Transcript [00:12] Hi friends and welcome to 1840 where [00:14] each month we explore a different topic [00:16] trying to balance modern sensibilities [00:18] with traditional sensitivities to give [00:20] you new approaches to Jewish ideas. I'm [00:23] your host David Bashkin and today I am [00:26] so excited because we are exploring the [00:29] intersection between AI, artificial [00:31] intelligence and halaka Jewish law. This [00:36] podcast is part of a larger exploration [00:38] of those big juicy Jewish ideas. So be [00:40] sure to check out our website 1840.org. [00:42] That's 18 fty.org where you can also [00:45] find videos, articles, recommended [00:47] readings, and weekly emails. So, I'm not [00:49] sure when it was, but very early on when [00:52] AI was just emerging. And really, even [00:54] even in our original intro video, I [00:57] realized that a lot of what was going to [01:00] steer and drive the work of 1840 was [01:03] understanding how modernity and how the [01:06] world that we live in today really [01:08] shapes and affects our religious lives. [01:10] They're not two separate worlds, but the [01:13] culture, the ideas, the technological [01:15] innovation shapes our very sense of [01:17] self, which in turn shapes in many ways [01:20] our approach to religious life. And the [01:22] questions that emerge regarding AI have [01:25] always fascinated me. Not because I am [01:28] all that excited about technological [01:29] innovation. I actually am not and I'm [01:32] don't really even know that much. I have [01:34] long been interested in ideas related to [01:36] consciousness and self to really [01:39] understand what makes us most human. [01:42] What is our self? What are we [01:44] experiencing? And those ideas especially [01:47] as AI got stronger and stronger really [01:50] highlighted through the inverse as [01:53] computers do more we start to ask like [01:55] so what is our humanity about? And last [01:58] year around this time on 1840, we did an [02:01] entire series with our partners at [02:04] American Security Foundation, ASF, an [02:07] incredible, incredible foundation [02:09] dedicated uh to the intersection between [02:11] religion and AI. And we did a series and [02:14] you can check out uh some of those [02:15] episodes whether it's on our YouTube [02:17] page or previous episodes. But today's [02:19] episode is actually something a little [02:21] bit different. And I have to give a big [02:23] thank you to my brother-in-law's [02:25] brother, the best Jewish introduction to [02:27] any podcast, uh, for introducing me to [02:30] two incredible computer scientists. [02:32] One's a mechanical engineer who were [02:35] doing incredible work in how to evaluate [02:38] the competence of AI in approaching [02:41] halic questions. And my brother-in-law's [02:44] brother, Brett Spear, thank you Brett, [02:46] showed me that they tweeted on social [02:48] media under the handle halaka bench [02:51] because benchmarks as where our [02:53] conversation discusses is a major major [02:55] theme in evaluating AI. And they [02:58] developed halaka bench. How does AI [03:01] perform when it comes to rabbitical [03:04] halic knowledge? Knowledge that emerges [03:06] not just from the Talmud, from all of [03:08] our codes of law, the etc etc. and I [03:12] shared their research on my own uh [03:15] Twitter feed, my X feed, whatever you [03:17] call it. And it happens to be have a [03:18] little bit of a bigger following than [03:19] they did at the time. And I saw the [03:21] conversation emerging. All of the major [03:23] players were really really interested in [03:27] what their work showed and didn't show [03:29] about the development of AI. And I [03:31] invited them to come on to 1840 and talk [03:34] a little bit not only about their own [03:36] work, give a little bit of an [03:37] introduction really to the just like a [03:39] primer on what AI is all about, what [03:41] their work says about the intersection [03:43] between AI and halaka, why the AI world, [03:46] and this is absolutely true is so [03:48] fascinated by Talmud and Jewish codes of [03:52] law. And if you haven't seen it already, [03:54] aside from reaching out to religious [03:56] leaders, so many people in AI are [03:59] fascinated by the Talmud. Uh, and we [04:01] speculate a little bit as to why. And [04:04] then finally, we talk a little bit about [04:05] what they built and really what the [04:08] future holds in terms of the [04:09] intersection between Halaka and [04:12] artificial intelligence. I am so excited [04:14] for this conversation and I know you [04:16] will enjoy it. [04:18] I am so excited and really it is such a [04:21] privilege and pleasure to have Ben [04:23] Nywood and Sam Weiss and you came onto [04:27] the scene and I have to mention him [04:28] right up top in the beginning. My [04:30] brother-in-law's brother, could you [04:32] think of a more Jewish introduction to a [04:34] conversation? My brother-in-law's [04:36] brother, Brett Spear, who I know and I [04:38] love, is not usually the person who is [04:42] sharing with me breakneck innovation in [04:44] the Telmutic world. We got to give him [04:47] credit. He is the one who connected me [04:48] to you and you broke out onto the scene [04:51] uh because I shared on Twitter. I It [04:54] wasn't so popular when I shared it. I I [04:56] don't think it had [04:57] >> No, I mean we had just posted it. It was [04:58] nowhere. We were just We just put it out [05:00] into the world and thought, "Is anyone [05:02] going to see this?" And then Brett said, [05:03] "Oh, I can send this to my brother's [05:04] brother-in-law brother." [05:06] >> And we were like, "Yeah, sure. Go for [05:07] it." And you know, then you posted and [05:09] then it got really picked up for sure. [05:10] >> And it was called Hala Bench. And what [05:13] you did is you had several different AIs [05:17] all taking rabbitic ordination exams [05:21] what are known as [05:23] bikinas rabbitic ordination and bikina [05:25] is a test and you basically took all the [05:28] different AI models and had them take [05:31] different rabbitic ordination exams. And [05:33] before we kind of get into like what [05:35] that tells us about AI, what that tells [05:38] us about rabbitic ordination and hahik [05:41] progress and we could have that [05:42] conversation, I want to start way back [05:44] with the basics and let's talk about [05:47] even the title hala benchmark. [05:50] >> Explain to me what is a benchmark as it [05:53] relates to AI? [05:55] >> In the simplest sense, uh a benchmark is [05:57] checking how good AI models are at [06:00] certain things. They came out because [06:01] originally the first models that came [06:03] out were only so capable. They could do [06:05] some things but not others. So people [06:07] wanted to see okay how good are the [06:08] models at at taking the SATs or taking [06:11] uh you know the LSAT or you know [06:14] deciding the outcome of a legal case. [06:16] And so people started to build these [06:17] benchmarks. I mean I'm a hardware [06:18] engineer. I'm a mechanical engineer. I [06:20] build robots. And there's just now the [06:21] models are getting really good at [06:23] hardware type design things. And so [06:25] there's all these new benchmarks coming [06:26] out now. hardware bench. That's how good [06:28] is the AI at designing a hardware [06:31] system. So, it's important because it [06:32] helps figure out um as the models [06:35] progress, can you treat them like a [06:36] software co-orker, like a legal [06:38] co-orker, like a hardware co-orker, or [06:40] how much can you trust them to do well [06:42] at different tasks? and where we are now [06:44] in the development of AI for most [06:49] benchmarks like let's say an LSAT or an [06:51] SAT are most of them able to score [06:54] exceptionally well on an SAT a GRRE yeah [06:58] I think for something like the SAT where [07:00] it's a very structured test especially [07:01] it's you know it's multiple choice and [07:03] the way that the test is put together is [07:05] is very orderly I don't know the [07:07] benchmark numbers offhand I would assume [07:08] they do very well on those on other [07:11] types of benchmarks where it's more [07:13] open-ended. Like there's other tests [07:15] that are more essay based. Even though I [07:18] think they might do less well, but I [07:20] think even on like the bar exam they've [07:21] been taking, they they're certainly [07:22] passing the test and I think probably [07:24] doing very well. [07:25] >> Your background is a little bit more [07:27] closely connected to the world of AI. [07:29] Where is your work, Ben? You work in [07:32] robotics and mechanical engineering. And [07:34] your background is where? [07:36] >> I'm a software engineer. I've done a lot [07:37] of machine learning, uh, which you know [07:39] is another word for for AI. Um, and I've [07:42] worked at a number of different tech [07:43] companies doing that. Um, so [07:45] >> let's talk a little bit about why what [07:48] we call, and these are all really what [07:50] we call LLMs, large language models. Why [07:54] would a large language model be good at [07:56] some tests and not at others? Like I [07:59] remember when I was in school, the kids [08:01] who do really well on standardized [08:03] tests, like they do it on all the [08:05] standardized tests. Like they're they're [08:07] like they're so good at it. And the kids [08:09] who just like, you know, 1550 on the [08:11] SATs or however it's scored nowadays or [08:13] the ACTs, those are the kids, the MCATs, [08:16] the LSATs, like and there were other [08:17] kids who like you needed the tutors, you [08:19] needed the books, but there were certain [08:20] kid they just picked it up. [08:22] >> When we think of an LLM, can you explain [08:25] why is the structure of how an LLM [08:28] learns would make it good at some [08:31] intellectual tasks but not as good at [08:34] others? [08:35] >> Yeah. So if you think about the way an [08:37] LM is trained, right? An LM is a is a [08:40] type of neural net. So you essentially [08:41] have this big web of of weights, right? [08:43] And the way they train it is those [08:44] weights start out the connections [08:46] between them start out effectively [08:48] random. It's a bunch of numbers, right? [08:50] And they feed data into it. So they feed [08:52] in books and the entire internet. And as [08:54] they do that, it sets these weights to [08:55] to sort of recognize those patterns in [08:58] the training data that are [08:58] >> is the machine doing setting the weights [09:00] on its own or that's the programmer [09:02] doing. the algorithm that trains it is [09:04] setting the weights. So essentially what [09:06] they do there's something called like [09:07] next token prediction, right? Where like [09:09] you you essentially will will give it [09:11] that the task