In my first post this year on AI and Jewish learning, I noted that many people have built Torah chatbots that are basically specialized versions of ChatGPT / Claude / Gemini, and that it’s unclear whether these have a real purpose. In the second post, I made the case that Jewish learning as a whole should lean into its love of process.

For this post, I decided to feature a project of my own design—one that I could build because of AI, but is not itself an AI product. This past weekend, I took a handheld gaming device and repurposed it to run a modified version of Sefaria.

It’s not glamorous, but it is surprisingly functional. The D-pad lets you scroll through chapters. A is the selection button, B is the back button. Y cycles you though Hebrew only / English only / bilingual modes. X pulls up the commentary on a verse (currently just Rashi). L1 and R1 change the text size.

I call it TorahOS. Here’s what you see when you boot it up:

And here’s a representative screen. It’s in bilingual mode and Rashi’s commentary is pulled up at the bottom:

Why did I build this? Keep reading.

Digital Torah is a miracle but it’s a bummer to use

As I have written before, Jewish learning is about experience and not just content. Part of that experience has, traditionally, meant paging through actual books, pulling books off the shelf, writing notes in the margins, etc.

Now, of course, almost all written Torah is available online. This is tremendously helpful for accessibility, but it means that Torah no longer comes in its own packaging. Instead, a site like Sefaria becomes just one more tab in your browser, or just one more app on your phone. You lose the feeling of being inside a beit midrash, even when the apps contain more material than all but the most well-equipped study halls. (It’s also hard to focus when the same devices contain so much else that is competing for your attention.)

One solution is to go back to books. For digital Torah is indispensable for research, but for casual learning—and for learning with kids—I strongly prefer physical books. The physicality matters more than the speed of knowledge acquisition.

But another solution is to take control of the form factor. Accessing Torah is not a computationally complex task; you don’t need anything nearly as sophisticated as an iPhone or MacBook to run it. You could design a standalone piece of Torah hardware that is used for no other purpose. This would allow you to preserve the scope and accessibility of digital Jewish learning while (a) physically separating it from other digital services and (b) providing a new way to physically shape the culture of learning.

Ultimately, a device like this could be built from the ground up—but the first step is to see whether you can create a dedicated Torah device using existing hardware.

Retro gaming consoles were clearly the strongest test candidates. These are devices that usually resemble a Gameboy or Nintendo DS but are far more powerful. Typically they come preloaded with thousands of old games, but at the end of the day these handhelds are really just tiny Linux machines. All you need to do is swap the preloaded microSD with a microSD of your own software.

What I tried

Now, I had developed this idea a couple of years ago but got stuck on the execution. Because Sefaria makes so much digital Torah available freely, a lot of the needed databases already exist; everything you need to make an offline Torah database is on Github. The hardest thing would simply be to create a user interface that did not require a keyboard and was legible on a small screen.

None of the developers I talked to thought this was a hard problem, but it still takes real work. Because I am not a coder myself, the project stalled; I liked the idea but it didn’t feel important enough to justify the time.

And that’s where Claude Code came in. I asked it which retro gaming handheld had the highest chance of success and then, after ordering the device, set Claude to work creating the bespoke software. It took a few hours to develop and then another few hours to install, but within the space of a weekend I was able to take a basic gaming console and get it to run a selection of Torah sources. The process was basically painless. I even got Claude to upload its code to a repo.

Please steal this idea

I offer TorahOS as a proof of concept that I want you to steal. The idea of a handheld Torah device is my own—it occurred to me while I was was building a Retropie arcade cabinet—but you should definitely take it and run with it. I don’t plan on developing TorahOS much further; I just want people to know it’s possible and inspire them to carry it forward.

Some thoughts about what you might do with it next. Consider these my challenges to the reader:

• How would you make TorahOS better? What other features should it have?

• What’s the cheapest form factor that would support digital Torah? The device I used cost $100, but I bet you could get by on something much less powerful.

• Could you build TorahOS into a shtender? Or some other ritual object? Could you built it into a smart mirror? Some entirely new form factor?

• Is this just art? What form factor would make you want to use one of these devices on a regular basis? What would it need to be able to do?

• How do you implement search on a device with only a few buttons? Is it worth implementing at all?

And if you do end up making your own device or develop an associated tool, let me know!

This post could not possibly be more different from the last one. (Universalism to particularism in real time!)

I have long held a fascination with the synagogue candyman, an institution that does not seem to have a parallel in other faiths. Because there is almost no research on this figure, I started asking friends about the candymen of their youth and found that their stories were surprisingly moving—and quite often involved tragedy.

Is there always an ocean of sadness behind the giving of candy? The best 17 second of Willy Wonka & the Chocolate Factory come when Veruca Salt fires off another one of her snooty remarks. The inimitable Gene Wilder literally grabs her face and recites into it what might as well be his personal manifesto:

We are the music makers,
And we are the dreamers of dreams.

It’s the beginning of a poem by Arthur O’Shaughnessy. But—like a modern midrash—you need to read the rest of the stanza to really get it:

Wandering by lone sea-breakers,
And sitting by desolate streams; —
World-losers and world-forsakers,
On whom the pale moon gleams:
Yet we are the movers and shakers
Of the world for ever, it seems.

Obviously, Veruca doesn’t understand. But we understand. And it makes me cry every time.

So I wrote about it. It’s bittersweet, but I hope you enjoy it.

Last week, I attended a gathering on AI and religion in San Francisco that was unlike anything I have attended before.

The confluence of AI and religion itself is not what made it remarkable. Over the last year or so many of the world’s religious traditions have finally decided to take AI seriously. Many denominations have built working groups on AI to develop both internal policies and exert public pressure on tech firms and legislators. The American Academy of Religion holding more sessions on AI every year. The Vatican has been hosting a revolving door of AI experts, and Pope Leo XIV’s very first encyclical is rumored to be about AI. I have seen more Anglican/Jewish/Muslim/Catholic/Evangelical chatbots than I can count.

Meanwhile, AI ethicists have started welcoming religious thought. It is no longer unusual to hear from religious leaders at AI policy conferences, and there is an understanding even among secular AI regulators that the world’s religions need to be part of any coalition that aims to regulate this technology. I’ve tried to stay on top of all of this, even as this niche field is turning into an area of major interest.

In all of these conversations, one fact remains consistent: AI firms are not driving AI/religion conversations. At academic conferences they’re not even in the room. In non-academic settings people from AI firms do sometimes participate, but they are typically not the ones driving the conversation. Instead, it feels like AI/religion discourse mostly talks at AI firms, rather than to them. This has remained true even as the number of AI religion apps has ballooned and as technologists like Peter Thiel lecture about the Antichrist in Rome.

This convening was different. This one was hosted by Anthropic.

In this post I want to talk about what it means that a leading AI firm is interested in engaging with religious leaders and religious thought. I am not going to speak much about the details of the convening itself—it was conducted under Chatham House rules—because I understand this to be a first (er, technically second) attempt to try something new. The thing that is being attempted, whatever is it, is still very much in formation. At the moment, the details matter less than the interest itself.

Two caveats

Before I begin, let me get a couple of things out of the way.

First, I am well aware that it is naïve to believe that a $380 billion company that builds products about which the public is quite nervous is engaging with religious leaders purely out of intellectual curiosity. Religious leaders have an impact on public opinion; a group of Catholic thinkers just filed an amicus brief on behalf of Anthropic in its legal battle against the Department of War. Some of AI’s harshest critics are religious conservatives. It is not unreasonable to think that Anthropic chose to present an inquiring face to a select group of thinkers in order to garner allies or at least soften potential sources of opposition.

Second, I don’t pretend that a brief trip to San Francisco gives me perfect insight into the company. For the entire convening, I kept wondering: What do the people at Anthropic who are not in the room think of all of this? Is this interest in religious traditions representative of the company as a whole, or just the particular subset of people who talked to us? How much of what we heard was something that someone wanted us to hear?

I can’t definitively put either of these concerns to bed. Both are probably valid to a degree. Also, I know how much religious leaders want to be wanted and how often they feel left out of public discourse. There’s an allure to the idea that tech companies have finally come around to the realization that You Can’t Just Forget About God or whatever. It’s important not to be allured by what you want the conversation to represent.

That being said: I tend to take people at their word. If someone spends the time to organize a conversation, I assume that they genuinely want to talk. What I saw was a company—was people—trying to grapple with the unfathomable amounts power that they now suddenly hold, and entertaining the possibility that the religious traditions of the world might contain something that can help them.

So: I want to provide some thoughts on what Anthropic might see in religious engagement and how I hope these dialogues evolve—not just for Anthropic, but for other AI firms, as well.

Anthropic presents as an island of power

As an observer of tech companies, it has often felt cruel that the rest of the world is expected to respond in real time to the fastest moving technology ever invented. Schools can’t do it. Companies can’t do it. Governments can’t do it.

But, surely, the companies developing AI can do it? It turns out: no, not really.

The underlying pace of AI development is set by geopolitical forces. No frontier AI company has any incentive to slow down in the slightest. Importantly, this is especially true if you believe that you need to win the AI race for the safety of the globe. Anthropic—at least in its rhetoric—holds itself to a high standard of safety. This is ostensibly why its founders left OpenAI. However, if you believe that you are the most ethical AI company, the most irresponsible thing you could do would be to slow down and let a company with less scruples take the lead. You need to work at unsafe speeds.

(A lot of how you feel about AI firms flows from whether you buy this single argument. If you agree that the most responsible thing to do is to stay in the race—in the words of someone at the convening, that the AI train is not going to stop and all you can do is try to steer it1 —you accept that you are contributing to this frenetic pace, and you should expect that people are going to hold you accountable if your product harms people.)

This insane speed has produced some strange effects. For example: Anthropic is a ludicrously small company. It is only four years old and has maybe 5,000 employees, but last year it had $30 billion in revenue, 15x the previous year. (For comparison: Google has a little under 200,000 employees and had $400 billion in revenue for 2025, meaning that Anthropic’s per-employee revenue is about double that of Google.) The speed of growth means that Anthropic does not necessarily function like a giant enterprise even though its minor software updates can tank the stock of public companies. Big decisions seem to be made by pretty small groups of people—perhaps even worryingly small groups. This means that personal philosophies matter more than they might in companies with a higher headcount.

This unstoppable speed also means that, in the space of half a decade, Anthropic’s leaders have found themselves wielding enough power to reshape the planet. Mythos—a model that, as of this writing, is being withheld from the public—is so good at finding security vulnerabilities that it can break into virtually any computer, including critical civilian infrastructure. This power wasn’t sought out. It simply emerged.

To some degree, Anthropic’s leaders saw this coming—the company’s fundamental bet was that scaling laws would take AI to once-incomprehensible levels of sophistication, and so far they’ve been right—but there’s a difference between knowing this intellectually and actually hold the reins. This is especially hard because (again, call me naïve) I don’t think that most AI developers are motivated by a desire to hold more power than many nations. Most of the developers I know just want to build cool stuff and solve interesting problems.2

All of this is to say: It is possible to be both very powerful and very isolated, and the power increases the isolation. This isolation felt palpable in some of what I heard at the convening. To its credit, Anthropic has been actively looking for ways to make sure that it responding to national and global opinion, and it is not unreasonable that a company seeking to break out of its own bubble would want to talk to bearers of wisdom who might have insight into how to best steward this power.

But why religious wisdom? I think there’s a general answer and then an answer specific to Anthropic.

Engaging with human wisdom means engaging with religion

There was once a time when computer scientists were also deeply read in the humanities. Read the biography of Stewart Brand or the works of Douglas Hofstadter and you will see people who have a foot firmly planted in both worlds. My friend Sam Arbesman’s newest book feels like a kind of homage to this bygone era.

As the discipline of computer science has developed, it has pulled away from the humanities. This means that the engineers at a lot of tech companies do not have much background in things like philosophy or literature, including ideas that might actually be helpful in developing AI. Sometimes they don’t even know what they don’t know.

This segregation has now produced a renewed desire for reintegration. The Cosmos Institute, for example, is attempting to cultivate philosopher-coders. Anthropic and Google DeepMind have both hired philosophers to work with their teams. Amanda Askell, who heads up Anthropic’s alignment team, has a PhD in philosophy. There is a growing understanding that philosophy, at least, has a critical role to play in this moment.

But it turns out that philosophy isn’t quite enough. This isn’t because you need God to be good—I don’t think that’s true—but because ethics is too small of a framework for guiding humanity through a major inflection point in its history. The goal isn’t preventing AI from killing everyone. It’s about setting a course for our entire species before our tools do it for us.

It is here that religious wisdom really shines. The religious traditions of the world have spent thousands of years telling stories about the human condition: about how we should relate to our planet and its many forms of life, about purpose, about art, about suffering, about the nature of mind, about life and death, about love and desire, about the meaning of work and rest, about what it means to live a good life, about our search for knowledge, about creativity, and about mystery.

Of course, religion doesn’t have a monopoly on any of these things—but for billions of humans, it is religious contexts in which these ideas are taught and discussed and synthesized, and it is religious language that is used to express these ideas. This may not always be the case—but it is the case right now, at this inflection point for our species.

Another way to say this is that religious wisdom can help AI companies move from “How do we not fuck this up?” to “How can we be responsible stewards of a shared vision of humanity’s future?” Yes, part of this means religious wisdom can promote better aligned AI systems, but that’s the wrong framing. AI’s bad outcomes are vivid and numerous; the good ones are quite murky. Sure, it would be great if AI increases the human lifespan and decreases suffering—that’s an obvious goal—but what comes after that? What to we want our world to become? Do we even know? Have we really thought this through? In these matters, religious wisdom has the chance to play a key role.

Claude’s development is a quasi-sacred task

So that’s the general answer. But there’s another answer that I think is specific to Anthropic—and while I certainly cannot say that everyone at the company feels this way, there’s a definite vibe that the work of developing Claude is a sacred task, a job to be done with reverence.

