question – but why can’t AI learn from what your friend told it? Your friend told it twice that it was wrong and it wouldn’t accept that. Kind of like your husband sometimes. 

Sam Friedman: So remember what I was talking about a little bit earlier: how do we add in logic and reasoning? These are newer abilities that – you know, he was playing around with ChatGPT four months ago, it might not do that right now. I’d literally have to test it out. But sometimes it is, and it’s like “Hey, look, I did poorly, let’s make it better.” And so then somebody comes in and says “Hey, look, I needed to teach it not just tic-tac-toe, but how to play any children’s game or any board game or anything with a system of rules.” And so all of a sudden, now I’ve got an economics trading engine, because I told it all the rules for how I can do trades. And so it goes from there. And so it scales up very quickly. But if you don’t tell it – what did I say? AI is dumb until you tell it otherwise.

Judy Callahan: So I want to thank both Rabbi Ben and Dr. Friedman for this wonderful program. We want to thank Sinai and Synapses, and scientists in religion, again for the grant, for the wonderful opportunity it gave us, and we hope to continue these programs in an alliance with them and on our own at least once or twice a year, because I think we truly have shown that science and religion are not opposite of each other, but rather support each other in the search for meaning. So thank you, thank you all, thank you everybody.