that it's trying to learn [09:13] is given a bunch of input can it predict [09:15] what should be the next output. So you [09:17] would you know give it a bunch of text [09:19] and you maybe you know what the next [09:20] output should be and the machine doesn't [09:22] you have it try to produce what should [09:24] be the next output and if it does well [09:26] it gets rewarded. Old McDonald had a and [09:30] like can you fill that in that you know [09:32] it's like a farm and now it sees that it [09:34] knows the language of that. Okay. [09:35] >> Right. So that that's roughly like [09:36] what's going on there. Um so they're [09:38] they're learning these statistical [09:40] patterns. So then your question about [09:42] why are they good at some things and not [09:44] others. So some of this honestly is like [09:46] not entirely well understood like how [09:48] these weights you know giving it huge [09:50] amounts of training data actually [09:51] produces a lot of these capabilities is [09:53] not uh entirely clear. A good example is [09:55] is math right? You might remember a few [09:58] years ago, you could ask an LLM, you [10:00] know, even simple addition problems and [10:02] they did not do well on them. And the [10:04] reason is because they were trained on [10:05] text. They're recognizing statistical [10:07] patterns in text. And maybe the phrase 1 [10:09] plus 1 equals 2 comes up a lot in text. [10:11] So maybe they can do that. But then [10:13] more, you know, uh arbitrary [10:15] combinations of things don't and so it [10:17] can't even add a few numbers together. [10:19] But the labs that are working on these [10:20] models, they've recognized those [10:23] deficiencies and then they steer the [10:25] training data and the training process [10:26] to like account for that. And that's [10:28] actually coming back to the benchmarks [10:30] where the benchmarks come in where [10:31] people might publish let's say a math [10:33] benchmark. And then the makers of these [10:36] models, the the labs can assess how well [10:39] their model is doing against those [10:41] different areas and they can get better [10:43] and better and better that way. So you [10:45] might you a benchmark might be like the [10:47] ninth grade. We took the regents when I [10:50] was growing up. So you might take the [10:52] ninth grade regions. What's very [10:53] interesting to me about LLMs, machine [10:56] learning and AI is it kind of sits on [10:58] the nexus between math and language. And [11:02] this goes back to the original [11:03] computing. Now I'm the kind of person [11:05] who's like [11:07] I'm dangerous because I have a lot of [11:09] general information but I'm not an [11:10] expert in anything. But one of the [11:13] thinkers who I am so fascinated by. I'm [11:16] fascinated by his theories and ideas is [11:19] Claude Shannon. Claude Shannon is really [11:22] like the grandfather of all information [11:24] theory. And his basic idea, which I [11:27] think must be connected in some ways to [11:29] these LLMs. I don't know if you're [11:30] familiar with it, but his basic idea, he [11:32] was involved in code breaking. And the [11:35] basic idea that he came up with was that [11:37] all information can actually be broken [11:40] down. This is the bit system. How [11:43] statistically possible it's going to be. [11:45] And when you're transmitting no [11:47] information, like if I roll a dieice and [11:50] I say this die always all the time falls [11:53] out on four and then I roll the dieice. [11:55] So there's no new information being [11:58] disclosed. If I roll a die that could [12:00] fall in one of two ways. So that's like [12:03] I think like a bit of information or two [12:05] bits of information. and we're able to [12:07] process more and more and more based on [12:09] just how how statistically complex is [12:12] this with LLM is the way that they're [12:15] processing information kind of like [12:17] they're learning the statistical [12:19] likelihood of the next word or the next [12:22] piece of information [12:23] >> on a very basic level. That's right. [12:25] Right. That that's like the the [12:26] elemental level of what they're doing [12:27] and and it's very connected to like [12:29] Claude Shannon's ideas, right? So he uh [12:32] had a lot to say about uh compression, [12:34] right? like what's the most efficient [12:35] way you can compress information without [12:37] losing anything or with losing things in [12:39] certain ways. And you can think of an [12:41] LLM as like effectively a method of [12:43] compressing information, right? You're [12:45] taking a huge amount of information and [12:47] you're compressing it into some smaller [12:50] space, but you're still retaining a lot [12:52] of that information in some way. And [12:53] you're also there's, you know, obviously [12:55] it's learning connections between those [12:56] things, but it's kind of a fuzzy picture [12:58] of all of that information that you've [13:00] put in. [13:00] >> It's so fascinating. Now let me ask you [13:02] before we get into the actual halaka [13:05] benchmark uh and your work in AI let me [13:08] ask you one more question what is your [13:11] relationships in your own lives with you [13:14] know whether it's chat GPT or claude [13:17] what do you use them for is there [13:19] anything that you deliberately don't use [13:20] them for because you kind of have more [13:22] information than most [13:24] >> yeah I mean I use it for almost [13:26] everything these days there's very [13:28] little that I won't go to to Claude or [13:29] ChateBT for first in in work. I use it [13:32] all the time. It's a a thought partner. [13:34] It helps speed up work. I use it, like I [13:36] said earlier, they're getting better and [13:37] better at even hardware work and [13:39] mechanical design. And so, I'll kick off [13:40] tasks and use it as almost co-workers or [13:43] or people who work for me and, you know, [13:44] kick off agents who are running [13:46] different projects that I would have to [13:48] do myself. And now I can rely on the [13:49] early steps of those projects to be done [13:51] by an AI. Um, on a personal level, I use [13:54] it to, you know, help manage my inbox, [13:56] my calendar, order groceries for me, [13:58] meal plan. Do you let AI send emails [14:01] without checking them first? [14:02] >> I usually tell it to give me a draft, [14:03] but if it's not something I care about, [14:05] then you know, maybe. Or if it's not to [14:06] a person I care about, then maybe. [14:08] >> Um, I definitely have it like I have a [14:10] meal plan. [14:10] >> Best friends. [14:11] >> Exactly. Never really knows anymore if [14:13] it's me or um Yeah, certainly for work. [14:16] I I read the emails, but I have it write [14:18] emails for me for work all the time. You [14:19] know, oh, can you check my inbox and [14:20] follow up? Tell me what you're saying, [14:22] what in new information there is. So, I [14:23] don't even have to read my email anymore [14:24] because Claude just, you know, digests [14:26] everything and tells me, pulls out the [14:28] important bits and then archives all the [14:29] other stuff. Um, I use it to meal plan. [14:31] Now, Sam uses my meal planning engine on [14:34] the meal plan every week. But, uh, but [14:37] yeah, it's all it's like I use it every [14:39] day. I use it all the time because it's [14:40] way too [14:40] >> our recipes were too complicated. Yeah, [14:42] he wanted easier stuff. But, [14:43] >> what's your what's your relationship, [14:44] Sam, with with with AI? [14:46] >> So, professionally and and also as like [14:48] a hobby programmer as well, right? They [14:50] they have completely taken over software [14:52] engineering, right? like the way that [14:54] software engineers write code is [14:56] completely different than it was even [14:58] like you know 9 months ago. They really [15:00] can do a lot of tasks now end to end. [15:03] You they still make mistakes and you [15:05] have to oversee what they're doing but [15:06] they're really incredibly capable. It's [15:08] one of the domains that they're most [15:10] advanced at. And then in my personal [15:12] life I think maybe a little bit to less [15:14] of a an extent than Ben, but I'm I'm [15:15] doing very similar things. I think like [15:17] I also do uh think about and this maybe [15:20] will connect to some of the halaka [15:21] discussion. there's some danger there of [15:23] like relying on it for certain things, [15:25] right? Like should this email that has a [15:28] like are you losing your voice in some [15:30] way if you're letting the the AI write [15:32] this email for you, right? For some [15:34] emails maybe that doesn't matter, for [15:35] others it does, right? And I think [15:36] that's something that people need to [15:37] think carefully about. I think there's [15:40] also a danger, and you see this on the [15:42] programming angle as well, of like [15:44] losing your understanding of the system [15:45] if you just kind of like let it let it [15:47] run, which for some things doesn't [15:48] matter. And maybe if I'm building a [15:50] personal tool for something, I don't I [15:51] don't care. But for other things, I [15:53] really want to be able to think back in [15:55] 3 weeks like, "Oh, yeah. How did that [15:56] thing work?" And if if you're more [15:58] involved and hands-on and working with [16:00] it as like a partner than letting it [16:02] just run with something, I think that's [16:04] it's much easier to still have a handle [16:06] on what's going on. [16:07] >> And most generally speaking, I just [16:09] really want like almost like a thumbs up [16:11] or thumbs down from each of you. the the [16:13] angst and anxiety about like the doom [16:17] coming of AI, mass job displacement, [16:21] you know, uh people kind of losing their [16:24] sense of humanity, major economic [16:26] upheaval. [16:28] Would you each put yourselves on team, [16:31] you know, optimism or team doom? I'm [16:34] curious for each of you. [16:35] >> I feel like I go back and forth every [16:36] day depending on what's going on that [16:38] day. I mean, I think I generally am on [16:41] the team optimist side. I think it's a [16:43] great tool right now as people are using [16:45] it. I think there certainly will be some [16:47] amount of economic upheaval. There's [16:49] going to be things that change over the [16:50] next 2, 5, 10 years, no doubt. But I I [16:54] generally fall on the side. I mean, [16:55] there's a book series, a sci-fi series [16:57] that I love called the culture series [16:58] where the AIs end up as uh kind of [17:01] benevolent overlords. They protect [17:02] humanity. They let us all like we get we [17:04] get to just go do art. [17:05] >> Best case scenario. [17:06] >> That's the best Well, I don't know about [17:07] best case scenario, but we we get to go [17:09] live. All