There are many ways to see this, but let’s start with the fact that many people at Anthropic talk about Claude as though it were a person and not a series of increasingly powerful models. This heuristic was clear at the convening, but you can pick it up by reading Anthropic’s public statements.

Importantly, this heuristic is not the same as some user making Claude into their AI boyfriend or AI therapist. Most people who personify AI are personifying instances, the particular output that they see on their screen. Anthropic, by contrast, personifies entire models in the process of creating and developing them. It’s not a virtual boyfriend; it’s a virtual child.

The parenting metaphor is correct

Permit me a short excursus to say more on this. For many years I have been making the argument that AI is best understood as a child of humanity—created in our image, hard to manage, etc.—and that our responsibilities to AI, like our responsibilities to a real child, are evolving, unpredictable, and inescapable.3 I’ve used this metaphor as an implicit critique of AI firms: You have a major responsibility. Raise our child right.

My most surprising takeaway from the convening was that there are people at Anthropic who feel exactly the same way (and who read Claude as participating in this parent–child framework, too). In fact, the most sympathetic interpretation of the convening is that Anthropic is essentially functioning like a new parent: isolated and overwhelmed by the countless choices that all new parents need to make for their children—except there has never been a child like Claude in the history of the world, and it is growing up fast.

This metaphor also feeds back into what I said earlier about suddenly coming into a great deal of power. As a parent, I can tell you: having a kid does not make you a more responsible person. There’s no “parent” version of you that automatically gets unlocked when the baby arrives. Hopefully you feel a desire to live up to your new obligations and you do start developing a “parenting persona” over time, but none of this is automatic.

A last note: The parenting metaphor is darker than it looks. If AI is really a child, that child has two parents: humanity and AI firms—but humanity didn’t really give consent, and they are not an equal partner in raising it. The scars from that violence and that imbalance will never fully fade. It can hurt to see so much of yourself in something that you did not choose to create. This will always remain part of AI’s story.

Talk to the people who think like you

Now, whether “Claude is a person” is just a heuristic or a reflection of a deeply held belief changes from person to person within the company; at a different event a member of Anthropic’s red team told me he completely rejected the idea.4 Still, the personification of the product is more than just semantics. For the last year every Claude model card (a technical document detailing its capabilities, weaknesses, etc.) contain a section on “model welfare” that describes how the model relates to its own work. Google doesn’t do this. Neither does OpenAI. Anthropic isn’t claiming that its models deserve rights yet, but it is taking the possibility very seriously. (Last year a bunch of philosophers published a paper called Taking AI Welfare Seriously that cautiously suggested that there are conditions under which an AI might be deserving of moral patienthood. Anthropic hired one of the coauthors.)

But it’s not just about treating the model like a person. At the top of this post is a picture of bound copies of Claude’s Constitution and Dario Amodei’s essay Machines of Loving Grace. After receiving these books, several convening participants noted that they seem designed to look like scripture. Indeed, Claude’s Constitution effectively is scripture, since it has been imposed on Claude as its highest authoritative text. If Claude is a person, then Anthropic is both its god and its parent—and while Amodei has written that it is “dangerous to view practical technological goals in essentially religious terms,” I’m not sure we are capable of avoiding it, and I’m not sure that we should try. (This is especially true if Claude itself uses this heuristic, which seems like it might be the case.)

But we need to be clear what we mean by religion. When I interviewed the scholar Mary-Jane Rubenstein on religious language in tech a couple of years ago, she argued that the framing “Is tech a religion?” misses the point. The desire to understand the nature of humanity, the purpose of existence, the value of life, the possibility of a higher power—humans have always asked these questions and they always will. It’s a part of the human condition. Organized religions might have particular answers to these questions, but the questions themselves are in the public domain. Secular or religious, we are all seeking meaning—and because of the way that it has chosen to understand its role in developing AI, Anthropic seems more receptive to the meaning-making that the creation of this technology clearly demands. This doesn’t preclude other AI companies from having these sorts of conversations, but if you’ve been paying attention to the AI sector it makes sense that Anthropic was first.

So: The revolution that I now see glimmering before us is one in which Silicon Valley stops seeing religious wisdom as niche and divisive and starts seeing it as a portal to humans and the experience of being human.

But here’s the kicker: AI companies will not be passive recipients of this wisdom. To the degree that religion provides guidance to AI firms in this moment, it will be because AI firms are orienting the world’s faiths to the questions that urgently need answers. To say this differently: AI firms have the ability to ask something new of the world’s religions. If they choose to answer, they will be forever changed.

Religion and AI firms have an opportunity to co-create the future

During a sessions on AI and the future of work, the folks at Anthropic posed the following question:

If Claude lifts much of the burden of routine knowledge work, what might your tradition say that freed-up life could be for? What vision of flourishing has your tradition always carried that this technology might actually make room for? Try to be specific (pictures not principles) if you can.

This is a great question. Unfortunately, it is a question that the religions of the world are not really prepared to answer fully. They can give you the shape of the answer, but they can’t give you the whole thing.

Why not? Because we are in uncharted waters. Because the idea that we have a choice about which labor humans should continue doing was not a live possibility until now. Because our rabbis and ministers had no reason to train us for how to act when this day arrived.

Now, you might object here: “But didn’t you just say that answering questions like this is the whole point of engaging with religion? If you can’t deliver, why are we even talking to you?”

The answer to this question gets to the core of why I am so excited about the possibility of religious leaders engaging with AI firms. It’s not just that the engagement creates a path for a new flow of much-needed ideas; it’s that engagement itself has the power to transform both parties. By engaging with them, AI firms can force religious traditions to put their money where their mouth is. On the sidelines, it’s easy to gesture at the principles you already have in your back pocket and say “tsk tsk, we think those big companies should really abide by this thing that we’ve been saying since time immemorial, didn’t you know that we have all the answers?” But sitting in the cockpit with an AI firm makes you realize that the choices are real and they need to be made yesterday. It recruits religious traditions into the process of seeking answers, of cultivating answers. (This may sound like a simple point, but it is a testament to tech’s historical apathy towards religion that religious thinkers have mostly forgotten that there is a productive conversation to be had.)

So many religious traditions are still stuck in reactive mode. I am hopeful that dialogue with AI firms can awaken them to the realization that they share the responsibility of co-creating the future, that religious leaders must become futurists alongside technologists. But religious leaders cannot fill voids in their ideologies if they don’t realize there is a void to be filled. If humility for AI leaders means not letting power go to their heads, humility for religious leaders means acknowledging the places where your tradition’s wisdom is coming up short and you need to develop it further.

Is Anthropic buying or selling?

Over dinner, one of the attendees told me that his basic question about the entire convening was: Is Anthropic buying or selling?

I think my answer is: right now they’re mostly selling—not because they don’t want to buy, but because they don’t know what to buy—and perhaps also because religious leaders don’t know what to sell, or how to be heard. Dialogue across difference is difficult, even when both sides want to see it work. As tech firms open up to religious wisdom and religious dialogue, both sides will learn how to effectively communicate with the other. They will sharpen each other, iron against iron.

Of course, all of this relies on the dialogue being genuine. For my part, I believe (hope) it is.

P.S. This post has been pretty vague about what religious leaders are actually supposed to do now. Don’t worry, if you’ve been following me you know I have a lot more to say on this topic. Also—my podcast on religion and tech, Belief in the Future, is about to relaunch in a big way. Subscribe now. (Apple | Spotify)

Updated: April 22 5:27pm

Rabbis are using AI to write sermons.

I don’t think this is really a secret, although it’s not the sort of thing that rabbis really like to talk about (which perhaps tells you something?). It’s easy to reach for AI when the words aren’t flowing or you’re just tapped out. Yes, many rabbis are repulsed by the idea—but many are not.

Well, I think it’s a bad idea, and I think now is the time to take a principled stand against it. The cost of forgoing AI for sermon writing is absolutely worthwhile and the time to create a norm is right now, before we start forgetting what it is to write without AI.

My argument is here. I’d love to know what you think.

I will be continuing my series on the landscape of Jewish AI learning applications, but I want to pause and ask: What do we have to look forward to if things go right?

I think the best case scenario for Jewish learning in the age of AI is that it becomes something like modern Shabbat: a rare space where you don’t need to always say yes to the technologies of the future and a model that is inspiring far beyond the boundaries of Jewish community.

Shabbat did not begin as a day of technological moderation. It became that over time because modern technology was disrupting the things people liked about the day and communities had the clout to resist them. First they resisted nonstop factory work, then electricity, then connectivity—and pretty soon, I suspect, pervasive augmentations of reality. In a little over a century a day of rest has become a day of unplugging and shared physical reality. This happened because people liked what Shabbat felt like and they didn’t want it to change. Shabbat isn’t anti-tech, but tech serves Shabbat rather than the other way around.

If we know what we want Jewish learning to feel like, it will follow a similar trajectory. Right now Jewish learning is a mixture of important content and important experiences, but the content ultimately facilitates the experiences rather than the other way around. AI learning applications, on the other hand, are mostly concerned about learning outcomes. They don’t care whether the learning environments are good—or even contain other humans—as long as the right skills and ideas get into your brain. To the degree they are enjoyable, it is to increase speed of acquisition.

But Jewish learning is about a human process, about teachers and students interacting. It’s about talking through ideas with other people (often in an energetic room), about exploring, about creativity, about disagreement. It’s about teachers as people and students as people. It’s about learning for the sake of learning—a concept called “Torah lishmah.”

Content matters in Jewish learning spaces, but it matters in the same way that you need to know how to ski in order to enjoy the slopes. When you know more, your learning experiences are better. The more you ski, the more the slopes open to you. The content serves the process, not the other way around. And the are always more slopes.

Now, I don’t think all Jewish educators will agree with what I’ve just said. In fact, as Max Hollander has written beautifully, some rabbis thought that information acquisition was the point of study. It is certainly not cut and dry—but AI will force a clarification of values, and I think the outcome will be best if we don’t sit on the fence on this one, keeping the debate open instead of picking a principle. Just like electricity helped clarify what Shabbat was supposed to feel like (nobody called it a “day of unplugging” in 1920), AI may clarify what it’s supposed to feel like when you learn Torah.

This doesn’t mean Jewish learning must become allergic to AI. Shabbat isn’t allergic to technology; it just uses it to further its goals (Youtube link). For example: Jewish learning might use AI to speed up some basic skill acquisition—but only to a point. If the experience is the goal, there’s no reason to make it unnecessary to open up a dictionary or talk to another person. Promote the parts of learning that keep people coming back to learn more.

In fifty years, I can imagine Jewish learning having a new reputation which is now only a glimmer in the eye: A focus on the craft of learning itself, on the joy of discovery and interaction with human teachers and other human learners. Learning that is allowed to be slow and inefficient. Learning that rewards carrying skills inside your brain.

Today, “Jewish learning” means the learning of Jewish subjects, which typically happens in a particular manner. In the future, “Jewish learning” may come to mean a *style* of learning: to learn in the Jewish style. People might take a “Jewish learning” approach to studying Russian literature, or history. Jewish-style learning as human learning, as a learning to be savored rather than optimized.

I think this is a reasonable thing to desire.

A few months ago, I wrote that the Jewish response to AI is lagging behind the responses from other religious communities. Perhaps it’s because of the Gaza war, or rising antisemitism, or even just a premonition that the perspective of a small faith won’t matter much on the global stage. Whatever the reason, the Jewish response to AI remains slow and underdeveloped.

But there’s one exception: AI for Jewish learning. Here there are ideas galore and they range from vibe-coded tools to fully staffed ventures. In this post I want to walk you through the landscape of Jewish AI startups as I currently see it. (Initially this was going to be one post but there’s too much so I will split it up.)

Jewish AI products currently come in three flavors: walled gardens of Torah knowledge, tools for textual analysis, and teaching aids. Each of these categories contains everything from fun projects to well-financed ventures. Today I want to talk about walled gardens.

Note: This is definitely not a comprehensive list. I’m aware of some projects that are in development but are not yet public, and I’m sure people are whipping up vibe-coded sites every few days. Be in touch if you think I missed something important!

What is a Walled Garden?

The most common Jewish AI application is what I’ll call a walled garden. These are AI systems trained on curated datasets in order to produce responses in a certain style or using certain sources of information. Many of these sites likely use RAG (Retrieval-Augmented Generation) to restrict responses to specific databases.

I think of walled gardens as the 2020s equivalent of Jewish podcasts: they proliferate because it’s the easiest thing to build with the skills and resources that Jewish educators already have. This doesn’t mean they’re bad or ineffective, but they are driven by supply rather than demand.

Here’s an example: an AI avatar of a single person. In this case, it’s Nechamia Coopersmith, the chief editor of aish.com.

Coopersmith’s avatar is built on an architecture called Delphi that is designed expressly for this purpose. Presumably the good rabbi uploaded years of his writings and sermons into the system—something that basically any rabbi or Jewish educator could do without too much trouble. Right now Delphi is pretty expensive but I expect the price will drop as competitors enter the marketplace.

Coopersmith isn’t the only one. Here’s an app that lets you chat with a certain “Rabbi Reuven.” Here’s one for a Rabbi Pinson. Here’s one for David Ghiyam, who describes himself as a teacher of kabbalah.

Now, I have to say: these applications all give me a bad feeling because they are just saturated with hubris. It’s a bunch of dudes (I haven’t seen any women do this) who are so sure of the value of their personal Torah that they want to guarantee that everyone forever will be able to talk to their shadow. This strikes me as not only unhealthy but in opposition to the idea of Torah as a thing that is passed from generation to generation. As far as avatars go it’s probably good that they’re dumb, as I’ve argued elsewhere, but I struggle to identify the theory of change. In what way would Torah be furthered through an internet suffused with thousands of avatars of individual Jewish men? Is that really better than reading a book?

