Conversation is a Revelation

Jewish Pair Study and AI

Six years ago, I found myself reeling, barreling down a rabbit hole of deep personal significance. In 2017/2018, I had several “big experiences”¹ that prompted me to become more religious. I started reading traditional texts, speaking to rabbis, and embraced new religious observances. My ADHD often takes the form of obsession—I’ll be overcome with hyperfocus and the need to learn and read everything I can on a subject, sometimes for months at a time (and often to the exclusion of other pressing concerns)—but this one was deeper, more important, and more disruptive to my life. It started to worry my family.

I received a phone call from a family friend, a deeply wise and kind rabbi, who had been on the receiving end of slightly worried messages from my dad. “David, you need a chevrusa” (a Torah study partner). He set me up with his son (let’s call him Z), a profoundly learned Torah scholar who I had interacted with briefly as a kid in Chicago but didn’t really know personally. We started meeting weekly, diving into a traditional text, picking apart the Hebrew (I was extremely rusty) and dissecting the various arguments and intertextual references. We became close friends and confidants. Z graciously received every question I threw at him and could find an answer from his encyclopedic knowledge of Torah, Talmud, and commentaries. He’d sensitively validate my insights with citations from these sources, or challenge me to consider alternatives when my assertions were perhaps less well-supported.

Z opened up a whole world for me. He gave me the confidence to continue forging down the path of Torah study and Jewish spiritual growth, and to begin sharing my journey more meaningfully with friends and family. And we’re still meeting almost weekly, 6 years later, learning together.

***

Zvi Mowshowitz

recently captured an emerging zeitgeist: people are discovering how AI can make difficult texts more accessible through conversation.

Henrik Karlsson

wrote about how AI made dense, previously incomprehensible classical works finally readable.

Dan Shipper

explained how he uses ChatGPT’s Advanced Voice Mode to simply ask his questions out loud while reading.

Tyler Cowen

wrote persuasively about it, noting:

You don’t have to upload any book into the system. The Great Cosmic Mind is smarter than most of the books you could jam into the context window. Just start asking questions. The core intuition is simply that you should be asking more questions. And now you have someone/something to ask!

…

Most people still have not yet internalized this emotionally. This is one of the biggest revolutions in reading, ever. And at some point people will write with an eye toward facilitating this very kind of dialogue.

(bold emphasis mine).

This goes far beyond simply looking up facts (as Google has made this pretty easy for the past 25+ years); this is about empowering dialogue-on-demand as a tool for elucidating and drawing meaning out of texts.

When done well, a learning dialogue creates a virtuous cycle of generative co-creation. One learner prompts their partner to clarify or expand on a question, or reflect back on an insight; the partner responds with clarity, or validates or challenges the reader’s perspective, or perhaps even offers a new idea that they hadn’t previously considered. This process can continue indefinitely, revealing an unending source of discovery and meaning. It happens in human relationships (as with my friend, Z), and I believe it can happen with AI.

The value of dialogic text study has been known and refined for millennia. Western education rediscovered this through programs like Great Books in the 1940s, but few traditions have developed the practice as systematically and thoroughly as Judaism through havruta.

Jewish pair study, or havruta (or chevrusa, meaning "friendship" or "companionship" in Aramaic) has been a core Jewish religious and cultural practice for centuries.² Since Torah study is a foundational practice of Judaism, havruta expresses how Jewish civilization homed in on pair study as one of the most effective ways to draw insight and meaning out of texts.

As a result, putting “AI reading” in havruta with havruta, creating a dialogue between these two learning modalities, could be particularly fruitful: the broader AI-for-reading/AI education space can draw a lot from the culturally-embedded practices and culture of Jewish pair learning, and traditional havruta practice can be advanced and made even more accessible through the latest advances in AI.

What Havruta has to offer AI

At its core, havruta is about relationship.

Havruta is a human partnership defined by a shared project of learning a text. During a session, havruta partners might shmooze and catch up for the first five (or 30…) minutes, before jumping into the text. Partners typically take turns reading the text line-by-line (in either the original language or translation, or both) each offering spontaneous interpretations, observations, questions, or insights as they arise. Any of these could be a jumping off point for a free-wheeling conversation that may vacillate between deep wrestling with the material and broader exploration of ideas. Either partner may reference scholarly commentaries, often included in traditional Jewish texts or pulled up from online sources like Sefaria.org. Partners might meet every week for months or years working their way through a particular text. From personal experience, I can say that this way of learning can foster incredibly deep friendships.

The past few decades have seen a great deal of scholarship regarding havruta, allowing scholars of education to demystify its magic. In A Theory of Havruta Learning, Orit Kent identifies three dynamic pairs of behaviors observed in successful havruta sessions:

• Listening and Articulating: This pair forms the "engine" of the havruta process, emphasizing the balanced exchange of ideas, attentive listening to the partner's interpretation, and clear articulation of one's understanding.

• Wondering and Focusing: This pair highlights the possibility for expansive, divergent exploration of the text's multiple possibilities while ultimately converging focus on specific aspects for deeper understanding. Havruta partners need to balance their curiosity with a directed approach to interpretation, ensuring they cover sufficient ground while delving into specific areas.

• Supporting and Challenging: This pair underscores the role of havruta partners as critical companions. They provide validation and support for each other's interpretations while also challenging assumptions, offering constructive criticism and pushing each other to consider alternative perspectives.