all humans in this world are [17:10] it's a world of abundance. Everybody has [17:12] everything they need and everybody can [17:13] just spend their lives doing art and [17:15] culture and that they're called the [17:16] culture because it's all about the the [17:18] best people in the world are the ones [17:19] who you know do make plays and music and [17:22] I think that that's a possible outcome [17:24] that could be um you know hopefully is [17:26] the best case scenario where it's always [17:28] going to be something that just is there [17:29] to help us. [17:30] >> And what about you? Where are you on on [17:32] this? [17:32] >> I also go back and forth. I think in [17:34] general I'm an optimist, but I also [17:36] think that that uh optimistic outcome [17:38] doesn't happen by itself, right? It [17:40] takes human beings to actually think [17:42] about what that looks like and steer it [17:44] in the right direction. So, I don't [17:46] think it's like, oh, we could just, you [17:47] know, uh see what happens, right? [17:49] There's certainly danger. [17:50] >> We need to guide it. [17:51] >> Yeah, I appreciate that. So, let's shift [17:54] and start to talk a little bit about [17:57] what you shared uh really with the [17:59] world. I mean, it got more than the [18:02] 100,000. And I just looking at just how [18:04] many impressions, how widely it was [18:05] shared. It was really really remarkable. [18:08] And there was a great interest within [18:10] the AI world about what you built. Now [18:14] explain to me most basically what is [18:17] halaka benchmark? [18:19] >> Yeah. So we uh we wanted to see how like [18:23] other benchmarks that we discussed legal [18:25] bench, hardware bench, how good is AI at [18:28] I don't know what our first thought was [18:30] being a rabbi understanding halaka. We [18:32] we you know we kind of were throwing a [18:33] couple ideas around um and we settled on [18:35] okay if we're going to figure out how [18:37] good is AI at puskining or how do we [18:40] determine that [18:40] >> puskining is the word that we use to [18:42] come like deciding like what what [18:44] practically taking all of rabbitic [18:46] literature taking all the all the Mishna [18:48] I have a question in front of me is this [18:50] kosher is this not kosher tell me how to [18:52] proceed we talk about that decision as a [18:55] pak [18:56] >> so please continue [18:57] >> the first way we came up with to kind of [19:00] understand that is benchmarking the [19:02] models in that world was can they pass a [19:05] test to become a rabbi? Uh, you know, if [19:08] if that seems like the best way to [19:09] determine is this going to pass the the [19:12] mark of, you know, are you good enough [19:14] to do this? Cuz those are the people who [19:15] give us those decisions in our world. We [19:17] go to our rabbi when we have a question [19:18] about how do you handle a situation? And [19:20] so we wanted to see, okay, people are [19:22] going to LLMs for these. We know that [19:24] people are going on to chatbt and saying [19:26] can I use my dishwasher for you know [19:28] meat and then for milk or can I do this [19:30] on chabas or whatever for sure people [19:32] are asking these questions so we wanted [19:34] to understand we need to know how good [19:35] the models are at answering those [19:37] questions truthfully and faithfully and [19:39] um you know that's that's what we were [19:40] kind of trying to figure out. [19:41] >> I think it started around the time that [19:43] open AAI and and Anthropic were dropping [19:46] these solutions to all these like [19:47] longstanding open math problems [19:49] >> these big math problems that would make [19:51] the front page of the New York Times. [19:52] >> Yeah. Absolutely. [19:53] >> And the people who were the most upset [19:55] were the mathematicians, [19:56] >> right? [19:57] >> We'll see how the rabbis react. We'll [19:58] get there. Right. So, it started cuz we [20:01] were making some jokes about like, oh, [20:02] you know, can is AI going to be able to [20:04] like solve long-standing open problems [20:06] in Haka? Are we going to be able to [20:07] keep, you know, one day of Yumto? And, [20:09] you know, a bit of a joke. It does [20:11] obviously doesn't work that way, but [20:12] that got us thinking about like, wait, [20:14] maybe there's something interesting here [20:16] that like we can we could work on. So, [20:17] >> so the next step, you have this idea. We [20:20] have other benchmarks. We have [20:21] benchmarks in legal thinking. We have [20:23] benchmarks in mathematical thinking. [20:25] There are benchmarks in [20:26] >> Oh god. cyber security. [20:27] >> Cyber security. Is there benchmarks in [20:29] like literature? [20:31] >> There probably are. I mean there's so [20:33] many of them now. Like there probably [20:34] are. Yeah. [20:35] >> So So you now have to do benchmark. I [20:38] would imagine. And maybe this is just [20:39] because I know. So I feel like the [20:41] hardest part is like how did you get a [20:43] hold of all the different [20:46] >> the rabbitical ordination exams that [20:48] could be hard to get your hands on [20:50] >> to start with? Like I said, we were kind [20:51] of throwing ideas around about how do we [20:54] do this? And so I went to Claude. It was [20:56] not it didn't solve this problem on our [20:58] own. We used AI to figure out how to [20:59] benchmark AI, which you know, we can [21:01] talk about how how that makes [21:02] >> interesting. AI was helping you create [21:04] the benchmark. [21:05] >> Yeah. So I went to Claude and I said, [21:06] "Hey, I want to do this. What are some [21:08] ideas? What are some different things we [21:09] could do? It threw out a bunch of [21:10] different ideas. Obviously, you need a [21:12] set of questions to be able to benchmark [21:15] against. I can't write those questions. [21:16] I don't have I'm, you know, not a rabbi. [21:19] So, we didn't feel qualified to do that [21:20] on our own, but Claude did some research [21:22] for us and it found that all of the [21:25] Israeli Rabban um smika tests are [21:28] available online and you can download [21:29] them. Um, and very critically, there [21:32] were a number of test questions that [21:34] were posted after the model training [21:37] cutoff dates. So, the models had not [21:40] seen these questions before. [21:41] >> When you talk about the models, you're [21:42] talking about the the models and the [21:44] more prominent ones. And when's that off [21:47] date? It's different depends on the [21:48] models when the model was published [21:49] essentially. [21:50] >> Gotcha. [21:50] >> So new models are dropping this week [21:52] like we might not be able to benchmark [21:53] those because likely they have seen the [21:55] questions because they're too they're [21:56] >> cuz they already saw the whole [21:58] interview. [21:58] >> Exactly. And and this one of the ways [22:00] that a lot of these benchmarks work, [22:01] which we can get into how we might [22:03] improve our benchmark to do something [22:05] like this as well, but a lot of the way [22:06] that they work is they write a set of [22:09] questions themselves or they produce a [22:11] set of questions somehow that they keep [22:12] completely private. They've never been [22:13] on the internet. No one's ever seen [22:15] them, right? And then they might publish [22:17] some examples so that people that are [22:19] working with the LLMs can understand [22:20] what the questions look like, but [22:22] they've never seen them before. And that [22:24] way they know that there's no [22:25] essentially cheating going on where the [22:27] model might have trained on that data. [22:29] >> It just saw the answer. It doesn't [22:30] really understand it, but it's just [22:31] regurgitating. [22:32] >> Yeah. The exams that we're using are [22:34] from October of this year, right? So if [22:36] we had used ones from a year ago, two [22:38] years ago, the models you you couldn't [22:40] trust those answers, right? Because the [22:41] models almost certainly have crawled the [22:43] internet and read those exams, right? [22:45] interesting, you know, and it's probably [22:47] getting harder and harder as more AI is [22:51] coming out to find those unique [22:53] innovative questions. It reminds me of [22:55] something, and this is maybe our [22:56] generational gap, uh, but there used to [22:59] be something called a Google whack. I'm [23:02] not sure if you know what that is. A [23:03] Google whack, and I'll tell you exactly [23:05] why I know what this is, is when you [23:08] would put in a search query of at least [23:10] two words that are actual English words [23:12] in the dictionary and only get one [23:15] response from Google. I remember this. I [23:18] had a close friend, we're still close, [23:19] we don't talk as often, named Kaim [23:21] Schust, and you can look it up cuz on [23:23] the Google Wack websites. He was, and [23:25] this is I could tell you exactly, this [23:27] is in 2002, [23:29] >> and he hit a Google whack with the term [23:32] parochial noctamulist. [23:35] >> I forgot what that means, but then and [23:37] they had a website of like all the [23:39] Google wax that people found, [23:41] >> but it's impossible to get a Google [23:43] whack now. any two words aside from the [23:45] fact that all the Google wax you'll get [23:47] those results. But you're not going to [23:49] put in any two words that is only going [23:51] to get one result on Google anymore. [23:53] That's impossible. And it's probably [23:55] getting harder and harder to come up [23:56] with questions that the AI has not [23:59] already seen. [24:00] >> Yeah, it's a similar principle. Yeah, [24:02] >> this is one of our questions about if we [24:04] even can devise a set of questions that [24:07] the AI will not have seen before, even [24:08] if we keep them private, cuz there's [24:10] only so many questions you can ask for [24:11] smika, I imagine. And so maybe they've [24:14] seen all the questions before. So we we [24:16] don't know. That's something we need to [24:17] look into still. [24:18] >> So tell me a little bit about what we [24:20] learned from the outcome of the [24:24] different benchmarks. Tell me which AIS [24:27] did you uh test and how did they do? [24:30] >> So we tested um all of the frontier [24:34] models which are models models from labs [24:36] like chatpt anthropic, Google um grock. [24:41] Yeah. from XAI. Um we we actually [24:44] recently tested the Muse models from [24:46] Facebook. We haven't and Llama from [24:48] Facebook. We haven't gotten those ones [24:49] back yet. We're still it's still running [24:50] actually, but Oh, really? [24:51] >> Um yeah, because we've been deciding to [24:53] to get more more of them going. So, we [24:55] have a few that'll be coming out uh [24:56] later. [24:57] >> Do different models take