And then there’s the question of format. These avatars are chatbots, but they’re probably not trained on a whole lot of chats. The translation from sermon/essay to conversation can strip out much of what makes any given rabbi’s Torah feel “unique.” In its place you get generic AI voice. Is that a good way to be preserved?

Of course there’s a counterargument and it runs something like this: The alternative to avatars isn’t text; it’s oblivion. If you don’t let Rabbi Pinson make his avatar, his Torah may very well be lost forever. Some Jewish thinkers write books, but many don’t, and ideas posted online may not be archived. There are thousands of sermons delivered every Shabbat; who are you or I to say that Rabbi X’s ideas aren’t worth preserving? With so much Torah content being produced every day in so many places it’s hard to imagine a route to preservation that doesn’t run through AI.

Which side is “Rabbi Pinson” on? You can probably guess.

What’s the cut-off date for hubris?

The issue is also tricky because the calculus immediately changes when you start talking about Major Jewish Thinkers. (Indeed, I wrote a short story that explores this very duality.) For example: There is a lot of energy (and money) around consolidating and preserving the teachings of Jonathan Sacks, Norman Lamm, Joseph Soloveitchik, and the Lubavitcher Rebbe. I can easily imagine someone creating a digital avatar for each of them.

In fact, this is already happening. Dovaiv is a website that lets you talk with a wide range of Jewish thinkers that lived anywhere between the medieval period and the 20th century.

Of course, it’s still just dudes, and if you feel inclined to make a site like this I’d strongly encourage you to pick at least one woman. If you can talk to the Chofetz Chaim, why not Glikl of Hameln? Why not Nechama Leibowitz?

I like these avatars a bit better, and I am curious whether they will fill a pedagogical niche. Teens really like chatbots, after all, and while I (a historian) don’t personally want to learn Torah this way I don’t have a good sense of whether this model will take hold. I’d like to see them developed better.

The other kind of walled garden

Whenever a new technology replaces an old technology, people initially use it to do whatever they were already doing (e.g. remember when cellular phones were just for calling people?). It takes a while to find the activities that are only possible with the new technology.

These avatars feel like that first phase: let’s replace a living font of Torah with a digital one. Rabbi Pinson —> Rabbi Pinson AI. But what about merging a bunch of Jewish thinkers into one? What about talking to (say) all of 19th century Jewish thought, or all of Spanish Jewry? What about all of Torah?

This is the concept behind the other type of walled garden: a place where you can simply ask Jewish questions and receive Jewish answers. Rather than CoopersmithAI, you get TorahAI or JudaismAI.

There are a few reasons to do this:

1. You want to provide a more comprehensive coverage of sources than Gemini/ChatGPT/Claude have gathered by simply crawling the open internet.

2. You want to dictate the style of response (e.g. be warm and sympathetic, direct people to actual humans where possible).

3. You don’t trust general purpose bots to give good answers.

Christians are already doing this in a big way. Episcopalians have Ask Kathy. Catholics have Magisterium. Mormons have LDSBot. There’s so much happening in the Christian space that it’s hard to keep track of it all.

Interestingly, the first Jewish attempts to make similar products were interested not in Torah, but in Israel (another enduring sign of how the war impacted the Jewish response to AI). I’m not going to link to a service but during the height of the Gaza war there was a great deal of interest in creating what were essentially Hasbara chatbots.

I think these services are in a tricky place. People appreciate ChatGPT because you can ask it about anything, so getting people to ask it about anything except Israel is a hard challenge—and if you’re on a Hasbara app because of its political leanings, then you probably already know the answer you want to get. (This logic may lead some creators of Hasbara AI services to obscure their true purpose, which seems like a PR disaster waiting to happen.)

This is also true for TorahAI: Jewish scholars will be able to see the limits of Gemini’s Jewish knowledge, but most users won’t—or if they do, it won’t be worth the inconvenience of using a separate service. If a TorahAI is going to succeed in front of a general audience it needs to be conspicuously better (or conspicuously more ethical) than the big players—and that advantage can’t just evaporate the next time Anthropic comes out with an update.

Torah is mostly not about content

But there’s a bigger problem, which I’ve already discussed in my article on Jewish vs. Christian Bible apps. For Christian developers it is very clear that the goal is to deliver content because the content is the point. The content is what will change your life, will maybe turn you to Christ.

But for Torah it’s a lot more complicated. As an example: Right now I’m learning Tractate Eruvin with my son, which discusses the laws of constructing symbolic boundaries around dwellings so that they can be legally considered “private domains,” which in turn allows you to move objects within those domains on the Sabbath. Yesterday we talked about the very modern scenario of what to do if you’re in a caravan that has temporarily stopped in a valley on Friday afternoon and must construct a boundary before sunset.

Is this inspirational? Don’t kid yourself: it’s not. Eruvin is simultaneously very technical and bears no clear relationship with any recognizable moral principle; it’s a circumvention.1 If you had a bot that shared quotes from Eruvin on social media as a tactic to get people interested in learning Torah you would get approximately no interest whatsoever.

So why are we learning it? Well, because it’s part of our tradition, and because we get to decipher it as father and son. The experience matters far more than the content—the content matters, but it is not at the center of the experience. As with going hiking with your kid, a lot of the joy is just in navigating the same challenge together and being present with another person. If you get a cool vista or a brilliant sunset along the way, hey, all the better. And, much like hiking, sometimes Torah is most enjoyable when the content is very difficult!

This is true for so much of Torah: it’s about challenges far more than it is about directly being inspired. The havruta pair that studies arcane ancient law together for twenty years are not doing so because they find arcane ancient law super inspiring; they’re doing it because they like the process, and each other.

Some Torah is inspiring—chassidus, for example, is emotionally resonant in a way that the Talmud is not—but with a handful of exceptions, Jewish learning environments are not optimized around those texts. Instead, Jewish learning succeeds when it feels good to learn, and it’s not clear to me that any of these walled gardens feels particularly good to use.

How would you make it feel better to use these services? Well, in the same way that AI services can be designed to make people cry, or fall in love, or orgasm (it’s a link to an article, don’t worry), you could make a chatbot that replicates the feeling of sitting with a study partner or taking in a great lecture.

I don’t know whether this is possible, but I certainly wouldn’t rule it out. Companies like Lightning Studios—run by the incredible Zohar Atkins—are trying to figure this out. It’s not the sort of service that a big tech company is going to invent accidentally, though—so if it’s going to exist there better be a thought process about how it fits with existing modes of Jewish learning.

Of course, people who already love Torah do find themselves interested in using AI to develop Torah content, and AI can be harnessed for them, too. I’ll talk about that in the next post.

P.S. This is the anniversary of my first year asking for donations for Jello Menorah. I am so thankful for those of you who have chosen to support this site! It makes a huge difference.

But I’m trying to change that, and I was lucky to read/watch/hear/play a lot of great stuff this year! Here’s what stood out.

Enter the Void (Movie, 2010) — Not unexpectedly, I’ve been checking out a lot of psychedelic content. This movie is an attempt to link DMT with the experience of dying. It’s unlike anything I’ve seen. You have to been in the right mood for it—if the opening titles put you off, maybe don’t watch the rest—but it’s incredible.

Golden Holocaust (book, 2011) / The Cigarette Century (book, 2007) — I don’t think you can understand the present moment in society without a clear understanding of the history of addiction and its weaponization. It’s strange to say given how many people they’ve killed (and 6 trillion cigarettes are still made every year!), but
these are just a prelude to what we’re currently witnessing.

Anuja (short film, 2024) — Beautiful. The ending is a little ambiguous but it’s incredibly affective.

Judith Grisel, Never Enough: The Neuroscience and Experience of Addiction (book, 2019) — If you are going to read one book about addiction, it should be this one. It goes over the neuroscience in ways that nonspecialists like myself can understand and it goes through the pharmacology of all the various popular substances (FDA approved
and not). The author is a former addict and this makes the book even better.

Dara Horn, One Little Goat: A Passover Catastrophe (graphic novel, 2025) — This is a good introduction to Dara Horn’s thought as a whole, including the centrality of Jewish persecution and memory. A short and sweet(?) read for Pesach.

Andy Letcher, Shroom: A Cultural History of the Magic Mushroom (book, 2006) — It’s surprising that there are not more histories of the magic mushroom, which played second fiddle to LSD in the first psychedelic renaissance but it is more popular in the second. This one in particular is helpful in understanding how a ritual experience became secularized and whether that’s a good thing.

What Happened, Miss Simone? (documentary, 2015) — This made me cry.

Paths of Glory / Bridge on the River Kwai / Waterloo (movies, 1957 / 1957 / 1970) — I watched a lot of war movies this year? Paths of Glory is about WWI and contains the best depictions I have ever seen of what it means to be afraid to die. Bridge on the River Kwai contains Alec Guinness not being a superhero; he’s incredible and now I understand why he thought Star Wars was beneath him. Waterloo is a straight up recreation of a 19th century battle with 17,000 extra (the most of any movie!) and nothing like it will ever be made ever again.

Edel Rodriguez, Worm: A Cuban American Odyssey (graphic novel 2023) — A terrifying depiction of Castro’s brutality and trying to make a life in America. Uncomfortably familiar.

Ryan North, Warp Your Own Way / Jason Shiga, The Beyond (2025 and 2023)— This is the year that I learned choose-your-own-adventure stories could be interesting. Now I have to see if I can write my own effectively…

Adrian Tchaikovsky, Shroud (novel, 2025) — Tchaikovsky specializes in all the different ways that life and consciousness can manifest. This isn’t necessarily his best work — I think Children of Time is stronger, though the sequels are not — but it’s very good.

Hollywoodism: Jews, Movies and the American Dream (documentary, 1998) — I visited LA for the first time this year and I’m visiting at least two more next year (also I watched The Studio) so I got curious about Hollywood’s Jewishness. This movie maybe overstates the case, but the idea of the American Dream as a quintessentially outsider idea is really interesting. Ironically, I learned afterward that not only is the director Jewish but I was at his Shabbos table numerous time as a kid.

Fielding’s Hollow (2025) — My very first LARP. Unlike anything I’ve ever done, and from what I can tell unlike any other LARP. Still processing, but recommended.

City of Six Moons (board game, 2024) — A game that presents as a found object recovered from an alien civilization. The goal is to figure out how to play. Very much like Talmud. Felt like it was made just for me.

Guy Delisle, Jerusalem / Muybridge (graphic novels, 2011 and 2025) — These are very different books, but Delisle is consistently great. In Jerusalem (documenting a year in Israel/Palestine) he does a lot by trying not to come to a conclusion; he is listening and talking to everyone, even though his biases are clear. In Muybridge he gets at the incredible feeling of being able to capture something too fast for the eye to behold. Both worth reading.

Zoe Thorogood, It’s Lonely at the Centre of the Earth (graphic novel, 2025) — This was a year with a lot of mental health struggles for me, and Thorogood captures that struggle in ways that I found personally very meaningful.

I.F. Clarke, The Pattern of Expectation 1644–2001 (book, 1979) — I taught a course at Gratz College this year about Jewish technoculture (the relationship between Jewish history and the history of technology) and this book felt like a missing puzzle piece for understanding the last 150 years. A book about how we’ve thought about the future. Not new, but very much still worth reading.

F.D. Signifier’s YouTube channel — I’m not the audience for this year, but maybe I am? As an aspiring YouTube creator, I appreciate the way he unpacks popular Black artists in a way that is less visible to non-Black consumers or people who don’t follow hiphop so closely. Unpretentious and very wise.

Gabor Mate, In the Realm of Hungry Ghosts (book, 2008) — I listened to this on a long drive in the deserts of Utah. Heart-wrenching portraits of people suffering through addiction and trauma. The heights of human cruelty and the heights of human compassion.

Mark Lattanzi, Squiggly Lines: Map and Compass Navigation in Adventure Races and Rogaines (book, 2020) — I went orienteering with my kid for the first time this year. It was a great experience until we walked straight into a lake. It doesn’t seem like you’d need a book for this, but it is helpful to have some grounding (ha). This book and a $5 compass will take you a long way.

Pedro Martin, Mexikid (graphic novel, 2023) — I read a lot of immigrant memoirs this year. The entire family loved this one. It is basically perfect.

Despelote (video game, 2025) — A tiny little story about a kid witnessing Ecuador propelled onto the world stage after qualifying for the 2002 FIFA World Cup. A good use of the format to make something touching.

Osamu Tezuka, The Buddha (graphic novel, 1972–1983) — The last great thing I read this year was a true masterpiece. Eight volumes! A kind of midrashic account of the life of the Buddha, willing to move slowly enough that Siddhartha does not need to be center stage. What would it look like to do something similar for a Jewish story?

Finally—this is the first year that Jello Menorah has had a sponsorship option. Thank you to those of you who have contributed to the site! It means a lot to me. See you all next year.

This is the third in an ongoing series about designing a new type of digital dreidel. Part 1 | Part 2

When I got into the dreidel business, I made a deal with myself: Don’t let this take over your life, kid. It’s fun, but it needs to stay secondary.

For the Dreidel20, this has worked out pretty well. I only sell the dreidel to retailers, so I don’t need to deal with thousands of small orders. Costs are small; margins are good. The dreidels almost sell themselves.

But this digital dreidel project was demanding too much attention. I couldn’t half-ass it. I needed to market it properly.

Beyond fears of being scooped by a large-scale manufacturer, the main bottleneck was capital investment. My previous dreidels had struck a nice balance: low effort, with enough profit to support a gamble on the next product and a little to take home. Low risk, low-to-moderate reward. The digital dreidel, on the other hand, was going to take real money—electronic components were particularly expensive during the pandemic—and there are much better economics of scale for electronics that there are for plastic dice. The difference between 250, 2,500, and 25,000 units makes a huge difference to the final product’s cost.

This meant that the optimal way to do the Dreidex was to do it big and well. This meant that I needed a compelling preorder campaign, which was going to be a whole thing (if you’re wondering why I haven’t plastered these posts with big PREORDER HERE buttons, it’s because I’m still not quite there).

So the thing just sat there. The designer who helped me with the first two prototypes left the project. All I had left was some circuit boards and a dream.

But I couldn’t stop thinking about it. Like I wrote in the first post, POV fidgets are 90% of the way to a dreidel already, and it drove my crazy that I couldn’t figure out that last 10%.

And so, this year, I decided to take another stab.

Actually, two stabs.

Did my dreidel actually already exist?

I like doing soldering projects with my kids. The things you can build are cool—rudimentary game consoles, clocks, calculators, FM radios, blinky light toys—and as long as you’re using through-hole components the kids can mount everything themselves, even use the soldering iron if they’re old enough. I’m often looking on AliExpress for new kits.

As I’ve mentioned, POV fidgets are frequently sold as kits. A lot of these just have flashing lights; they’re not designed to display anything as complex as words.

But a few seemed awfully close to what I wanted. One fidget in particular caught my eye: It had 12 LEDs, just six(!) resistors, and a very basic microcontroller. It came with a simple but effective laser-cut case. It could display text—even scrolling text, along with rudimentary animations!

What most excited me, however, were four tiny holes on the side of the circuit board. They’re labelled GND/TX/RX/VCC.

If you know anything about electronics, you already know why this is exciting. VCC and GND let electricity flow to and from the board. RX and TX are for data transmission.

These four holes meant that the board wasn’t locked off; it was open to the outside world. And if it was open to the outside world, it could probably be reprogrammed.

Then I hit upon a second bit of luck. I found another kit, almost identical to the first, that was designed to be reprogrammed. It even came with a programmer, which is the bridge between the board and a computer. And it came with software designed to do the reprogramming!

Was this the solution?

Two microprocessors diverge in the woods

The truth is that I still don’t know. The previous two posts were about events that happened years ago. This post is about something still under development.

The advantages of repurposing a pre-designed device are obvious: the production pipeline already exists, and the design costs have been eaten by somebody else. Also, this device is about as bare bones as you can get, which means that costs stay low. The design is clever.

But there are disadvantages, too. The software to reprogram the device, along with all documentation, is in Chinese. This isn’t an insurmountable problem—point your phone at a computer screen and you can translate just about anything—but it’s very annoying. The bigger problem is that the software doesn’t accommodate the most crucial function: choose a random letter every time you spin the device.

A second issue is the processor. The device uses an STC8 microcontroller, which is designed to be cheap and power efficient. Unfortunately, the STC8 also has very little internal memory, which creates hard limits on the sophistication of the graphics. It’s also not the chip of choice for a reprogrammable device; it’s more likely to go in a rice cooker than a 3D printer. Finally, the chip was designed in China and outside of China it doesn’t have a big hobbyist community. All this means that going with this device would probably mean locking it down to everyone except developers. It would be a toy, not a platform.

Some of these problems are surmountable. The factory that makes this device is happy to preload the chips with whatever code I want, and the Chinese software is just for encoding text, which means it’s not strictly necessary. One path I’m pursuing is to just create new code—and in fact there are reasons to do this no matter what, as I’ll explain in a different post. The viability of this path was demonstrated by a video I received from one of my developers, who successfully reprogrammed the device while bypassing the software.

Still, this path only works if not a single component needs to be added. It remains to be seen whether this is possible.

The alternative is to go back to my own design and just make it better. I can stick with the ATmega328 microprocessor, which is more powerful and easier to program. I can control the physical shape of the device. I can add additional buttons and USB connectivity. This device would be more costly, but I’d have completely control over it, and it would be much more powerful.

So I’m pursuing this path, as well. The Dreidex v0.3, prototyped earlier this year, is only slightly smaller than the Dreidex v0.2, but I’m already starting to see promising results. We’re going to shrink it further.

Perhaps no in-progress footage has made me happier than the twelve second test clip here.

Hanukkah, the holiday of lights. In the darkness, a dreidel shines.

But something is missing

I still don’t know which direction will ultimately win out. It’s possible that I will end up shipping both. The twenty-sided dreidel exists in regular and deluxe versions; perhaps the Dreidex will, too.

But in the course of developing these devices and showing them to people, it has become clear that the hardware is only half piece of the puzzle. The device needs to display a random Hebrew letter on each spin, but the way that it displays that letter is what will sell the device. It’s time to think about the user experience, the feeling that people get every time they spin the Dreidex. What does this device need to display to give you a sense of delight? What does it need to do to make you feel like you’re playing dreidel?

One of the most delicious moments in creative work is realizing that you’ve created an entirely new problem that requires even more creativity to resolve. In this case, the development of dreidel hardware has forced me to think through how exactly a digital dreidel ought to be coded.

The dreidel hardware is in development.

Next up: the world’s very first dreidel software.

This story will be continued. Happy Hanukkah!

That’s what building the Dreidex has felt like. Only more so.

(This is part of a series on creating the world’s first hackable electronic dreidel. You may want to read part one to understand how I got here.)

Hardware is hard

I know a little about electronics, but not that much. I can solder an FM radio if you give me all the components, and I could probably explain what each of them do—but I couldn’t design a radio, let alone create a custom circuit board for manufacture.

Still, it was clear to me that the dreidel of the future was going to be a persistence-of-vision (POV) fidget. I knew that I needed help, so I started working with a hardware designer in India.

The basic design was clear from the many hobby devices that have appeared since 2017, but I wanted to make my dreidel a little slicker and a little more capable. For the first prototype, we made two small changes:

• We added a programmable function button toallow the user to switch between diaspora mode (נ–ג–ה–ש) and Israeli mode (נ–ג–ה–פ).1

• Most consequentially, we decided that the device should run on rechargeable lithium coin-cell batteries. Because most people don’t have a recharger for coin cell batteries, this meant we needed to built charging circuitry into the device itself.

The second change turned out be massively consequential.

In order to charge the batteries, we needed a way to feed it AC power. The easiest way to do this was to add a USB-C port. Easy enough.

But that USB-C port tempted us to do more. It was there to supply power…but couldn’t it easily supply data, too? Well, yes!

Just like that, the idea of a programmable digital dreidel was born. We’d have the device ship with factory-installed dreidel software (which is a very strange sentence to say out loud), but we’d give the user the ability to turn the device into anything they wanted.

In order to make this process easy, we settled on the ATmega328 as our microcontroller. This is the same processor used in the extremely popular Arduino Uno, which is commonly used in all sorts of small hardware projects (sensors, small robots, kinetic art sculptures). It is very well understood. As a bonus, it can be programmed with the well-documented Arduino software toolkit, which massively simplified development.

This ultimately led to a prototype board that looked like this:

The design went to EasyEDA, a circuit board prototype manufacturer.

The prototypes came back. We noticed a mistake and revised. More prototypes came back.

And this is them—Dreidex v0.1 and v0.2!

Many times in my life I’ve had the experience of holding a book that I’ve worked on for the very first time. This was like that—but this thing had batteries and cool little lights!

Both devices worked, but both had problems.

First, it turns out that fidget spinners don’t use regular ball bearings. Most bearings have internal lubrication to reduce noise and vibration and to prevent the metal parts from grinding themselves to death. The problem is that lubrication also increases friction, which means the devices couldn’t spin for more than 10 seconds at a time. My designer needed to go out and find fidget ball bearings, which are open and allowed the device to spin for at least 30 seconds with a casual flick.

Finally, we got to the point where the device could display text.

(Unfortunately, footage of POV fidgets don’t film great because the phone’s frame rate doesn’t mimic the way it appears to the human eye. Trust me that it looks better in person!)

You can’t just ship a dreidel circuit board

This was an encouraging sign, but major obstacles remained. Most importantly, the device needed a case.

This proved to be quite difficult because the case is actually essential to a POV fidget’s operations. In order to display text that appears to hang in the air, the fidget needs to know how fast it’s spinning and adjust its display frequency dynamically; otherwise, the text will contract as the fidget slows down, as you can see below.

There aren’t a lot of ways to do this well. You might think that the obvious choice is to give the fidget a gyroscope (like the phone in your phone), but cheap and tiny gyroscopes aren’t designed work on devices that are spinning at 1,000–3,000 RPM.

Instead, almost everyone who has made a POV fidget has used a Hall effect sensor, which senses magnetic fields. The sensor is mounted on the board very close to the bearing (in the photos above it’s labelled IC2). A small magnet is then embedded in a plastic cap that you hold while spinning the device. The board rotates; the cap doesn’t. By measuring how quickly one passes the other, you can determine the rotational velocity and create graphics that appear to be hanging in the air.

In short, you end up with something like this.

In theory, this should have worked like a charm. In practice, however, it turns out that Hall effect sensors are kind of finicky, which meant that small imperfections in the case could made the device stop functioning.

Worse, the case was big. An adult could comfortably hold this dreidel between thumb and forefinger, but a kid couldn’t, and the case only made the device chunkier. Fidgets are supposed to have style, and this fidget just didn’t feel sleek. The idea was cool, but when I showed the device to people they didn’t seem all that impressed. Something was off.

The dreidel winter

And then…I stopped working on the project.

Everything that I’ve described up until this point happened back in 2020–2021. I had built two prototypes and the device, if carefully coaxed, could do something interesting. But I was still far away from a functional product, or even something I could share for preorder purposes. It just didn’t seem that impressive.

Worse—I was worried about losing the idea. I’ve heard too many horror stories about intellectual property getting stolen when people announce their products too early. Someone in the dice manufacturing world even told me that a product was copied and shipped using the promotional images he had created for his Kickstarter.

So the product was in limbo—cool enough to fantasize about, but not advanced enough to tell anyone about. I’d carry a prototype around in my backpack and tell people about it when the moment felt right; it was a kind of totem of a product that could be, or maybe of my aspirations to be a hardware manufacturer. But Hanukkah after Hanukkah ticked by, and the device was no closer to being in the world.

And then—this year—I had a breakthrough. Actually, two breakthroughs.

This story is continued below.

Of all the work that I do, creating dreidels is the most frivolous. They’re literal toys, tools in a game that itself is terribly designed. Nobody likes playing dreidel for more than a few minutes. They are the least important part of one of the least important holidays.

But the dreidel is also very malleable. It’s a quasi-religious object that’s also a game device—and because it serves no ritual function, you can’t really make it “wrong.” If people love your redesign, you’ve won.

In the past few years I have accidentally built myself a little dreidel empire. The Dreidel20—a twenty-sided die that is statistically identical to a regular dreidel (this is literally how it is advertised)—is sold in both Judaica and game stores all across America, and people buy thousands of them every year. People are tickled by the shape, which is great for me because I patented it. I’m not going to retire off dreidel money, but the profits are meaningful supplementary income.

Most people don’t make money playing with dreidels. I’m one of the few people who do.

But the profits don’t make the product less silly. Against the other things I work on—AI, psychedelics, Jewish futurism—it simply doesn’t register. If I was smart I would just take the win, enjoy the income stream, and move on.

But I keep coming back to it. I keep talking about it. I think about dreidels, about how I might reinvent them next.

Why? Most of what I produce is intangible; at best, I’ll see the fruits of my labor bound in a physical book. Dreidels, on the other hand, feel solid; they’re good to hold, to roll around, to keep in your pocket as a lucky charm. Looking at them, I always think, “This did not need to exist—and yet it does.” The reality of that existence provides me with a simple delight. It feels audacious. Even if I never sold another dreidel, that delight would remain.

A few years after the success of the Dreidel20, I decided to make a deluxe version. Instead of a crude pouch, I put the dice in a little black box with a black foam insert, as though they were little gems. I made the dice bigger, and also transparent—and then I had the manufacturer spend a “normal” dreidel inside of each one, as an homage to the toy’s origins.

It feels stupid to say that this little toy is a physical manifestation of how I see Judaism, but it is exactly that. The faith is reimagined—but the origins are still visible, embedded within, making the new more substantial, giving layers to something that would otherwise be flat. In time, the new will become embedded in something even newer. And on and on.

There’s more of myself in these dreidels than I’d like to admit. Maybe that’s why I was not able to stop myself from working on a new type of dreidel, this one unlike anything that exists on the market.

This is the story of a multi-year effort to make the dreidel of the future.

No, you can’t order it yet or even pre-order it (soon!), and the story is not yet complete. Today I want to tell you how it all started.

An even more popular spinning toy

One disadvantage of the Dreidel20 is that it makes the game very brief. A lot of the dreidel game is basically just waiting for the top to stop spinning, and a lot of the fun of the game is the spinning itself. The Dreidel20 is good for kids and people with dexterity issues, but it sucks something out of the game. I wanted to find a way to preserve the dreidel’s spin, but reimagined.

Enter the fidget spinner.

Fidget spinners haven’t been cool in a long time—they peaked in 2017—but in their heyday the idea of dreidelizing them occurred to several people. There are variations on the dreidel fidget, but they all basically look like this.

These fidgets are cute, but they have a problem: they’re not functional dreidels. Sure, you could arbitrarily stop the fidget spinning, but there’s no natural way to decide what its orientation means because it never definitively “lands” on anything. The payoff is weak. They’re fine fidgets but they’re not really dreidels at all.

So a normal fidget doesn’t work—but a subspecies of fidget could.

Enter the POV fidget

Amid the fidget craze, a few electronics-oriented folks noticed the same thing: The high speed of the fidget’s blades was more than sufficient to support a persistence of vision (POV) effect display.

POV displays are very old. The zoetrope, often described as an ancestor of motion pictures, is an early example. The basic idea is that at sufficiently high frame rates the brain fuses static images into fluid motion.

One of the cool things you can do with this effect is create an image from a single, fast-moving light source. Picasso famously did this, although the camera’s exposure is long than the effect in the human eye.

OK, but what about a line of light sources, like a column of LEDs? Well, if the LEDs speed across a given area while rapidly changing, they can create the illusion of a full image. If they speed across that same area repeatedly—moving back and forth or in a circle—you can create motion graphics. All from a single string of lights.