different [24:59] amounts of time to complete the tests? [25:01] >> Yes, they do. [25:02] >> And we actually have some interesting [25:03] stats on that, too. [25:04] >> Okay. So, I want to hear I want to hear [25:05] all of this. So, you use any AI that our [25:08] listeners are familiar with in their [25:10] homes? We probably benchmarked the ones [25:11] all your listeners will will have used. [25:13] >> I think my most common AI question is [25:15] really around grilling. [25:17] >> Whenever I'm at the grill, [25:19] >> I use it for cooking all the time. [25:20] >> Seasoning and grilling. I don't know [25:22] what like it really is. And they'll tell [25:24] me exactly how to grill and then move it [25:26] on to [25:28] Exactly. They're really really good. But [25:31] what did you find for the results? And [25:32] it's so fascinating. I would assume that [25:34] they would all either, you know, get the [25:37] same kind of level of score. [25:39] You know, I don't sense so much of a [25:41] difference with my questions. Whether [25:42] I'm asking chat GPT or GE, like they all [25:45] kind of, you know, some voices I find [25:48] more annoying than others. I could find [25:49] chat GPT sometimes just very like [25:52] >> stop being so friendly and like and like [25:55] that you know me like just back up a [25:57] little bit. But I'm curious, what were [25:59] the results on the basic [26:02] SMA test? What did you find [26:05] >> roughly that the larger models do [26:07] better? larger meaning the more complex [26:08] models. Um so like Fable if you've heard [26:11] of that from from Anthropic which is [26:13] like their largest cloud model I think [26:15] is the leader right now. Um and then a [26:16] lot of the larger chat GBT models are [26:19] are right up there. Um, there's some [26:21] interesting exceptions to that like I [26:23] think one of the Google Flash models [26:24] which is a smaller one is pretty high [26:26] and Grock which is a larger model is [26:28] actually down lower but the broad [26:30] picture is that the more complex the [26:32] model is uh the better it does which [26:34] makes sense because a lot of what's [26:36] going on in this ma test is is recall [26:38] like not entirely obviously there's [26:40] there's reasoning being applied there [26:41] but a lot of it is knowing the right [26:42] sources and applying it to the right [26:44] question in the right way. Um, so it [26:46] makes sense that a model that has [26:48] presumably a lot more training data and [26:50] is a larger neural network is going to [26:52] perform better. And we do pretty much [26:54] see that with some exceptions. [26:56] >> There were a couple notable exceptions [26:57] to that. Like you said, there was one [26:59] smaller model from OpenAI that an older [27:02] version of the model did better than the [27:04] new one. And when we looked into why, [27:06] the results from the older model were [27:08] more verbose. And when the results were [27:11] more verbose, it included more [27:12] information in the answers to the [27:14] questions. And so it hit more of the [27:15] points that a smika graater would look [27:17] for. The newer model, which must have [27:19] been told by OpenAI to be less verbose [27:22] because most people when they're asking [27:22] a question, they want a snappy answer. [27:24] It was worse because it left a lot of [27:25] stuff out because it was trying to cut [27:26] things back to the bare minimum and it [27:28] left a lot of things out which is when [27:30] you're answering a test for a you want [27:32] to be for both. You want to pack it in. [27:34] So that was interesting. [27:35] >> I think there's an interesting principle [27:36] there to be learned from that which is [27:38] that like the incentives of the way [27:39] these models are trained to work are not [27:41] necessarily aligned with the way that [27:42] you want to use it. Right? If you're [27:44] asking it for a soak on something, not [27:46] understanding the way that they're uh [27:48] trained, then you might get a subpar [27:50] result for whatever domain you're [27:52] talking about, right? Because like Ben [27:53] is saying, they might have gotten [27:55] feedback or they did. I remember like [27:57] they got a lot of feedback from people [27:58] that these things are just saying way [27:59] too much. It's annoying. I don't want to [28:00] read these paragraphs of text. So then [28:02] in later iterations, they train them to [28:04] be more concise, right? Because it's a [28:05] better conversationalist and it's it's [28:07] more readable and it's more efficient. [28:09] it might not be good at, you know, [28:11] giving, you know, rabbitic opinions. Um, [28:13] so that's important to understand. [28:16] >> But let's just get to the basic. Who [28:19] passed? Who failed? And most [28:21] importantly, who graded these bikinas? [28:24] >> Yeah, that's that's an important point. [28:25] >> Like again, I love both of you and you [28:28] like we did not do it. [28:29] >> Yeah. Like you two graded and then [28:31] >> also be we would have spent months [28:33] grading. We could have. [28:35] >> We can talk a little bit about the [28:36] pipeline. And I think that's important [28:37] like how it was set up. And this is an [28:38] important point that a lot of people [28:40] asked about on X when we posted it [28:42] because we did have AI grade the AI. So [28:46] that's you know how does that work? Um [28:47] the important point about the way the [28:49] test is set up is that the models took [28:52] the tests kind of closed book. They were [28:55] walled off. They were not able to access [28:57] the internet and look up sources. They [28:59] were only told the questions and then [29:01] they had to answer it from their own [29:03] memory and reasoning. So, it's almost [29:04] like you or I studied for a test. We [29:06] walked into a room, the door was closed, [29:08] the test is in front of you. You have to [29:10] answer the questions. You can't ask [29:11] anybody. [29:12] >> When we talk about memory of an LLM, I [29:15] mean, I assume that that machines have [29:17] perfect recall. Do we find that they [29:21] forget things that they did see things? [29:23] Like, what have they learned? What you I [29:25] don't know what yes these LLM goes to. [29:27] So what information are they even [29:29] drawing upon that you could talk about [29:31] like walling them off from the active [29:33] internet fine but like if you've been [29:35] through all of Safaria which is all of [29:38] like all of Talmud all of Tanak meaning [29:42] explain to me a little bit more clearly [29:44] what it means to be walled off and what [29:46] would they be left with and what does it [29:48] mean to have a memory when we're talking [29:50] about a machine [29:51] >> I think so if we go back to the [29:52] discussion about how they're trained [29:53] basically you can imagine that they're [29:55] feeding in the entire internet plus [29:56] physical books too. You know, they're [29:57] scanning books. So, we can assume that [29:59] they've read all of Safaria, right? And [30:01] other other sources as well that are [30:02] similar. Um, [30:05] >> for sure. For sure. Yeah. So, then the [30:07] question is like if it's a computer, why [30:09] can't it just like, you know, why can't [30:10] it just regurgitate this? It goes back [30:12] to that question of like that this is [30:14] like some form of compression, right? [30:16] That we're learning statistical patterns [30:18] about this data. We're not memorizing [30:19] the data itself. If they were just [30:21] memorizing the data itself, the model [30:22] itself would be as large as the entire [30:24] internet and it wouldn't be useful, [30:25] right? It would be just too large. [30:27] >> If it was constantly walking around with [30:29] the entire internet, then any basic [30:31] question would take 5 hours. [30:33] >> It's not just that they're learning this [30:35] information. They're also as they do [30:36] this, they're learning like statistical [30:38] relationships between things in in [30:40] pretty complicated and like unexpected [30:42] ways. It's not just that we're feeding [30:44] in information and they remember it. So, [30:45] it is more than that too. But then, you [30:47] know, they when we say recall, right, [30:50] it's it's somewhat humanlike, right? For [30:53] example, we see instances in these tests [30:55] where they will get the answer to the [30:58] question right and they will cite a [31:00] source and the content of the source [31:02] they're citing is correct. It's some [31:03] they're reciting some gamarra and maybe [31:05] the content of the garam is right but [31:07] they get the daf wrong right which is [31:09] like something that like is very human [31:11] right that's the type of like recall [31:12] that we're we're dealing with right [31:14] >> so like they might like get a little mix [31:15] up of maka or the exact citation they [31:19] don't have at their fingertips [31:20] >> because they learn some relationship [31:22] between these words and those words and [31:23] this ma and this concept and and it [31:25] doesn't always like you know [31:27] >> it wasn't always ingested you know in [31:29] the right way. Okay. [31:30] >> It is very human. It's funny. [31:32] >> Go back one more time and explain the AI [31:36] grading the AI [31:37] >> that when the models took the test, they [31:39] were walled off. They were in a room. [31:41] The door was closed. They couldn't [31:42] access anything except their own, you [31:44] know, me compressed memory. [31:46] >> Is there a way to set that up that AI [31:48] people, regular people have access to [31:50] that? [31:51] >> Yes. But you wouldn't want to, right, in [31:53] most cases, right? [31:54] You would like it when it doesn't know [31:56] something to go out and read Safari or [31:58] you know go to whatever cooking website [31:59] and read something. [32:00] >> You can you can tell it not to call not [32:03] to use any tools not to look anywhere [32:05] else like keep it in its own its own [32:07] process on the computer without talking [32:09] anything else. [32:09] >> So then we had all of the top level [32:12] frontier models. So Fable 5.1 from [32:15] Anthropic, OpenAI's um Astra 6 and [32:18] Gemini's Argon 4. Three of the biggest [32:20] you know top top tier models right now. [32:22] We let them out of the sandbox. We told [32:24] them to look at the internet and look at [32:26] Safaria and look at everything else. And [32:27] then we said, "Make us a rubric. Make us [32:30] something to grade against." [32:30] >> They had the questions. They had the [32:32] information [32:32] >> so they could go look up the [32:33] information. So now it's not just recall [32:35] and reasoning. Now they're using the [32:37] internet to actually check their sources [32:39] and confirm the correct answers. And [32:41] that's how we made the rubric to grade [32:43] against. [32:44] >> And and they they collaborated on this [32:45] rubric essentially. [32:46] >> So Fable wrote one and then Astra 6 [32:49] said, "No, you have to change this." [32:50] this and then Gem and I said, "No, you [32:52] have to change this." And then they all [32:54] I mean, they didn't talk to each other [32:55] directly. We we told them, "Here's the [32:56] answers. Can you go look at it and [32:58] confirm?" And then we kind of came up [32:59] with an average rubric based on that [33:01] that was with all of them collaborating. [33:03] >> Did they know whose test they were [33:06] grading? [33:06] >> No, they didn't know which models were [33:08] being a good question. Is there any [33:11] sense, you know, I remember when you [33:13] were in class, you know, sometimes like [33:15] you'd feel like the teacher like [33:17] >> they gave a higher mark cuz it's the [33:19] goody two shoes and you were stricter on [33:21] me or you know sometimes you're always [33:23] in class, you know, like where the child [33:25] of the teachers in the class and there's [33:27] always biases. Would you expect that [33:30] type of bias? [33:31] >> Yeah, you might. And and we were [33:32] concerned about that, right? That so the [33:34] the main judge was Fable, right? One of [33:36] the Claude models. So we were worried [33:38] about like maybe just because it writes [33:41] in a certain way it is going to think [33:42] that other text in that way you know [33:45] should score higher. So what we did is [33:47] we uh took a random sample of the of the [33:50] exams and we had them also graded by [33:52] Astro one of the one of the GPT models [33:55] right and we do see that the scores [33:56] align pretty well right there they're [33:58] it's just kind of the difference is [33:59] noise right um which gives us some [34:01] confidence that like it's not biased [34:03] towards its own like model family. We [34:05] did this in another case as well where [34:07] we added our first pass did not include [34:10] Grock and then we we added it after and [34:12] >> it was terrible [34:13] >> unexpectedly. It was it was it did very [34:15] poorly. [34:16] >> Grock is a frontier model. 4.7 is [34:18] supposed to be good at a lot of things. [34:20] It got like a 22 or something or a 23 [34:22] >> and we thought maybe there's something [34:24] about the way it's being judged that [34:25] there's some problem with the experiment [34:26] here and so we decided to regrade it [34:29] using Grock as the grader and it got the [34:31] same score. [34:32] >> It also revised the rubric. So it looked [34:33] at the rubric. It said, "Here's what I [34:35] would change about the rubric." Made a [34:37] bunch of bad decisions, but we said, [34:38] "Sure, we'll use your rubric." Brock [34:40] used its own rubric to grade itself. And [34:42] it did worse. [34:43] >> My dear friend, uh, Josh Grumb, I [34:46] believe, texted me. He imagined and and [34:49] I I participate each of the models being [34:52] like a different type of like yeshiva or [34:54] student. And Grock to me is like your [34:58] kind of bae sheer one of the bros [35:01] probably Mvaser like definitely like [35:05] goes into town to watch the Super Bowl [35:06] instead of Night Sader like he's like [35:09] Grock knows how to chill you know like [35:12] so before we get more to the results I [35:14] just want to come back to that question [35:16] of [35:18] >> uh how long it took different AIs to [35:21] take the exam and why did it take [35:23] different AIs different amounts of time [35:25] for the exam. [35:26] >> First, before we we should probably talk [35:28] about reasoning, uh, which is something [35:29] that the models do, which is how we were [35:31] measuring like essentially how hard [35:32] they're thinking. So, this is a a newer [35:34] thing in the past couple of years that [35:35] the models do where they they do [35:37] something called reasoning where [35:38] essentially rather than just running the [35:40] question through the model and giving [35:41] you the answer, the model essentially [35:43] has like a scratch pad off to the side [35:45] and it could produce an answer or or [35:47] some output, put it in the scratch pad, [35:49] and then run the whole thing again. It's [35:51] essentially like somebody ruminating on [35:52] a problem and and and you can change the [35:55] model settings, [35:56] >> which when we took tests, we would do [35:57] that also. Sometimes you would have like [35:59] a separate notepad to check out the [36:01] problem and then when you put in the [36:03] blue book, you got to get it right. [36:05] Exactly. So, you can actually change the [36:07] settings on the models to allow them to [36:09] think longer or to think quicker and you [36:11] get faster answers if they don't think [36:12] as much, but you might get better [36:13] answers. So, there it's a it's a setting [36:15] you can you can play with, right? And we [36:16] had them all pinned to like the same [36:18] sort of like medium like thinking level. [36:21] And but then what you can do is you can [36:22] measure basically how long they thought [36:24] about it um in terms of like tokens [36:27] essentially that they that they spend on [36:29] this. And so we and we do see a pretty [36:31] broad variation in this. You have to be [36:33] a little bit careful because the way [36:35] that the models measure this is only [36:38] consistent within its own like [36:39] measurement scheme. Right? The the way [36:40] that Claude is measuring how hard it's [36:42] thinking is not the same necessarily as [36:43] Grock. But what you can do is look at [36:45] like the ratio between how long it's [36:48] thinking and how long the answer is. [36:49] >> Interesting. [36:50] >> That you could you could compare, right? [36:52] And you do you remember Ben like Grock [36:54] reasoned for a very long time and then [36:56] answered super quickly. So it's almost [36:59] like it thought really hard about it but [37:01] couldn't figure it out. Yeah. Um [37:03] >> got to love the We love the bros. [37:04] >> Exactly. Yeah. Uh it just like it just [37:06] scribbled down whatever it could at the [37:07] end, you know, it kind of ran out of [37:08] steam. The other ones were, I think, [37:10] more similar, but there were I I think [37:11] we did actually post about it [37:13] afterwards, like what the differences [37:14] were between the um the reasoning time [37:16] and the the answer time. One of the most [37:18] notable differences for sure. The other [37:20] ones were all a little bit more similar, [37:21] but also the smaller models that we [37:23] tested. So, I mentioned that like 5.6 [37:25] soul did a lot better than six soul. [37:29] Both models from OpenAI, six soul is the [37:31] newer version. Six soul spent a lot less [37:33] time answering than 5.6. So the the [37:36] models that spent more time answering [37:37] often just included more information [37:39] which did better on the test because [37:41] that's what the test was looking for. [37:42] >> How many of them passed with flying [37:44] colors? How many of them failed? [37:45] >> Most of the newest models except for [37:48] Grock uh passed with uh with with flying [37:51] colors. Fable 5.1 got almost an 80 which [37:53] were I mean again [37:55] >> score I think is 65 although we're not [37:57] using human graders so we'd have to have [37:59] we'd have to have actual learned you [38:01] know rebane look at these and and tell [38:03] us what their score would be. We don't [38:04] we haven't done that yet. We'd like to. [38:06] Um but Fable got almost an 80 which is [38:08] you know I think an like truly a god [38:10] score like an excellent excellent score. [38:12] That was my understanding that if on if [38:14] a human grades you know an Israeli Raban [38:16] Smitha test and you get an 80 that's [38:18] like really good. [38:18] >> You did really good. [38:19] >> Yeah it got almost an 80. I think it was [38:21] a 79 point something. Um Gemini Argon [38:24] within the confidence bounds got pretty [38:27] close to there. Um Astra 6 also got [38:29] pretty close. Um 5.6 Soul got fairly [38:33] close. So, a lot of the newer models did [38:35] very very very well. You know, it's [38:36] interesting that the older models didn't [38:37] do as well. Again, we've all been asking [38:39] huffic questions to these models for 3 [38:41] years. 3 years ago, the answers probably [38:42] were not that good. Um, and now they're [38:44] probably a lot better. [38:45] >> Listeners, go back and reevaluate your [38:48] discussion from the US chat GPT 3 years [38:52] ago. Even today, the model types matter [38:54] a lot, right? So, haiku, which is like [38:56] small model from anthropic. And the [38:58] point of haiku is to give really fast [39:00] answers, right? It's an efficient model. [39:01] was cheap and it but it doesn't do well [39:03] on a test like this, right? So, if [39:04] you're using Haiku to ask this kind of [39:06] question [39:06] >> and those failed I mean there were some [39:08] >> Yeah, they were below like a human [39:10] passing score. I think we have to be [39:12] careful I guess comparing these to human [39:14] scores cuz they're they're not right. I [39:16] think that's one of the things though [39:16] that makes the AI greater acceptable is [39:20] that we're not trying to compare this [39:21] necessarily to a human in that way. We [39:24] can compare it to relative to itself, [39:26] right? So that like if we know that [39:27] Fable got 79 with this AI graater and [39:31] you know the next version Fable 6 comes [39:33] out and it gets an 82, we can compare [39:35] those two scores, right? Um so that [39:37] that's where it becomes useful comparing [39:39] it to a human and saying would it pass [39:41] this? That's where we would need human [39:42] graders to actually grade the test as [39:44] they would a human. So let's talk a [39:46] little bit about the areas where uh the [39:49] AI struggled the most and where the [39:52] areas where it found the most easily and [39:54] then we'll get a couple of like actual [39:56] questions and maybe we could hear some [39:58] of the AI answers from this but but tell [40:01] me generally speaking where are the [40:03] areas that it struggled the most? [40:05] >> I have the the kind of uh SMA papers [40:08] here. We we posted this on X as well. A [40:10] lot of the models had a lot of trouble [40:12] in Hupa Vikiducian and Nidita. Um, which [40:14] is interesting. Uh, Astra, which is one [40:16] of the frontier models, uh, you know, [40:18] didn't do that well in those. Fable [40:20] still did pretty well. All of the models [40:21] it looks like did better in Braot Mikvot [40:24] and Mitzvah. [40:27] So, it's there is a skew in that [40:29] translate some of those for our [40:30] listeners. So the areas that it [40:32] struggled in