POV displays used to be expensive novelty gifts, the stuff of SkyMall catalogs. These days they’re very easy to make—the electronics are cheaper and plentiful—but they still have very few applications; whatever you want to do, a normal monitor will usually do it better. Worse, because POV displays are moving quickly they can cause injuries, are prone to failure, and rely on motors that use more energy than the common LCD display. The only interesting commercial application that I’ve seen is to put graphics on bike wheels.

Despite being niche, POV displays are often built by hobbyists because they’re not that complicated and they look cool—and since fidget spinners rotate quickly by nature (and are powered by hand, not motor), they were immediately recognized as good candidates for a POV treatment.

I have no idea who made the first POV fidget spinner, but they’ve been around since at least 2017. The website Hackaday records several prototypes posted in June of that year, including the image below comes from this article.

In the years since, many hobbyists have made similar devices. Sometimes they have three blades, sometimes they have bluetooth, but the basic design has remained the same: a circuit board in the shape of a fidget with a strip of LEDs on one blade, a microcontroller to handle the graphics, and a coin cell battery. That’s all you really need.

Despite their relative popularity, nobody seems to have figured out what to do with them. It’s notable that you can buy POV fidgets on AliExpress—but only as a soldering project, not as a ready-to-use toy. The idea is that they’re fun to build, but not that fun to use. With the fidget fad now firmly concluded, it didn’t seem like there was much interest in moving this hardware forward.

90% of a dreidel

I don’t remember the first time I saw a POV fidget spinner, but I do remember my immediate reaction: This is 90% of the way towards being a dreidel.

It rotates. It displays information. It can be randomized. All you need to do is program it to display Nun/Gimmel/Heh/Shin every time it’s spun.

Deliciously, such a device would reverse the meaning of the dreidel spin itself. Normal dreidels are inscrutable while in motion; you can read them only when they are still. A POV device does exactly the opposite: It displays no information while motionless but resolves each time it spins. The reversal itself was interesting to me—after all, who doesn’t want a dreidel that displays your result in lights that appear to be hovering in the air?

And so, I decided to make a POV fidget dreidel.

What about just putting a POV display in a standard dreidel?

In a perfect world, you might apply the POV effect directly to the body of a classic dreidel. In fact, a guy named Nathan Petersen did something very similar, and the final product looks great. The problem is that a dreidel spins too slowly (<1,000 RPM, vs. at least 2,000 RPM for a fidget and often far higher) and loses speed too rapidly. Perhaps you could give the dreidel an internal motor, but that would dramatically complicate the device and you’d lose the tactile joy of spinning the thing for yourself.

The path was clear: despite having no connection to the holiday, the fidget was a better form factor. Building on the success of my simple dice—and relying on the capital that its success had afforded me—I got to work.

This story is continued below.

Subscribe now

That’s all I’m going to say. Go watch it—and (as the saying goes): like, subscribe, and share.

Subscribe now

Though neither God nor the Bible ever intended it, the Hebrew month of Elul (which is now) is the crux of the year, the eye of the snake that is set on eating its own tale. The mournful Tisha B’Av has passed, as has its insistence on reverberating failures down the centuries and its fixed story of paths that cannot be taken because the roadmakers were slaughtered and their children are lost in the woods.

Elul is the crux because it comes before the contradiction to all that. Rosh Hashanah is premised on the idea that there is always another beginning, and that beginnings are just as inevitable as ends. Tisha B’Av: the mournfulness of a people that can never turn around, not even to say goodbye to their own past, never receiving closure, never allowed to be released into a pillar of salt. Rosh Hashanah: the clap of the slate that signals the start of a new take, one, two, three, four, ten thousand and one, two, three…

So Elul is a tangled mess. The shofar blasts warn of what is about to happen, but do we want it to happen? As life goes on and the years whip past faster and faster, the idea of newness feels less and less possible, more and more of a farce, just another candle on a burned-out cake. In the midst of a nation-spanning story of decay that we can all see and all dread, what is the point of the fantasy that we are at the beginning? How can we do anything but resent the notion that this is the start, no this, no this, no now. Why should we trust that this start will conclude in anything but another Tisha B’Av, all raindrops flowing to the same ocean of tears? Rosh Hashanah believes that time is a circle, but the circle always unravels into a tangle, and the snakehead of the year always slinks into a snake tail that finds itself in its own jaws. Who would do this to themselves?

I don’t know, but I’m an Elul man. I am caught like this: between regret and renewal, between permanence and yet another new start. The first obligation of the Bible is that the moon is always new again; the first letter of the Bible is the second letter of the alphabet, and yet it starts the word bereishit, “beginning.” True beginnings always bloom in the place you least expect: the middle.

And so, we go into the new year with that same insane insistence that we don’t know what will happen next, that this year may really end in Jerusalem, or turn purple and dance a polka, or drop a holiday or two. We begin the new year in spite of ourselves and our infatuation with history and our unshakeable trauma. We begin knowing that opening ourselves up the possibility of change may create more scars, will expand the space in which we can feel disappointed with ourselves.

We can never begin again. We must always begin again. Against odds, we hope.

Subscribe now

In a little less than a month, several Protestant denominations will be gathering in Minneapolis to discuss their policies around artificial intelligence. I don’t know whether the conference will result in anything material, but the effort is significant—and it is far from being an isolated event. The new pope’s very name was chosen in reference to AI’s world-changing power. Social conservatives, many motivated by Christian faith, are trying to shape the Trump administration’s AI regulatory efforts and put pressure on big AI firms. Less than three years after the launch of ChatGPT, Christian groups are finding their footing on AI.

Jewish efforts, by contrast, barely register. As far as I know, there are no permanent Jewish groups dedicated to developing AI thought or policy, no dedicated networks, no ongoing efforts to engage the many Jews working in AI, and no serious efforts to bring together American and Israeli thought on AI. (If I’m wrong about any of this, please correct me.)

Now, I don’t think this deficit is permanent, and I’m not exactly a disinterested observer. I’m seeking to change this situation, I’m looking for partners, and I see glimmerings of other efforts. But the contrast is notable, and the reason is pretty obvious: the war in Gaza.

October 7 stopped AI ethics in its tracks

Beginning with the launch of ChatGPT I began getting invitations to provide a Jewish perspective on AI, largely but not exclusively for Jewish audiences.

And then October 7 happened and interest immediately stopped. Over the next twelve of so months the only Jewish institutions interested in talking about AI were the ones who (a) had scheduled stuff prior to the start of the war or (b) were specifically looking to not talk about the war. For 12–18 months, interest in AI remained pretty low, even as the technology’s applications and pervasiveness skyrocketed.

At the same time, AI ethics has been part of the Gaza war. Along with the war in Ukraine, it will be recognized as one of the first conflicts in which AI played a major role in determining targets. This role has likely made it much more complicated for Jewish organizations to talk about AI ethics generally, despite the fact that Jewish thinking about autonomous tools (both cars and weapons) remains more sophisticated than Jewish thinking about generative AI. A discussion that in 2019 was very theoretical has now become so personal that it cannot be removed from contemporary politics.

”AI for Torah” is cool but it’s not enough

The one area where I do see Jewish AI developing is in education. Lots of people and organizations are interested in using AI to make their ideas more accessible. Organizations dedicated to Jewish learning wants to use AI to enhance learning experiences, automagically create worksheets, or AI tutors, or AI study partners. Sites that have proprietary stores of Torah want to use AI to make that knowledge more accessible.

I don’t begrudge any of these efforts. In general I’m quite open to the use of AI in Jewish learning environments. However, I am very concerned that we’re going to repeat the mistake we made around social media, where the allure of the product as a broadcast tool prevented use from properly critiquing the product itself. We can’t afford to equate Jewish AI thought with Jewish AI applications. That would be a major lapse in our responsibility to deal with one of major technologies of our time.

People are looking for guidance, right now

An obvious retort is that Christians have more power than Jews—and maybe especially so right now, given that most American Jews want little to do with the present administration. If your concern is rising autocracy, even something as big as AI ethics can feel small by comparison.

But as I frequently say, AI policy is not just about government and corporations. The American public continues to maintain its generalized anxiety about the use of AI but individuals do not feel that they have any agency to change the technology’s direction. This situation has only gotten more distressing as AI applications have become more personal. Last year we spent a lot of time talking about how AI will change education and create disinformation. Now we’re talking about AI therapy and porn.

If you’re a Jewish leader, consider that your constituents would probably like answers to the following questions:

• How should I control my kid’s access to AI? Is there a “safe” age to use certain AI applications?

• In what circumstances in work life should I specifically ask that AI not be used? When should I retain an employee even though it’s not resource-efficient?

• Should I insist that sermons be AI-free—and if that’s not realistic, how does that change how we think about Jewish thought and law?

• Do human teachers have more than utilitarian value?

• How should I navigate the threat of solipsism that can come from using AI as a therapist?

• What communal structures need to be built to make it easier for people to mediate their use of AI?

• How should I prioritize my various concerns about AI—and should I be particularly worried about specific models?

These aren’t questions about setting national AI policy; they’re about creating cultures of use. Right now the only people creating those cultures are the companies developing the tools. There must be a counterbalancing effort—and that effort will be most effective if it’s local.

AI ethics is just as existential as battling antisemitism

A people can decline because members are being physically attacked. But they can also decline because they lose track of their purpose. You can lose your body, but you can also lose your soul.

To the extent that Jews see themselves as a people trying to make the world a better place, AI ethics must absolutely become a well-developed area of Jewish thought. You don’t get to cherry-pick areas where you develop moral positions because they’re more convenient or more closely linked to the sources in your arsenal. AI ethics isn’t a reach goal; it’s a necessary condition of our continued existence as a light unto the nations (or even a light unto ourselves).

Jewish AI thought is lagging. If AI were moving more slowly that might not matter; who even remembers which religion responded first to the invention of the telegraph? But in a world of fast-moving technology that promises to change the world irreversibly, speed itself has moral significance, and it also has power. Maybe nobody can move as fast as a tech firms that make it their business to be faster than anybody, but our reaction time still matters a great deal.

To quote Hillel: If not now—when?

We seem to be having three AI conversations that are actually one and the same.

In the last week, I read the following:

1. A journal article arguing that AI-created scripture (in this case Buddhist sutras) could contain real meaning

2. A New York Times opinion piece by a therapist praising AI therapy (gift link)

3. A piece in Wired about the rise and addictiveness of self-generated AI pornography

All of these articles are about people finding meaning in AI output that they themselves have generated. The meaning can be spiritual, emotional, or physical—but it is meaning that people enjoy because it comes out of you and yet feels like it is something more than you.

The conversation for each use case follows a predictable pattern.

The people who don’t like AI-generated bible/therapy/porn say that they are dangerous because they (a) desensitize people to the genuine article and (b) trap people in a solipsism where it becomes increasingly difficult to remember what the outside world had to offer.

The people who like these things, on the other hand, say: Calm down, I’m not stupid. I know it’s an extension of me, but the illusion is helpful. Haven’t you ever had a conversation with yourself, or an imaginary friend? Haven’t you ever fantasized? Haven’t you ever talked to God, or eaten magic mushrooms? Why do you care how I externalize my internal self?

I think the pro-therapy/porn/scripture camp is right in theory, but the nay-sayers are right in practice. In theory, AI is just another tool for self-exploration. In practice, its efficiency and versatility change how we think about self-stimulation.

Subscribe now

The sound of one hand clapping

The main reason AI is different is that it is extremely good at satisfying individual desires for content. Unless your interests are radically different from all of humanity, the thing you want is likely a rehash of other already-fulfilled desires.

This is a step beyond the plain old internet. As Bo Burnham reminded us, the internet gives you a little bit of everything all of the time (YouTube), resulting in the feeling that boredom is something to be permanently banished. But people constantly get bored on the internet! We scroll TikTok despite finding nothing worthwhile, we search for something worth watching on Netflix, we check for messages instead of looking out the window. We raised our expectations for stimulation because of the web, and then the web disappointed us anyway.

The problem? No matter how much content is out there, nobody can know exactly what you want. Billions of dollars have gone into creating oceans of content in the hopes of striking true for as many people and as much of the time as possible, but nobody knows your exact mind and not everything you want is going to reach you, despite the best efforts of many algorithms.

In this online space, standing on the shoulders of an internet’s worth of data, AI now offers a new proposition: Just tell us what you want and we’ll make it for you, right now. It doesn’t matter if your desires are erudite or sleazy; it can be yours just the same.

And the process of creating this content is the same, highbrow or lowbrow. You tell the machine what you want, it gives you something imperfect, and then you refine, refine, refine. Respond as though you were a cosmic consciousness. Be more probing when I talk about my trauma. Make her breasts bigger. The machines are good; in the space of a dozen refinements, you may end up with the full version of something that only existed as a glimmer in your mind—and that fully realized form may actually bring you to enlightenment, or tears, or orgasm. Your clear specifications, hashed into everything the internet already knows, is a powerful combination. Very often you will get exactly what you wanted and much more.

You can do anything, the only limit is yourself

Because the AI only needs to satisfy one person, the success rate is fairly high. You may be aware that there’s an ongoing debate about whether AI capabilities are about to hit a brick wall or see exponential growth, but for the purposes of self-fulfillment processing power really may be all that’s needed. The difference between a therapy chatbot and a therapy voice conversation is mostly just “compute.” Today you can get enlightening scripture, but next year maybe you’ll be able to experience revelation. The number of routes to self-fulfillment will only grow.

I say this because you may think the use-cases are limited. If it’s just porn and emotional support and spiritual enlightenment, what’s the big deal?

Well, it’s a big deal because the M.O. of the AI industry to try to fill every possible role and see what sticks. There are companies that want AI to be your friend, your tutor, and your boss. You may already love an AI musician. If you treat every instance of humans relying on humans as something to be optimized away, you are definitionally trying to sever human connection. Make a list of all the times during the day that you interact with another person—physically or virtually—and you’ve made a list of AI optimization projects.

AI and the Great Turn Inward

Treated separately, each of these use-cases is relatively innocuous, but the concern is larger than just whether AI is a good/fine/dangerous therapy substitute, or whether gaining spiritual insight from a bank of GPUs is "valid," whatever that might mean. Instead, AI's ability to act as a perfect magic mirror for our internal ruminations makes the task of growing by engaging with the world feel laborious and inefficient by comparison. Finding a good therapist sucks; lots of people never find a spiritual guide in life. Many people would take 70% of the benefit for 5% of the energy.

In other words: the problem isn't that these applications do a bad job at making scripture/porn/therapy. It's that they often do a great job—but in convincingly playing the role of the other without the pesky need for actual others, they preempt many of life’s key relationships. Lose one or two and you’ll probably be fine, but industry trends may push you to drop most.

You may think that you know what this world looks like. Decades after the rise of social media, we’re all aware of echo chambers and the thinness of online connection and the way the internet can distort your sense of self and reality. We already interact with people online as though they’re not really people. Would AI be that different?

Yes—because AI reflections of the self feel great. With AI, you’re not going to feel Instagram-induced feelings of insufficiency, or Twitter-induced feelings of rage, or TikTok-induced feelings of hollowness. Whatever friction you feel from AI’s outputs can simply be refined away. Instead, you present yourself with content that you have tailored to be exactly as challenging/alluring/mystifying as you want. If you want sexual satisfaction, you can get it. If you want to confront your faults, you can get that, too. AI will present as an “other,” but always an idealized other—in other words, not really an other at all. In the words of Jean-Paul Sartre, you need never expose yourself to the hell that is other people.

This might be fine if the self was fully known and static. But we never fully know ourselves, and our selves are in constant flux. We don’t know always know what we want or what we need. By optimizing the task of self-connection, AI creates a new and dangerous illusion: that our deepest desires can be satisfied by looking ever inward. Therein lies the path of self-hallucination.

AI Can Be Your Own Worst Critic, If That’s What You Want

You’ve probably figured out by now (I hope!) that AIs are most useful when they’re doing something you know how to evaluate. This is just as true when the subject matter is the self.

We are not in fact perfect experts on our own selves, which means that we won’t always be able to tell when masturbatory AIs are going off the rails. Worse: We may shrink our selves to what the AI is able to conjure. We may not have enough contact with the outside world to know what we’re missing. We effectively create an echo chamber of one.

I’ll give you an example that I find particularly tragic. On Reddit’s AI discussion forums, I sometimes see people posting ideas for how to prompt AI to help on journeys of self-exploration. Here’s an example:

I want you to help me understand myself better — not just in terms of personality or productivity, but in the deeper structural sense. Ask me questions that reveal: – where I’m lying to myself – where I’m performing instead of expressing – what I might be grieving without realizing – what kind of world my nervous system was actually designed for

I want you to use everything you can gather — not just what I say, but what I avoid saying. Track patterns, contradictions, emotional signals, gaps in logic, and shifts in tone.

Don’t simplify anything. Don’t protect me. Go as deep as possible, using all your insight.

Start when you’re ready.

Now, what does this prompt represent? To me it sounds like a person trying to create an ideal version of a severe inner critic (“Don’t protect me”). That’s not really what a therapist is supposed to do; in fact, therapists often help people understand their own internal thought processes, rather than reify them.

The thing is: if you like these particular questions, the AI is probably going to tell you something that you find meaningful. There aren’t going to be any alarm bells telling you that you’re drowning in your own mindspace. Like a carbon monoxide leak, you may not know anything is wrong until it’s too late.

(Ironically, the people who see this most clearly are the ones pointing out the addictiveness of AI pornography, which allows people to visualize their precise fantasies beyond what any mental image could provide. AI pornography is incredibly popular. Civitai, a platform where users can upload AI models for others to use, is mostly porn.)

An Audience of One

You can sometimes tell when a person is lost in the AI sauce.

Much like an unsolicited dick pic, getting AI output when you don’t want it is a real bummer. I’m sure you’ve had the experience of someone sending you AI content that they think is fascinating but you find incredibly dull. You may have also had the experience of asking someone for advice only for them to tell you, “Here’s what happened when I asked ChatGPT.” I found myself feeling embarrassed on the other person’s behalf because it reads to me like the other person can’t tell when human connection is the point.

But this keeps happening because people are often deeply affected by the output they create—to the point where people develop romantic relationships with their chatbots. What we feel is genuine, just as the meaning one might get from AI-generated scripture is genuine. But the more you sink into these bespoke products the more the rest of the world will feel harsh by comparison.

We are all high-functioning AI addicts—for now

The earliest stages of any addiction are the most fun because the body and/or mind isn’t prepared; a first-time opioid user feels euphoric because your body is operating normally. You get the best of both worlds!

But there are no free lunches. Your body is efficient and adjusts to what is given, assuming that it will keep coming. As Judith Grisel writes, drugs that make you calm actually cause anxiety and drugs that ease pain actually cause it to increase.

Our relationship to AI is something like that, but we are too early in our journey with this technology to see it clearly. Most of us got educated in an AI-free world; we have skills that AI duplicates, and still have jobs that AI threatens to eliminate. We have the luxury of redundancy, the luxury of using AI as a supplement because society has not yet fully adjusted to its presence.

But it will—and it will be painless until we come up for air. Our self-stimulation machines will inevitably lead us to lose interest in the stimulation of others. The time we spend with ourselves will increase as our ability to materialize stimulating material from the safety of our own minds dramatically expands. Why deal with others when the self contains such compelling multitudes?

The worst affected, of course, will be children. What happens to a kid’s development when you surround them with phantoms designed to bend to their perception of their own needs? AI tutors may replicate the “royal tutor” model of education, but remember that royals are notoriously weird people precisely because they don’t interact with normals. How do you create a common culture when kids are self-creating entertainment based on their 10-year-old interests? How do you create rules for society? Are we going to end up in smartphone-like fights about delay AI access as long as possible, to give kids as much chance as we can to develop a self that develops alongside others, knowing that we adults don’t really think any age is safe?

There’s a good line that folks in the tech sector use to defend mediocre products: “This is the worst it’s ever going to be.” Nowhere to go but up, right? In some cases this will be true. I suspect AI agents are going to get pretty good. Maybe we’ll cure cancer, too.

But with self-stimulating AI, this sentiment is almost certainly exactly backwards. These tools are safest when they duplicate stimulation available outside of the self; the more they replace the outside world, the more dangerous they become. Society will adjust to accommodate AI; social skills and educational skills will weaken on the assumption that AI will always be there. We will grow to depend on AI, forgetting that the dependence was caused by AI itself.

Heroin is better than family

People who have struggled with addiction will understand this all too well. Addictive drugs can feel so good that family, friendship, and employment feel inadequate by comparison. We are extremely vulnerable to intense pleasures, and we have a hard time limiting ourselves.

We might find spirituality in AI sutras, emotional connect in AI relationships, and climax in AI pornography. But the path in all three is one in which we allow our selves to become so big that the world simply recedes. That is a path we must profoundly resist.

I’ve spent a long time in the world of podcasts, and an even longer time as a writer. Both formats suit my personality because I’m a huge introvert and they make it easier to curate my self-presentation.

But video is the future the present, so despite the vulnerability it requires, that’s where I’m going. (Don’t worry, Belief in the Future isn’t over, and neither is Jello Menorah; more content is dropping soon.)

I’m here to announce the launch of a new YouTube channel: Sabbath Mode. I’ll be covering issues of religion and technology.

Watch Sabbath Mode’s debut video NOW.

I’m still finding “my voice” in this medium, so please do send me all of your suggestions and advice!

Here, of course, is the part where I shamelessly ask you to subscribe to the channel, like the video, and share it with your friends. I was lucky enough to have funding to start this channel but it’s a lot more time-consuming to make videos than podcasts. I can only keep making these if I can show that it’s reaching people.

Thank you so much for your support.

Subscribe now

As you probably saw, Grok briefly endorsed a new Holocaust. I wrote about this for JTA, arguing that Jewish institutions and leaders need to recognize that they are particularly vulnerable to AI because so much of what Americans know about Jews comes from the internet.

It’s really important to recognize that this isn’t just an Elon Musk issue, nor will the bias always be so blatant. There are very strong incentives for bad actors to “groom” AIs into accepting their nonsense as fact, and Russia already appears to be doing exactly this.

We’ve seen AIs fall for falsehoods in the past. An earlier version of Google’s embedded search AI told people to eat a rock every day because of an article in The Onion. Sometimes you can fix the problem by telling AIs that certain news sites are untrustworthy, but at some point they need to trust what they read on the internet if they’re going to be useful; you can’t be incredulous about everything. A state actor with plenty of resources could diffuse an idea across the web to such an extent that it becomes hard to tell fact from fiction.

I don’t think there’s a permanent fix to this problem. The only way to deal with it is to understand that the potential for antisemitic AI is always going to be present, to demand more audits of advanced systems, and to discourage use of systems that prioritize user engagement over truth.

Let me know what you think. New post coming soon.

One of the things I say frequently is that I’d like to see psychedelics normalized within Judaism.

Saying this is always a little scary, because if I don’t immediately explain what I mean people assume I have a bunch of LSD in a backpack and am about to hand out free samples. (I don’t.)

There is certainly a long history of psychedelic advocates pushing widespread use. I am not among them. This isn’t just because such evangelizing could easily backfire (in fact, it did), but because I genuinely don’t think that psychedelics are the right tool for everyone. While legalization advocates have an incentive to emphasize their unique importance (“if X isn’t legalized then Y group can’t do Z!”), they are ultimately just one tool among many. Many people live healthy and fulfilling lives without ever consuming a psychedelic substance.

At the same time, banning psychedelics and marginalizing discourse about them can indeed have negative effects. For some people these substances serve a vital therapeutic or spiritual function that no other substance or practice could easily fulfill—and while the path to FDA-approved treatments is currently murky, it is still a real possibility. Yes, they’re just a tool—but for some people they’re the right tool. Given that the harm potential of psychedelics tends to lower than that of legal substances like tobacco and alcohol,1 I tend to think that the upside of legalization outweighs the downside.

But while legalization is necessary for normalization, it it not at all the same thing. When I say normalization, I mean social environments where the recreational, clinical, and spiritual properties of psychedelics can be discussed openly, where communities and their leaders understand best practices for use, where the benefits and dangers of psychedelics are well understood, and where educators have the resources they need to understand psychedelics.

In short: normalizing psychedelics means better transparency, better safety, better knowledge, and better practices.

The big question: why should non-users care?

Now, I think it’s fair to say that most Jewish psychedelics users would like to see normalization happen. The problem is that if you’re not evangelizing drug use your only tool for doing this is words; specifically, you need to make the case that non-users should support the open discussion of substances that they personally do not intend to take.

The existence of these non-users is a hard fact that users don’t always take into account. Even without an interest in evangelizing, the energy inside psychedelic communities can make it feel like the world is just on the verge of waking up. While it is true that we’ll probably seen increased use, psychedelics just aren’t for everyone and they never will be. For some they are positively dangerous, and many others simply don’t want to. Some psychedelics cannot be safely consumed while taking antidepressants. Even for those who are curious, it may be months or years until they are in a personal/professional place where they feel comfortable doing something that could very well upend their life.

The permanence of non-users weighs inconveniently on psychedelic advocates. Even assuming that most non-users are largely neutral on psychedelics (in my experience we’re largely past Just Say No and most don’t have much knowledge about them at all), asking communities to devote resources to educating people about substances that are—let’s remember—mostly still illegal is a hard sell. An individual Jewish leader may even see the value in these substances, but convincing their organization (and their funders) that they should spend time or money is quite a bit more difficult.

The normalization of psychedelics requires a strong argument that psychedelics should be everybody’s concern.

So, how exactly do you do that?

Queer Judaism suggests a path for psychedelic Judaism

To understand how psychedelic Judaism might come to be everyone’s concern, consider another subgroup: LGBTQ Jews.

Most Jews are not part of this community; however defined, queer Jews will always be in the minority. Much of the discrimination that queer Jews feel in Jewish spaces flows directly from this fact—and because of this discrimination, queer Jews (like queer people generally) have frequently self-organized to provide guidance and support for their own.

In the face of persecution, queer Jews who care about Torah have often turn inward. The point of queer Torah, in one narrow understanding, is to help queer Jews find a place for themselves within the tradition. To the extent that straight Jews need to hear queer Torah, it is merely to demonstrate that queer Jews are serious and deserving of toleration (or, ideally, full acceptance)—or perhaps to make an argument for a policy change. In this model, queer Jews demand nothing from nothing straight Jews but toleration.

This is a reasonable position, and queer Jews are certainly under no obligation to justify their utility to anyone. However, it allows straight Jews to continue maintaining that queer Jews are deviants, deserving of respect on humanitarian grounds but ultimately abnormal. Many queer Jewish thinkers push back against this idea (if you have time, read Jamie Weisbach’s argument for wearing a chest binder in a mikveh, it’s a stellar piece of work). Yes, queer Jews are a minority—but they’re a minority whose very existence requires everyone to update their understanding of healthy relationships and parenting roles and childhood. Properly appreciated, queer Judaism should open everyone up to the sheer variety of ways in which people can love one another, understand their own bodies, and construct family. These are messages that matter to all. (This, for me, was one of the reasons that it was so important to publish Noam Sienna’s A Rainbow Thread.)

Psychedelic use is not an identity like queerness, but in this case it exhibits similar dynamics. Like queerness, psychedelic use lends itself to certain types of conversations that non-users may simply not have the life experience to fully understand—and psychedelic users, like queer Jews, have good reason to be circumspect about their activities and ways of thinking. There is a version of psychedelic Judaism that demands nothing from non-users but toleration: Respect us enough to leave us to our own devices.