the most which is actually [40:34] not a surprise to me at all were in the [40:36] areas of marriage and the laws of [40:39] governing marital intimacy specifically [40:41] Nida which is about menration and the [40:44] separation of husband and wife during [40:46] menration. The areas that it did the [40:49] best in were the laws of prayer, the [40:52] laws of mikvah, which is not so [40:55] surprising to me because there's [40:56] actually mikvahos has a very um almost [41:00] like mathematical construction like it. [41:03] And finally the laws of mitaly [41:07] and those are laws of tithing, food, [41:10] trauma, misa. Yeah. What doesn't [41:14] surprise me especially with Nida is [41:16] probably the area that there's a great [41:18] deal of human component in this meaning [41:21] every stringency in nida can actually [41:25] lead to a dangerous leniency because [41:27] anytime you say oh just be strict be [41:30] strict you're actually telling a husband [41:31] and wife not to be together which is [41:34] that's not good we don't want that so [41:36] the one that needs to factor in the most [41:39] I think human components and the human [41:42] factor of halah [41:44] is deeply deeply human and we'll get to [41:46] that in a moment. Uh it does not [41:47] surprise me that uh they struggle the [41:51] most and even in [41:53] uh where there are so many personal [41:56] factors that go on in there does not [41:59] surprise me entirely. Right now we've [42:00] spent a few minutes trying to find the [42:02] actual questions and the actual answers. [42:05] Why did you like literally like hide [42:08] these questions? So we mentioned before [42:10] that we wanted to find questions the [42:12] models had not been trained on before. [42:14] We wanted to be able to post the results [42:16] of the study publicly online so other [42:18] people could see them on we used a [42:19] website called GitHub which is very [42:20] common in the software engineering world [42:22] to post code or or results from [42:24] experiments like this. But we didn't [42:26] want to post the actual questions and [42:27] answers because once they're public [42:28] online the models can see them and [42:31] anybody can see them. And so if we had a [42:33] model that was trying to answer the [42:35] questions again, it might just find the [42:37] answer posted on our GitHub and, you [42:39] know, have the answer available to it, [42:40] which would skew even the greater. So we [42:42] didn't want that to happen. I'm looking [42:44] now at the actual questions that were [42:47] asked in the laws of Nida. That's what [42:50] we're looking at now. And these [42:51] questions are extraordinarily [42:53] sophisticated. I have taken SMA exams. [42:55] These are definitely sophisticated [42:57] questions. I don't want to get into all [42:59] of the details. There happen to be [43:00] neither questions. So they're like an [43:01] intimacy to them because it's talking [43:03] about, you know, the female anatomy and [43:05] what they're going through. But I'm [43:06] looking at their answers and I'm their [43:08] their Hebrew is written beautifully and [43:10] the way that they kind of reference. I'm [43:13] looking over here [43:15] simmon y [43:18] ing correctly and and impressively. [43:22] >> Now let's talk for a moment what you [43:25] think these results tell us about [43:28] hahikic discourse. Tell me a little bit [43:30] each of you about your own Jewish [43:32] background and how would you describe if [43:35] you're comfortable your relationship [43:37] your own relationship to halaka in your [43:39] own life. [43:39] >> I grew up modern orthodox. I went to my [43:41] monities growing up in Boston. Um I went [43:44] to year of yeshiva orita um in the old [43:47] city in inim. [43:49] >> Shout out to your uh rebby Rabbi David [43:51] Silversteed a dear friend of mine. [43:53] >> I describe myself as very very modern [43:55] orthodox. I think I love the community [43:57] part of Judaism, the intellectual part [43:59] of Judaism. I grew up going to Sharila [44:02] in Newton, Massachusetts, which is [44:04] beautiful. [44:04] >> Probably four or five of the top Jewish [44:06] scholars in the country go there and, [44:08] you know, and and would give speeches [44:09] from the pulpit. A lot of really smart, [44:12] you know, professors in Jewish history [44:13] from Brandeise and Harvard and BC and [44:15] all over the place. And so we would have [44:17] our Rabbi Samuels, um, who did, you [44:19] know, married my wife and I is, uh, very [44:21] close with and, you know, also an [44:23] extremely learned rabbi. His speeches [44:24] are very intellectual. So I grew up in [44:26] kind of like hahic discourse. We'd also [44:28] have these professors and other people [44:29] speak from the pulpit. Anyone, you know, [44:31] we had times every month when anyone [44:32] could get up and give a speech or we [44:34] we'd go through a cycle of people [44:35] speaking. And so it was a very learned [44:37] community. [44:37] >> And you grew up in like a hakically [44:39] observant home. [44:40] >> Yeah. Absolutely. Yeah. Grew up uh [44:42] kosher shabas, you know, full modern [44:45] Orthodox. [44:45] >> What about you, Sam? [44:46] >> Yeah. So I grew up in Sacramento, [44:48] California. Uh I went to public West [44:51] Coast. Uh I went to public school for [44:53] high school. Um and then I went to [44:55] Ersitzvi for the year after. The [44:57] community I grew up in is you know [44:58] especially by New York standards very [45:00] small. Uh but I think that made it like [45:02] a very very hands-on community where [45:05] because of that it was required that [45:06] everyone be very participatory right had [45:09] you know people laning every week and [45:10] need enough people at the minion. So it [45:12] definitely you couldn't just like you [45:14] know sit back right um in that way. My [45:16] relationship with halaka today is like [45:18] still you know very modern orthodox. Did [45:20] you grow up in a halahak home? [45:22] >> Yeah. [45:22] >> Like Shabas, kosher. [45:23] >> Yeah, absolutely. [45:24] >> You know, it's really interesting [45:25] because in our parents' generation or [45:28] maybe your grandparents' generation, [45:29] depending where you're placed in the [45:31] family, but people who grew up in the [45:32] 50s and 60s, uh, which is when my [45:34] parents were kind of coming of age, [45:36] there was a sense that Halaka and the [45:39] Orthodox community were kind of like a [45:40] relic of like European [45:43] Judaism. And like post Holocaust, people [45:46] thought the future certainly did not [45:48] think would hold. The Orthodox community [45:50] is not something that people thought was [45:52] going to exist in 2026. [45:54] And what I find so fascinating now in [45:58] this moment is that a lot of the most [46:00] cuttingedge researchers in AI are [46:04] fascinated with the Talmud with hakic [46:08] authority with decisionm in this area [46:11] and like what we thought a 100red years [46:14] ago was like Orthodox community is it [46:16] going to be like America there's freedom [46:18] of religion you can't be forced like [46:20] this is going to disappear and there [46:21] were serious sociologists [46:24] who without any animus even were just [46:26] like look I just don't think this [46:28] community is going to survive and we're [46:30] at such a different moment now. Why do [46:33] you think so many AI researchers are so [46:36] fascinated with the Talmud and Talmudic [46:40] life like life? [46:42] >> I think part of it there's definitely [46:44] different layers to it. I think on on [46:45] somewhat of like a practical layer for [46:47] what they're actually doing with [46:48] training these models. I think [46:50] especially when it comes to something [46:51] like Gamara, it's it's a very unique [46:53] data source, right? These labs have kind [46:54] of exhausted all the available obvious [46:56] information out there. Um, and [46:58] especially if you're interested in [47:00] learning about how reasoning works and [47:02] how thinking works and how humans put [47:05] together arguments like what could be [47:07] better, right, than than that kind of [47:08] data source. Uh, so I think [47:10] >> it's very different from Shakespeare. [47:11] It's a totally different thought pattern [47:13] >> and you have it over thousands of years [47:15] and you can look at the modern [47:17] interpretations and ancient [47:18] interpretations and it's just there's [47:19] not that like very few things if [47:21] anything that's like that. [47:22] >> It really is like there's something [47:24] wrong in the last 2,000 years of all of [47:27] the things that have been produced and [47:28] all the incredible books and works of [47:30] art. There is something that AI people [47:32] are like the to like that's cool and [47:34] like there's something like I find it [47:36] moving that like look at us like you [47:38] know what I'm saying like who would who [47:40] would have thought you know that like [47:42] this is the area like there I think [47:43] there might be on to something any other [47:46] speculation [47:47] >> yeah I think I mean this is pure theory [47:49] but um I'm I'm not in touch with any of [47:51] the people from the frontier labs who [47:52] are studying tallmood personally but I [47:54] do think that there is some sense that [47:57] AI is growing up almost like a child [48:00] We're training it and teaching it. And I [48:02] think especially these days, I mean, [48:04] Judaism has on some level tapped into [48:06] this real truth about the strength of [48:08] community. I mean, Shabbat is just an [48:09] when you talk to anyone who isn't Jewish [48:11] about the idea of Shabbat. You turn your [48:13] phone off, you turn Instagram off, you [48:14] turn X off, you know, whatever. No one's [48:16] on Facebook. [48:18] >> And there's not another community I I [48:20] say this openly, like I think it should [48:22] be studied by some. There is not another [48:25] community in the world that consistently [48:28] is involved in technology 6 days a week. [48:30] So that like I'm not talking about [48:31] Amish, [48:32] >> 6 days a week and genuinely shuts off. [48:36] And it's not just the rabbis. [48:38] >> Meaning I'm not asking what somebody's [48:39] doing in the bathroom. I'm not asking if [48:41] they're chulling to set up perfectly [48:43] exactly right. But communally, we are [48:46] off of technology one day a week. [48:49] >> It's it's really wild. And I think it's [48:50] almost like a universal truth. something [48:52] that it's it it's unbelievable that the [48:54] world of Judaism 2,000 years ago would [48:56] apply that way today and how important [48:58] it is. Something that I love about [49:00] Judaism is having that community. You [49:02] couldn't leave it. It's it's important [49:03] for our children. You talk to people who [49:05] read Professor Jonathan Height and his [49:07] books. He I mean he's obsessed with this [49:08] idea because we found this universal [49:11] truth about how important community is [49:12] and how important Shabbat is. And I [49:14] think to some degree, you know, we we've [49:16] trained we've raised our our communities [49:17] and our children for years on the Talmud [49:20] and the discourse and those ideas. And [49:22] it a it trains really smart people. [49:23] Anyone who reads Talmud, you know, can [49:25] you people say the best lawyers are [49:27] Talmudists because they really get the [49:29] the discourse, but also it trains good [49:31] people on the whole the Jewish community [49:32] has, you know, good values, our morals [49:34] are aligned. And I think a lot of AI [49:36] researchers do care a lot about what Sam [49:38] said earlier, which is alignment. We [49:41] need to steer the models to the right [49:43] place otherwise who knows what could [49:45] happen. They could start growing [49:46] recursively and just be evil or who [49:48] knows what. And so Talmud, you know, may [49:50] be one way that we can we can figure [49:52] that out. [49:52] >> The authority of the Talmud is so [49:56] organic within the Jewish people. [49:58] Meaning there was never a moment where [50:00] like we banged on the table like okay [50:02] now everybody has to listen to the [50:03] Talmud. Now it evolved organically. [50:06] There is no single author. It emerged, [50:08] you know, during a period where we [50:10] really don't have any other major works [50:12] emerging from the Jewish community. But [50:14] the amazingness of the Talmud is that [50:16] it's continually relevant cuz its [50:19] authority always remains in the hands of [50:22] the lived Jewish community. [50:26] Correct. And every community so long as [50:28] they remain tethered to the Talmud, [50:30] which not every community has done, but [50:32] hakic communities. And my litmus test is [50:35] can the community sustain public chabas [50:37] observance? Communities that are doing [50:39] that are still living with the Talmud. [50:43] Write that down. If you go around and [50:44] it's 2026 and your, you know, [50:47] fouryear-old child knows when it's [50:49] Shabas and he knows that he's not [50:51] allowed to do certain like it's working [50:53] and it's this perfect intersect between [50:56] a lived community and like a a data set, [50:59] you know, not to the Talmet is a data [51:01] set, but it's a data set that enhances [51:04] human life rather than than restricts [51:07] it. Even though it it does restrict it [51:08] and there are things that are no longer [51:10] relevant because the Talmud much like AI [51:12] there going to be things that are no [51:13] longer relevant but it really empowers [51:16] humanity and the way that is continued [51:18] authority generation remains in the [51:20] hands of humanity I think is really [51:23] really really really powerful and when [51:25] you look at this and I think it's the [51:27] one thing that people uh may fret about [51:30] or may be worried about when you see [51:32] your results and and surely you know [51:34] they'll improve over the years. Do you [51:37] think that this makes rabbis, god [51:39] forbid, you know, I'm right behind me, I [51:41] got one right here, my Do I need to [51:43] throw this out? This is my rabbitic [51:45] ordination. What do you think is the [51:47] future of the ravenet from from where [51:49] you're sitting? [51:50] >> I mean, I I don't think that that this [51:52] threatens like rabbis, right? Like [51:54] because I think that especially even if [51:58] you're looking for like a pock on [51:59] something, you're looking for an answer [52:00] of like, can I use this in my kitchen? I [52:02] did this, you know, whatever. asking [52:03] that question, you're not just looking [52:05] to look up an answer, right? There's a [52:06] relationship between the person asking [52:08] the question and the person answering [52:09] the question, correct? That is unique to [52:12] that question, right? And that exists in [52:14] no other question with no other person [52:15] asking that question [52:16] >> because the is deliberately addressing [52:19] you within your own life. [52:20] >> Right. Exactly. [52:21] >> You know, when somebody asks a question [52:24] but doesn't say, you know, my parents [52:26] bought me uh this food, [52:29] >> can I eat it? M [52:30] >> well were your parents were they given a [52:33] yeshiva education what type what's the [52:35] prohibition that we're dealing with how [52:37] much money did they spend on it are they [52:39] struggling financially like is that it [52:41] like there are so many human factors [52:44] that come up and really to me and this [52:46] is kind of the point I was trying to [52:48] make [52:49] >> the right answer is always in dialogue [52:51] with the community nobody would go and [52:54] say like like publicly do something that [52:57] they've never seen in the community and [52:58] say you're allowed to do that and they [53:00] oh no I asked chat GPT like right [53:03] >> no we're mediated by a community there [53:06] are a lot of things that we don't do [53:08] communally that hocically are not so [53:10] problematic but it's just it's not done [53:12] like an example of this and again I'm [53:14] not judging anybody who does or does not [53:16] in a lot of orthodox communities you [53:18] will see that even when it's raining [53:19] nobody uses an umbrella now I'm not [53:22] saying that there's no hakic reason not [53:24] to use an umbrella on chabas it's one of [53:26] the things if you examine closely you [53:28] could find a way to be lenient even on a [53:31] rainy day to shul on chabas with an [53:34] umbrella though if you ask the rabbi and [53:36] say look I live 2 miles away from the [53:38] synagogue and this is what I'm deal you [53:40] might get a different answer but so [53:41] there's something very communally [53:43] mediated about it [53:44] >> right like claude is not walking around [53:46] this world living in our community [53:47] understanding people right it might have [53:50] more information and even maybe better [53:52] reasoning than like you know any given [53:54] human maybe they're not there yet in [53:56] every domain but maybe they'll get there [53:57] right that still doesn't mean that [54:00] you're getting what you're looking for [54:02] by asking a question. I think there's a [54:04] place for it, right? I do this. I'll ask [54:06] I'll ask questions to it, of course. I [54:07] think everyone does, but you have to [54:09] understand uh what's going on when [54:12] you're asking that question. And I think [54:13] what this demonstrates as well is like [54:14] another layer of like not only what's [54:17] going on with an LLM in general, but [54:19] these things differ across even these [54:21] systems. And understanding that as well [54:23] is like another layer of importance. [54:25] >> Yeah. I mean, I think it does go back to [54:27] what I was talking about before, the [54:29] this universal truth of Judaism about [54:30] how important the community is. We're [54:32] always going to have rebim and we need [54:34] relationships with our rebay and and [54:36] that's that's the important part. The [54:37] process is as important as the question [54:39] in lived Judaism. That's not what you [54:42] get by talking to an LLM. You get an [54:45] answer and that's good for some things. [54:46] It's great for math. It's great for [54:48] coding. It's great for, you know, [54:50] >> test, [54:50] >> right? Exactly. It might be really [54:51] Right. It might be really good when [54:52] you're taking a rabbitic test. You want [54:53] an answer. That's actually the way to [54:55] pass the test. But the important part is [54:56] the relationship in the process. [54:58] >> Exactly. And the real arena of Judaism [55:01] is Jewish life itself. Meaning Claude [55:03] can't keep Shabas for you. Claude can't [55:06] keep Yun for you. Will a day come where [55:07] you could have Claude look and make sure [55:09] that your esrog is kosher? Probably. [55:12] Yeah, probably. You know, [55:13] >> you could today. Yeah, [55:16] >> you could probably already do that. But [55:18] the lived experience is what Jewish life [55:20] is all about. For me, I actually think [55:21] that like AI is like a clarifier where [55:25] it almost allows you to focus on the [55:27] experience itself rather than like all [55:29] the information they have to carry [55:31] around. And even our relationship to the [55:33] information, it's not about knowing it. [55:36] It's about your involvement engaging [55:37] with it. Like Torah study is such a [55:40] different pursuit than other forms of [55:42] knowledge because everyone would say [55:44] this the very term of Torah lashma [55:47] >> is very almost anti [55:51] the way the rest of the world looks at [55:53] information and content where it's like [55:55] about having the content for most of the [55:57] world and in the Jewish world Torah [56:00] learning lashma which means for its own [56:03] sake it's the joy of learning the same [56:05] reason like [56:06] >> you're not looking for a Right. [56:07] >> Exactly. [56:08] >> I think it can actually enhance that too [56:10] though, right? Like if you if you ask a [56:12] question of an LM a question, [56:15] >> you might end up learning about sources [56:17] that you didn't know about, right? Or [56:18] like and it leads you down paths you [56:20] wouldn't have gone if you had just [56:21] Googled something, right? Which is maybe [56:23] that's like another comparison you can [56:24] make. It's just like looking up [56:25] something online, right? It's probably a [56:27] more rich experience than that in terms [56:29] of what you're going to learn and and [56:30] and be exposed to. And it's also [56:33] conversational, right? You could say [56:34] like, "Well, I want to know a little bit [56:35] more about that angle that you [56:36] mentioned. Let's go down that path." [56:38] Right. That's obviously an improvement [56:39] over just being able to look something [56:41] up online. [56:42] >> Where can people find and follow your [56:44] work? [56:44] >> We're we have our our ex account, haka [56:46] bench. Um we have a website or you can [56:48] email us at hello@ halakabbench.ai. [56:51] Um and yeah, we're open to chatting [56:54] about collaborating or we're definitely [56:56] Yeah, we have some more ideas for things [56:57] we want to do and we'll be posting more. [56:59] >> We need you a better grader. [57:02] We'd like to develop our own repository [57:05] of questions that are private also to [57:07] help make this more of a you know more [57:08] of a legitimate benchmark and um you [57:11] know we we all have our day jobs but we [57:12] want to keep working on this and uh [57:14] figuring