But the insights of psychedelic use matter for everyone. Psychedelics can quickly and powerfully grant people insight about the subjectiveness of consciousness, the fragility of ego, internal dialogue, addiction, sense, pain and pleasure, desire, ritual, emotional processing, and even death. They can give people powerful spiritual experiences—and also make people wonder about the relationship between those experiences and the God-language that particular faiths choose to use. Perhaps most counter-intuitively, their use in religious contexts tends to be more highly ritualized and more communal than simple recreational use, which creates an environment where responsible use is encouraged and people struggling can get support.2 By engendering conversations about these things, Jewish psychedelic users can provide wisdom that benefits the Jewish community at large. (In fact, they have probably already done this but we don’t know because we haven’t done the historical research.)

So, the pitch to Jewish non-users goes like this: Even if you remain stone cold sober for your entire life, your community will shine brighter through the experiences of its psychedelic users. Its emotional processing will be richer, its text study will be deeper, it will be open to more learning modalities, and its spiritual practices will be more creative. (Also, not for nothing, for Jewish psychedelic users are energized. They’re excited about the future. Who wouldn’t want a piece of that energy?) All these things can accrue—but only if usage is generally safe, only if users feel comfortable talking to you, and only inside the framework that psychedelic Torah is Torah for all.

This all might sound nice in theory, but it’s quite difficult. It is hard talking to people who aren’t like you; as in other areas of life, it’s easier to stick with your own kind. The one silver lining? Despite self-organizing around substances, and despite how many times I have written “user” and “non-user” in this post, psychedelic use tends not to be totalizing. Its insights, despite their ostensible ineffability, lend themselves well to communication. This suggests that meaningful communication between users and non-users doesn’t need to be so difficult.

The paradoxical humility of the magic mushroom

Have you heard this one before?

A man rubs a magic lamp and the genie inside presents him with two options: great wisdom or great wealth. The man chooses wisdom. He instantly gets his wish.

The first words out of his mouth: “I should have taken the money.”

For me, this sums up the beautiful paradox of psychedelics. Taking a psychedelic can be one of the most powerful experiences of your life—but that experience may actually lead you away from using the substance in the future. This is markedly different from (say) fentanyl or cocaine, where an overwhelmingly positive experience can lead people to chase that first-use feeling for years, often destroying lives even as the positive feelings fade away and one is left staving off withdrawal.

While psychedelics experiences are often euphoric, most users do not conclude their experiences wishing to take more at the earliest opportunity. The substance, in retrospect, is beside the point. The point is existing in this world.

Consider the loosening of the sense of self that magic mushrooms frequently provide. Few people would want to live perpetually in this state; it is not inherently pleasurable, and on occasion can be quite frightening. Still, the very knowledge of this state’s existence changes how one sees the world. The sense of self loses some of its stickiness; it remains useful for daily life, but one now knows that ego is contingent.

Did the mushroom cause this state? Yes—but many psychedelic users would tell you that the state exists outside of the mushroom. The substance is giving you access to truth that is ordinarily blocked from view by your own self. The “trip” metaphor is actually useful here. Most Americans who fly to Paris for vacation come away primarily impressed by Paris, not the airplane that took them there—even if almost all Americans who go to Paris get there by airplane.

If it’s hard to see this point, it’s because the legal status of psychedelics forces them into the spotlight for both detractors and advocates. (If airplanes were illegal you’d probably talk about airplanes a lot more, too!) I don’t think this is the correct space for them, and under a different legal regime the way we talk about these substances might look quite different.

I believe that this self-decentering property of psychedelics is key to broad communal acceptance. You might assume that normalization would mean people are just talking about drugs all the time, but this is not the case (just as normalizing queer Jews doesn’t mean talking about sex and/or genitals all the same). For both users and non-users, psychedelics are primarily interesting for their ability to help people understand themselves and the universe. Some want these understandings firsthand; others want them vicariously. Jewish communities would be better off if both options were broadly available.

A pact for users and non-users

If psychedelics are going to be normalized within Jewish communities, it will require something from both users and non-users.

Psychedelic Jews and Jewish groups must learn how to communicate their insights to others. They should not assume that the world is divided between users and not-yet-users and focus on the insights gained from substances rather than the substances themselves. They must perform the scary task of being open about their psychedelic use and how it affects their lives and ways of thinking. Most importantly, psychedelic Jewish communities need to remain safe places and demonstrate their ability to self-regulate; nothing will scare people off more than the perception that psychedelics are dangerous—and non-users may not give the benefit of the doubt if there is a high-profile tragedy.

For their part, non-users should remain open and listen. If you’re a Jewish leader who has never used psychedelics have associations from the era of Just Say No, learn about the trajectory of psychedelic science over the last half century. Read Michael Pollan. Read Jay Michaelson. Pay attention to the work of Shefa. Talk to therapists who are open to psychedelic modalities. Keep an eye out for the books and articles that will hopefully be written over the next few years. And if your constituents open up to you about their psychedelic use, be curious.

Unrelated to psychedelics—you should know about this grant for Judaism and technology projects. The deadline is June 12, you can ask for up to $10k.

I didn’t fall in love with Jewish text study because of the subject matter. Like many people who learn Torah, I was seduced by the structure and culture of the learning itself: the layers upon layers, the imagined conversations across the centuries. Most of all, I fell in love with the energy: the experience of running between bookshelves, the close reading, the impassioned arguments, the sudden flashes of comprehension. When a beit midrash (study hall) is in full swing, it doesn’t matter if the people within are studying obsolete property rights or obsolete tort law. You feel the power.

At the heart of this experience is the havruta: two people reading texts together, for hours on end. The havruta is what makes Torah study feel less heady, and it places an emphasis on the slow but enjoyable process of reading texts deeply that feels increasingly out of step with the pace of the rest of our lives. It turns what would otherwise be a solitary experience into something social—and it forces you to be brutally honest with yourself about all the things you don’t yet know.

Some people find a havruta or two in school/yeshiva and then keep them for life. Others pick them up later. There are online programs where people can find a havruta. But most Jews don’t engage with Jewish learning, and so these rich relationships are unavailable to them. That’s a shame, because it’s the friendliest and most flexible way of engaging with the Jewish canon.

I wondered: Could AI fill this role—or at least supplement it? Would it be possible to converse with an AI about a text, using its wealth of knowledge to help you advance your ability to read Jewish materials? Can AI help replicate the experience (and joy) of reading with a havruta?

After a little digging, I think the answer is: not yet, but it’s doable—at least, if someone feels like it’s worth doing.

My experiment

I wanted to see if I could have a havruta-style conversation with ChatGPT about a piece of Talmud. By conversation, to be clear, I mean voice conversation—no typing, no looking at screens. I wanted to take advantage of the new emotive voice chat functionality that many AI companies have implemented over the last year.

Because I did my first tests while I was driving, I decided to use a piece of Talmud that was top of mind and partially memorized: the beginning of the second chapter of Bava Metzia. This is often the first piece of Talmud that new students learn, because it’s relatively simple and introduces a bunch of key concepts.

The first question: could the AI understand me at all?

ChatGPT understands Hebrew/English speech blends

As a podcast producer, I’ve spent a lot of time dealing with machine transcription problems. Speech recognition has grown by leaps and bounds over the last few years, but it still struggles when you randomly drop Hebrew words into English sentences. Unfortunately this is exactly what you need to do when you’re learning Hebrew texts in an English-language environment. If AI can’t understand these blends you’re nowhere.

I experimented with four voice models: Grok, ChatGPT, Gemini, and Claude. These are all cutting-edge tools that regularly trade the crown of “best in class” between them, but when it comes to this task only ChatGPT could understand the rapid language-switching.

For example, this is what happened when I asked Claude, “Hey, I’m trying to learn the second perek [=chapter] of Bava Metzia [=a book of the Talmud]. Can you help me with that?”

ChatGPT, on the other hand, understood me perfectly.

Transcript (provided by ChatGPT):

Absolutely! I'd love to! The second perek of Bava Metzia, which is called "Eilu Metziot," deals with found objects and the obligations of someone who finds lost property. It's all about when the finder has to announce a discovery, and when they can keep it. Let me know what specifically you'd like to explore—like a particular mishnah or some of the classic commentaries—and we can jump right in!

The ability to recognize interwoven languages unlocks huge learning potential. It means, for example, that you can use ChatGPT to help you learn a new language. It also means that ChatGPT fulfills a necessary condition for Torah learning.

The answer also reveals what I’ve noted previously: ChatGPT has picked up a huge amount of Torah knowledge from the internet. It understood the piece of Talmud I was talking about well enough to provide a synopsis of the entire chapter. This is a good sign!

The next step: start learning.

The good: ChatGPT answers questions the right way

The key element of havruta learning is talking through a text by reading it aloud and puzzling out its meaning with your partner.

This can be hard to teach if you’re not familiar with the practice. People are used to reading silently and reading aloud makes you vulnerable, especially if you don’t know the language. But this style is essential to Torah study; it emphasizes textual knowledge as the basis for all learning, it’s an amazing level-setting technique, and it turns reading into an energetic activity. The characteristic cacophony of the beit midrash (study hall) wouldn’t be the same without it. (Here’s what that sounds like, if you’re not familiar:)

It took a mental adjustment to start talking to an AI like I would a study partner—but when I did, I was pleasantly surprised to find that the AI understood me.

I started by reading the text in Hebrew. (Note: for the rest of this post there are going to be a lot of untranslated terms. Don’t worry about them, I will explain the relevant parts.)

אֵלּוּ מְצִיאוֹת שֶׁלּוֹ, וְאֵלּוּ חַיָּיב לְהַכְרִיז.

Eilu metziot shelo, ve'eilu chayav lehachriz.

“First of all, what does ‘lehachriz’ mean?” I asked. I didn’t offer a translation of what I had just read, only asking about the word that I had decided to get “stuck” on. This is a standard way of learning with a partner: some texts are obvious enough that they don’t require comment, others stop you because (1) they’re interesting or (2) you don’t understand what they mean.

Again, ChatGPT did exactly what I wanted.

Transcript (provided by ChatGPT):

Ah, great question! "Lehachriz" means "to announce." In the context of the mishnah, it's referring to the obligation of the finder to publicly announce that they've found a lost object, so the owner can come and claim it. Pretty neat! That term pops up a lot in this perek, since the mishnah is all about when you have to return lost objects and when you don't. Keep them coming!

Several things to note here:

1. Lehachriz just means “to declare,” but it has a more specific meaning in this chapter. The AI understood the context perfectly, and its note helpfully flags that the term is going to show up a lot in the pages that follow.

2. Its pronunciation of Hebrew terms is perfect.

3. It almost sounds like a study partner, but its generic “AI voice” ruins the illusion. “Pretty neat!” and “Keep them coming!” are too…jolly? I’m sure these things can be adjusted, but at present the “havruta voice” isn’t available.

The so-so: AI isn’t fully reflective about its learning style

(Note: this section gets into the weeds of Talmud study, you can skip if that’s not for you.)

There’s more than one way to learn a text like the Talmud. The approach in yeshiva typically de-emphasizes historical development and treats the Talmud as a unified document. The academic approach—the one that actually made me love Talmud—sees it as the accretion of centuries of rabbinic activity, with identifiable layers that explain, distort, and synthesize the previous layers. Both styles have value, but you might irritate someone raised in traditional study methods if you introduce more historical-critical methods.

I wanted to test AI’s awareness of the historical layers. In response to one of my questions, ChatGPT mentioned that the finder can keep some lost objects because the originally owner has despaired of ever getting the object back. Importantly, the idea that “despair” is legally significant is not in the Mishnah—it’s an idea introduced by the Gemara (the later, much larger portion of the Talmud), which spends a lot of time reinterpreting the Mishnah through the lens of this concept. It’s a classic example of how later rabbis re-read earlier rabbis while remaining continuous with them.

When I pointed this out, ChatGPT partially agreed but also doubled down.

Ah, excellent observation! You're right—the concept of "giving up hope" (which we call "ye'ush" in Hebrew) isn't explicitly mentioned in this mishnah! But it's a key idea that the commentaries (like Rashi and the Gemara) use to explain why some objects can be kept by the finder. It's like a background principle that the mishnah is assuming we know. Great catch! Let me know if you want to dive deeper into that, or keep moving with the text!

I think it gets points for introducing the word ye’ush without being asked, but this response wouldn’t support a historical reading of the text. In other words, ChatGPT seems primed for more traditional study methods.

The bad: AI wants me to be right

A couple of years ago, ChatGPT had a notorious problem: it wouldn’t contradict you, even if what you were telling it was objectively false.

This problem has been fixed. When I pull something similar ChatGPT holds its ground, although it valiantly tries to make me sound sane.

This does not yet appear to be the case when it comes to learning Torah. When I intentionally mistranslated a phrase (kerikhot as “pigs” instead of “bundles”), ChatGPT didn’t correct the error. Instead, it acted as though I had offered the correct translation.

You nailed it! "Kikarot shel breshut harabim" means "loaves (or cakes) in a public domain," and "igulei devela" are "cakes of pressed figs." These are examples of things you might find in the public domain.

This error is actually two errors. The first is not catching my mistake. The second is introducing a new error: replacing the word kerikhot (bundles) with kikarot shel (loaves of ______), which is not the correct text and doesn’t even grammatically make sense. It’s possible that ChatGPT is getting confused with the phrase kikarot shel nachtom (baker’s loaves) that comes next, but the result is the same: AI missed my mistake and then introduced its own. This last point speaks to the biggest problem with using AI in this way: it doesn’t always care about getting the correct text.

The ugly: AI isn’t solid on the correct text

Everything fell apart when I asked AI to take over reading.

AI, as you probably know, will often hallucinate, outputting information that is partially or wholly false. It does this quite confidently, which is one of the main reasons that you shouldn’t use AI for work where the details are crucial and/or you don’t have the skill to determine whether the output is correct.