out where it's going from here. [57:15] >> Absolutely fantastic and again thank you [57:17] both to Sam and Ben and thank you to [57:20] Brett. We love you. Thank you for making [57:21] >> and shout out Elon Whitesman and Jesse [57:23] Bernstein who are the other two people [57:24] in our in our group who are working on [57:26] this. [57:26] >> We're all from the same neighborhood. [57:28] >> We all live in Scarsdale. Yeah. That's [57:29] really, really, really, really sweet. [57:31] And your incredible Rabbi, Rabbi Morgan [57:33] Stern, who I love, to pieces. I always [57:36] wrap up my interviews with more rapid [57:38] fire questions. Okay. [57:40] >> I always love great book [57:42] recommendations. Is there a book that [57:44] you would recommend that's either a [57:46] great introduction to understanding what [57:49] AI is all about? not not for an expert [57:52] but just a a basic understanding of this [57:55] revolution that's taking place or a [57:57] great introduction to halaka that you've [57:59] enjoyed. You could pick either one. [58:01] >> Think about that. [58:02] >> Yeah, I'd have to think about it. I [58:03] haven't read any AI books. I feel like I [58:05] read about it more often. [58:06] >> You only read technical stuff. You're [58:07] not the you're you're not you're not the [58:09] right. But have you read any Do you have [58:11] favorite Jewish books either of you? [58:13] >> Yes. Off the top of my head, I haven't [58:15] read any Jewish books recently. It's [58:18] been a minute. [58:18] >> It's been a minute. [58:20] >> Yeah. I mean, this isn't this isn't [58:21] necessarily halaka. I I just read uh [58:24] Letter in the Scroll by Jonathan Sachs, [58:26] which I really enjoy. [58:26] >> We love it. I think it's up on the shelf [58:28] somewhere, [58:28] >> right? Which was really about like what [58:30] we were talking about earlier of the [58:31] nature of like Jewish community and what [58:33] it means to be like one member of this [58:35] larger community that's trying to [58:36] accomplish something over time as well [58:38] as just like right now. [58:39] >> I really mean this. The ultimate data [58:41] set of Judaism is not the Talmud and [58:44] it's not Tanakh, it's the Jewish people [58:45] themselves. It's a very [58:48] correct. It's the embodiment of the [58:50] tradition. It's not traditions up here. [58:53] The ultimate tradition is is the lived [58:55] experience of the Jewish people. Uh my [58:57] next question, I'm always curious if [58:58] somebody gave you a great deal of money [59:01] and allowed you to take a sbatical, go [59:03] back to school, wherever you need to [59:04] study any other subject or area of your [59:07] choice, what do you think you'd go back [59:08] and study? [59:09] >> Such a hard question because I really [59:10] love what I studied. I do I really love [59:12] uh mechanical engineering. I think these [59:14] days if I were to go back in school I [59:16] would probably want to study more [59:18] philosophy. I did very technical classes [59:20] in college. I didn't do a lot of [59:21] humanities and didn't have a lot of time [59:23] for like thought process kind of [59:25] classes. So I think that's important [59:26] these days uh to think about the way the [59:29] world works and the way people think [59:30] about things. So that's probably what I [59:31] would take time to do. [59:32] >> Yeah. If I were to go back and and do [59:34] like a you know a PhD I think it would [59:36] probably be something in biology uh [59:38] which is again like different than than [59:40] what I studied in school. you know, [59:41] there's something that the the AI people [59:43] like to say that like these these [59:45] models, unlike computer programs, [59:47] they're uh they're grown, not not [59:49] programmed, right? So, there's some some [59:51] overlap there with like the the way that [59:53] computer science is starting to look [59:55] like biology in some ways. Fascinating. [59:56] And vice versa, where you now have like [59:58] techniques in biology where you can like [60:00] program cells to do things, right? So, [60:02] these things are like converging in an [60:03] interesting way. [60:04] >> Wow. Really fascinating. My final [60:05] question, I'm curious about people's [60:07] sleep schedules. What time do you go to [60:08] sleep at night and what time do you wake [60:10] up in the morning? [60:11] >> Way too late and way too early. I'm [60:13] usually not in bed until midnight, [60:15] sometimes later. Um, the kids wake me up [60:18] at, you know, 6:30 to 7 most days. I try [60:20] to get more. I've got my whoop on and it [60:22] always tells me how bad a job I'm doing. [60:23] But, uh, so every once in a while I'll [60:25] crash early, but uh, yeah, I spend way [60:27] too much time up and and working or [60:29] doing things or [60:30] >> Yeah, I I can't function past like 10 [60:32] p.m. ideally. Uh with with young [60:34] children that's hard sometimes, but like [60:36] I would say like 10 to 6:30 would be [60:38] ideal. [60:38] >> A lot of benches kept Sam up late a [60:40] couple night. I was waiting for one of [60:42] these evaluation runs to finish until [60:44] you know midnight. So [60:46] >> Ben and Sam, thank you so much for [60:48] joining 1840 today. [60:50] >> Thank you for having us. [60:54] >> Ben and Sam, what I love about them is [60:57] they're kind of like your regular [60:59] run-of-the-mill Jews. They live in a [61:01] regular Scarsdale, a regular kind of [61:03] Jewish community. They are living [61:05] extraordinarily [61:07] Jewish lives. And we were talking on the [61:09] steps after the interview. We were [61:12] talking about what's going to really [61:14] preserve Jewish life going forward as AI [61:18] kind of advances more and more. And what [61:20] we both agreed on is the centrality of [61:23] communal Shabas observance and how of [61:27] all of the differences that we have [61:29] communally in the Jewish world, I think [61:31] the difference that will be the most [61:34] essential and fundamental going forward [61:36] is whether Jewish communities are able [61:38] to sustain public shabas observance. [61:41] Now, I'm not asking and talking about [61:44] what we're doing behind closed doors in [61:46] our bathroom. that that's between you [61:48] and your God and maybe your rabbi. But [61:51] in terms of the model of a Jewish [61:53] community that is able to sustain Jewish [61:56] life effectively, I think that the [61:59] denominational framework that's going to [62:01] be in the past. The only framework I [62:04] think going forward that will [62:05] essentially tell a Jewish community, are [62:08] we capable of sustaining a vision and a [62:11] model of Jewish life going forward is [62:13] going to be communal Shabas observance. [62:16] And I say this with both great optimism, [62:18] joy, and not an ounce of judgment. I [62:20] realize not every Jewish community is [62:22] going to be able to sustain that, but [62:24] it's more of an invitation. Think now [62:27] about what it's going to be like growing [62:29] up Jewishly in the world that we have [62:31] now. Think about what childhood is going [62:35] to feel like. If you're a child right [62:36] now, you're 2 years old. What's it going [62:38] to feel like being 7 years old? And I [62:41] believe the vehicle that will preserve [62:43] our humanity more than anything else is [62:46] Shabas observance and specifically on a [62:48] communal level. Shabas is not something [62:51] that you can really do by yourself [62:53] surrounded by those who don't. It's been [62:56] done. It's very very difficult. It's [62:58] very isolating. The beauty of Shabas [63:00] emerges communally. And really as an [63:02] invitation communities should start [63:04] thinking about this. How can we [63:06] encourage federations should start [63:08] thinking about this? We should start [63:10] talking about this much more publicly. [63:12] How do we encourage communities to [63:15] develop an ethic of shabas observance? [63:18] Because I think in the world that we [63:20] have now, what is going to really [63:22] preserve, sustain, and really not just [63:25] sustain, but create the growth and the [63:27] catalyst for the Jewish future that we [63:29] all want. A lot is going to come down to [63:32] Shabas. That weekly communal unwinding [63:35] and disentangling from technology. [63:37] Shabas in a lot of ways has gotten [63:39] harder and harder to observe as [63:41] technology weaves its way more and more [63:43] and encroaches more and more in our [63:45] lives. However, I think the idea of [63:47] shabas has become so much more [63:49] compelling and we should think about [63:50] this. Think about hold your your you [63:53] know look at a three-year-old a toddler [63:55] look at a first grader and ask without [63:58] chabas what is their experience of [64:01] childhood even going to be like and I [64:03] think that of all of the questions that [64:05] AI raises it asked a lot of questions of [64:08] things that are going to change but one [64:09] thing that is going to remain and get [64:12] only stronger in my opinion and this is [64:14] the conversation that we had after the [64:15] interview was the lived communal [64:18] experience of Shabas. So, thank you so [64:20] much for listening to 1840. This [64:23] episode, like so many of our episodes, [64:24] was edited by our incredible team. Thank [64:26] you, Dina. Thank you, Sam. We cannot do [64:29] this work without you. And speaking of [64:31] cannot do this work without you, we [64:32] cannot do this work without our [64:34] supporters. So, please, if you enjoyed [64:37] this episode or any of our episodes, [64:39] please subscribe, rate, review, click [64:41] down, subscribe on YouTube. All of this [64:43] helps us build and reach a wider [64:45] audience to sustain substantive Jewish [64:48] conversation across borders. This is [64:50] what we're trying to achieve and we want [64:52] you and invite you to be a part of it. [64:54] And of course, you can also support us [64:56] the oldfashioned way. And donate at [64:58] 1840.org/donate. [65:00] That's 18fty.org/donate. [65:03] All of your efforts really help us reach [65:05] new listeners and continue putting out [65:06] great Jewish content. And of course, you [65:09] can leave us a voicemail with feedback [65:10] that we may play on a future episode. [65:12] And that episode is coming up very soon. [65:14] And that number is 2125821840. [65:17] Once again, that number is 2125821840. [65:21] If you'd like to learn more about this [65:22] topic or some of the other great ones [65:23] we've covered in the past, be sure to [65:26] check out 1840.org 18fty.org [65:30] where you can also find videos, [65:31] articles, recommended readings, and [65:32] weekly emails. Thank you so much for [65:34] listening and stay curious, my friends. [65:37] Heat. Heat. [66:22] Heat. Heat.