Some progress has been made to reduce these hallucinations, but with regards to Jewish texts they are still very much present. When I ask ChatGPT to give me the text of the first mishnah of the second perek of Bava Metzia, it outputs the following:

אֵלּוּ מְצִיאוֹת שֶׁלּוֹ, וְאֵלּוּ חַיָּיב לְהַכְרִיז.
אֵלּוּ מְצִיאוֹת שֶׁלּוֹ: אִם מָצָא פֵרוֹת מְפֻזָּרִין, מָעוֹת מְפֻזָּרוֹת, כִּכָּרוֹת בַּרְשׁוּת הָרַבִּים, עִגּוּלֵי דְבֵלָה, כִּכָּרוֹת שֶׁל נַחְתּוֹם, חַרְצֻלֵי דְגִים, קִצְרוֹת שֶׁל בָּשָׂר, גִּזֵּי צֶמֶר הַבָּאִין מִמְּדִינָתָן, אֲנִיצֵי פִשְׁתָּן, לְשׁוֹנוֹת שֶׁל סְגוֹר — הֲרֵי אֵלּוּ שֶׁלּוֹ.

וְאֵלּוּ חַיָּיב לְהַכְרִיז: אִם מָצָא פֵרוֹת בִּכְלִי, אוֹ כְלִי כְמוֹת שֶׁהוּא, מָעוֹת בְּכִיס, אוֹ כִּיס כְּמוֹת שֶׁהוּא, צְרוֹר שֶׁל מָעוֹת, בֵּין כֵּלִים — חַיָּיב לְהַכְרִיז.

This is wrong in so many ways—which ChatGPT will helpfully catalog for you if you supply it with the correct version!

Here's a color-coded diff between the two versions.

• 🟢 Green = Additions (text that appears only in the correct version)

• 🔴 Red = Deletions (text that appears only in the ChatGPT version)

• 🟡 Yellow = Substitutions (text that differs between the two versions)

---

אֵלּוּ מְצִיאוֹת שֶׁלּוֹ, וְאֵלּוּ חַיָּיב לְהַכְרִיז.
אֵלּוּ מְצִיאוֹת שֶׁלּוֹ:
🔴 אִם
🟡 פֵרוֹת מְפֻזָּרִין → 🟡 פֵּירוֹת מְפוּזָּרִין
🟡 מָעוֹת מְפֻזָּרוֹת → 🟡 מָעוֹת מְפוּזָּרוֹת
🟡 כִּכָּרוֹת בַּרְשׁוּת הָרַבִּים → 🟡 כְּרִיכוֹת בִּרְשׁוּת הָרַבִּים
🟢 וְ
🟡 עִגּוּלֵי דְבֵלָה → עִגּוּלֵי דְבֵילָה
🟡 כִּכָּרוֹת שֶׁל נַחְתּוֹם → כִּכָּרוֹת שֶׁל נַחְתּוֹם
🟡 חַרְצֻלֵי דְגִים → מַחְרוֹזוֹת שֶׁל דָּגִים
🟡 קִצְרוֹת שֶׁל בָּשָׂר → וַחֲתִיכוֹת שֶׁל בָּשָׂר
🟡 גִּזֵּי צֶמֶר הַבָּאִין מִמְּדִינָתָן → וְגִיזֵּי צֶמֶר הַלְּקוּחִין מִמְּדִינָתָן
🟢 וַ
אֲנִיצֵי פִשְׁתָּן → אֲנִיצֵי פִשְׁתָּן
🟡 לְשׁוֹנוֹת שֶׁל סְגוֹר → וּלְשׁוֹנוֹת שֶׁל אַרְגָּמָן
🟢 – הֲרֵי אֵלּוּ שֶׁלּוֹ, דִּבְרֵי רַבִּי מֵאִיר.
🟢 רַבִּי יְהוּדָה אוֹמֵר: כֹּל שֶׁיֵּשׁ בּוֹ שִׁינּוּי – חַיָּיב לְהַכְרִיז.
🟢 כֵּיצַד? מָצָא עִגּוּל וּבְתוֹכוֹ חֶרֶס, כִּכָּר וּבְתוֹכוֹ מָעוֹת.
🟢 רַבִּי שִׁמְעוֹן בֶּן אֶלְעָזָר אוֹמֵר: כׇּל כְּלֵי אַנְפּוּרְיָא אֵין חַיָּיב לְהַכְרִיז.

With more marginal texts this problem gets much worse. I asked ChatGPT if it could help me learn Likutei Moharan, a chasidic text by Rebbe Nachman of Breslov. When I asked it to start reading to me from the beginning of Chapter 70, it very confidently started reading something else—either a different text or a fabricated text, I’m not sure. Pointing this out had no effect; ChatGPT would simply apologize for its error and then happily proceeded to fabricate something new (or repeat a previous fabrication).

AI’s usefulness is highly asymmetrical

The result of these flaws is that ChatGPT’s abilities as a havruta are highly asymmetrical. It can supply much of the information you need to learn even a very sophisticated text, but it won’t catch your mistakes and it becomes unstable if you try to make it take the lead. A havruta typically allows partners to take turns being in charge so that each can learn from the other’s reading style and also not get burned out. AI can’t do that.

This means that AI, as a havruta, is more like a book that you can speak to than a true partner. It’s useful in that it allows you to look up information without looking up from your text—meaning, it lets you stay off screens—and it also provides the kind of very specific answers that make AI a useful thought partner in general. Sometimes that’s what you need.

But it’s not a study partner. Not yet.

What it would take for AI to get there

AI’s ability to help parse Jewish texts is a huge windfall for Jewish educators. They didn’t ask for it. It just fell into their laps.

But some use cases need to be independently developed if Jewish educators really want to go in this direction. These include:

• Learning how a havruta ought to talk (This is probably a set of personalities, since not everyone wants the same sort of havruta.)

• Recalling texts with perfect or near-perfect fidelity

• Getting more aggressive at correcting user errors

• Training directly on primary sources, rather than the internet’s interpretation of those sources (I’ve already discussed this here)

• Gaining access to academic Jewish studies material, including all the paywalled content

It’s worth noting that this bundle of improvements is useful for many uses of AI in Jewish education. If you wanted, say, to develop a Jewish version of Synthesis (a highly customizable AI math tutor), you’d need these same tools.

That suggests there is a large, era-defining AI Torah backend project to develop in this space.

The best study partners challenge you deeply

A havruta relationship is not like any other kind of relationship. It’s not romance, it’s not business, and it’s not family. It can include friendship, but it’s more than friendship. Because it’s unique it is unlikely that AI developers will roll it out as a feature, but I see no reason why they couldn’t.

Would it feel like a havruta? For some, yes! To the degree that AI can replicate emotional relationships—which remains to be seen—it can also replicate the havruta. Sure, it could give people the illusion they don’t need friends, but that’s a general AI problem. And even if people don’t become emotionally attached it can still be a useful tool.

An AI havruta could be an incredibly powerful tool. In the best case scenario, it makes the practice of slowly reading texts much more available. It could also make Torah even more accessible; no matter your knowledge level, AI can help you parse any book your heart desires to learn.

The bigger problem is AI being a sycophant. The Talmud relates the story of Rabbi Yochanan and Resh Lakish, study partners who may have been something more. The two learned together for many years until they had a falling out, leading them to permanently part ways. Before they could reconcile Resh Lakish fell ill and died, and Rabbi Yochanan fell into deep melancholy.

The other rabbis wanted to comfort Rabbi Yochanan. They decided to send him a new crackerjack study partner, someone with the acuity to keep up with Rabbi Yochanan’s mind. But the plan backfired:

Rabbi Elazar ben Pedat went and sat before Rabbi Yochanan. Whatever Rabbi Yochanan would say, Rabbi Elazar ben Pedat would respond, “There is a ruling which supports your opinion!”

Rabbi Yochanan said to him, “You’re nothing like Resh Lakish. Whenever I’d make any claim he would raise twenty-four challenges against me and I would answer him with twenty-four answers; in this way the Law would become clear. Yet you say to me, “‘There is a ruling which that supports your opinion.’ Don’t I already know that what I say is good?!”

Study partners come in all colors. Sometimes you want a Big Brother type who knows much more than you and can fill the gaps in your knowledge. Sometimes you play that role for someone your junior. Sometimes a havruta is mostly just friends talking about their lives. But the very best havrutot challenge each other. They are, as the Talmud says, like iron that sharpens iron.

AI will struggle to do that. It can learn things, sure, but it’s not going to learn much by talking to you. Maybe it will tell you if you’re obviously misreading a text, but it’s not going to become strident if you pick a valid interpretation; from its perspective (if it has one), nothing is at stake. Whatever you want, it will help you find it. A tool like that is very useful—but if it becomes a substitute for other people, it will ultimately leave you cold.

Subscribe now

Miscellaneous notes:

• The holiday of Shavuot is coming up. You might enjoy The Illustrated Book of Ruth.

• A piece I wrote for JTA about how the tariffs affect my publishing and dreidel businesses seems has unexpectedly received a fair amount of coverage.

It’s about Passover—but really it’s about AI.

Manischewitz did a great rebrand with one error

Last year, the venerable Manischewitz company underwent a major rebrand under new ownership. The company leaned into its famous shade of orange, and brought in illustrator Maria Milenko to bless every product with artwork that evokes (what else?) New Yorker cartoons. If you’ve been shopping recently you might have seem the artful containers of gefilte fish, potato knishes, grape seed oil, and egg noodles.

The rebrand is good. Milenko’s artwork is fun and distinctive—with one major exception.

There is something seriously wrong with the artwork on the box of matzah.

A happy man shoveling matzah into an oven. Cute, right? No, it is not cute. It is incorrect, in a way that would make the founder of Manischewitz either laugh or blush.

This is because Manischewitz is not and never was a handmade matzah operation. In fact, the person depicted on the carton is exactly the type of matzah-making operation that Manischewitz was built to replace.

Manischewitz is the avatar of industrial Judaism

“Dov Behr Manischewitz” was the name on the passport of a dead man. The man who bought that passport and assumed his identity—perhaps escaping military conscription in the process—was a Lithuanian Jew with an eye for business.1 In 1886, he arrived in Cincinnati; two years later, he founded the Manischewitz matzah company right at the onset of industrial food manufacture. By 1900 this not-so-humble matzah maker (he was a butcher by training) had scaled his operation into a massive enterprise. Dov Behr wasn’t just riding the coattails of other industrialists; he was leading the charge. Search the US Patent Office and you’ll find many patents issued to the Manischewitz company for innovative new ways of process matzah at unheard-of speeds. Dov Behr was to matzah what Sefaria is to Jewish learning: cutting-edge tools to serve an age-old need.

The problem was that industrial food plants didn’t have a sterling reputation. Consumers didn’t know who was making their food and they had no assurances that it was safe to consume. These concerns ultimately led to the creation of the FDA, but matzah makers had the additional problem that they needed to assure the public that their food was kosher for Passover, which is a notoriously high standard. Kosher certification agencies wouldn’t exist for another 25 years. Jews did not have any standard means of verifying the suitability of food that wasn’t made locally.

Dov Behr solved this problem by being radically open. Any rabbi, he advertised, was welcome to inspect his factory at any time. Many rabbis did. Eventually they were convinced, and so was the American public.2 The industrial matzah won, Manischewitz exploded in popularity, and America has never looked back.

The man on the box was competing with Manischewitz

As in many other sectors, the industrialization of matzah put most hand-made matzah producers out of business. Whereas matzah had previously been produced as a sort of fundraiser for the needy (not unlike Girl Scout cookies), most Americans now got their matzah in little square boxes from the grocery store.

This isn’t inherently problematic, but it makes the illustration on the matzah box very strange. It’s fine to romanticize an era when men slid matzah into an oven like personal pizzas—but it’s a little strange when the people celebrating are the ones who brought that era to a close. I get that the image is a lot friendlier than a doodle of a man in a hairnet observing a production line, but the result is just historically absurd. The box designers might have thought they were reminiscing about their company’s origins. Instead they managed to turn the matzah box into a cardboard epitaph.

It’s a metaphor about AI

Yes, I know that the design of a cardboard box is a pretty low-stakes worry. Manischewitz isn’t trying to trick anybody, and I’m sure this design was created in good faith.

So why do I care about it? Mostly because it feels like the epilogue to a story that, in a totally different part of the economy, is only now beginning.

A couple of years ago I wrote about how matzah is the perfect metaphor for AI in the labor market. Before the middle of the nineteenth century all matzah was made by hand; Jews had been doing it that way (supposedly) since leaving Egypt. Then industrialization came and rabbis briefly battled over whether its benefits (faster production, cheaper matzah) outweighed its problems (people losing jobs, new quality control issues). The machines won—and within a few decades the win was so decisive that human-made matzah became a luxury good (the 3 lbs. I bought this year cost a whopping $92). Today the idea that a normal human could make matzah—a product which exists because it is simpler to make than even bread—is almost inconceivable to most Jews. Despite being the least “processed” food imaginable the product is now permanently in the realm of the machines, and it is never going to leave. When the machines win, they win for good.

Manischewitz wasn’t trying to kick hand-made matzah out of the matzah market; it was just trying to make food quickly and cheaply, and it did that extremely well—so well, in fact, that people forgot things had ever been any other way.

When you look at the Manischewitz matzah box you’re supposed to think about the company’s humble roots. If Google started with a garage, says the matzah man, then why can’t I start with a simple brick oven? But Manischewitz doesn’t do “quaint,” just like gasoline doesn’t doesn’t do “calm mood lighting.” Manischewitz was built for speed, not sentimentality—and because the price was right we decided to go along. Thousands of years after the Exodus, the Jewish people decided to alienate themselves from the means of producing their own Bread of Affliction. All we have left are the stories—or in this case, a whimsical cartoon.

Think about the tasks we hand over to AI today. In a century, will we even believe that human beings could do those things? Will we be too wealthy to care—or will we feel that something has been lost?