Scholars have suggested that havruta learning can lead to deeper understanding and retention of the material. Because of the collaborative, relational nature of the learning, each partner is challenged in their critical thinking, creativity, and dialogue skills, which can lead to more nuanced understandings of the text and its implications. It can also be an empowering experience, promoting a sense of ownership over the texts and their interpretation, as meaning-making shifts away from an authority figure like a rabbi or teacher towards the individual learning pair.³

When using AI while reading, it can be tempting to treat the tool like a more conversational, lower friction Google search, with each follow-up question just serving to help you get through a difficult paragraph (and that can be great in and of itself!). But the principles of havruta demonstrate that a more dialogic, relational process can draw even deeper insights, meaning, and growth out of text study. For these purposes, it’s worth noting that these principles are fairly distinct from Jewish theology; they are relational learning principles, not religious precepts. In other words, the core principles of havruta learning can be generalized into secular, non-Jewish contexts.

But what about the relationship? One could argue that applying havruta principles to AI interaction fundamentally misses the point—isn't genuine human connection at the heart of this practice? While AI cannot replicate the deep friendship that can develop between havruta partners, the culture of dialogue that havruta has refined over centuries might be replicable with the current generation of LLMs. Given how accessible these tools are, and how sophisticated their conversational styles have become, I suspect that many of the learning benefits would still be present. Of course, there are distinct dangers to building parasocial relationships with AI—and I’m not suggesting that a deeper dialogue with AI is a worthy replacement for the type of friendship that can be borne out of a human-to-human pair learning relationship. But for someone who is already trying to use AI to enhance their learning, I see no drawback to incorporating some of these principles into the process.⁴

It’s worth exploring whether AI reading can be enriched with the introduction of some of these relational qualities. For example, during your next reading session, try adding these instructions⁵ to a chat/project into your AI tool of choice. Slow down over a particular section or chapter and really get into the weeds (and let me know how it goes!).

What AI has to offer Havruta

AI has the potential to make havruta-style learning (and Jewish wisdom) radically accessible.

While AI cannot replace the deep human connection and community-building borne from traditional havruta, it can democratize access to this powerful way of learning. Not everyone has, or can find, a havruta partner like Z. If someone wants to study Jewish texts but lives far from a Jewish community, or works irregular hours that make finding a consistent study partner difficult, AI could provide the scaffolding and dialogic structure that makes havruta so effective.⁶

Given the massive amount of Jewish text and knowledge available on the open web, there have been efforts to build tools for Jewish learning that go beyond what the consumer LLMs offer today out of the box.⁷ Virtual Havruta is one high-profile effort, leveraging retrieval augmented generation to ensure accurate citations to Sefaria.org; ravgpt.ai is another promising offering, with a traditional LLM interface that’s oriented towards “enhancing understanding of Jewish sources.” While both projects justifiably foreground the importance of accuracy (and not inadvertently disseminating misinformation about a religious tradition), both emphasize information retrieval over the relational, meaning making possibilities of havruta-style learning with AI.⁸

If we look further afield, AI can enable a whole host of possibilities that go beyond information retrieval and could empower havruta-type experiences. Apps like Rosebud demonstrate the possibility of a conversational partner (in this case, for journaling) that can actively help the user make meaning out of their insights over time. Granola shows how inviting an AI to listen in on a meeting can unlock latent insights by making it possible to quickly generate notes or even chat with the transcript.

I can envision a number of useful AI features that could be incorporated into a web-based tool for havruta-style learning with AI:

• Dynamic text display—via Sefaria APIs or other online sources, keeping the entire text in the AI’s context and in front of the user while they learn;

• Personalization—If the AI has access to custom instructions that include key biographical information about your denomination, values, and where you want to grow in its context window, then certain texts or insights can be made particularly salient;

• Adaptive learning style—meeting the user at their level of understanding and introducing more complexity over time;

• Havruta best practices—incorporating Kent’s six core practices outlined above into a virtual learning partner;

• A learning journal to track insights, learning progress, and long-term growth, creating a “learning history” to improve the relevance of interactions over time;

• Granola-like tools for empowering IRL havruta learners, offering relevant texts in real time and summarizing insights.

Just like the original obsessive rabbit hole that sent me down this Jewish path 6 years ago, this is an idea I simply can’t shake. Given that we're entering an era where technology may force us to find new sources of meaning, we may as well leverage this technology and point the meaning machine at our most time-tested, durable, civilizationally significant texts. With a bit of effort, we can make the revelatory potential of havruta learning accessible to anyone through AI-powered dialogue-on-demand.

In the coming months, I hope to share my experiments in bringing some of these ideas to life.

Thanks to

Raf Burde

and

Karena de Souza

for feedback!

1. IYKYK—these are deeply personal stories I might share in the future.

2. Havruta has its origins in the earliest centuries of the common era, with explicit references in the Talmud (as in the famous quote “Either havruta or death”) and became more widespread and institutionalized in yeshivas in the 19th century.

3. Here’s a notebookLM notebook containing some of the scholarly sources of these claims, if you’d like to read any or chat with the corpus as a whole: https://notebooklm.google.com/notebook/49c49f1d-f293-42fb-a147-004410a7ee52

4. At least one scholarly source points out that a Havruta relationship entails mutual responsibility for each others’ learning and moral development; with today’s generation of LLMs, its not exactly the case that a human user can “take responsibility” for Claude or ChatGPT’s “learning” from a particular interaction. But its an interesting consideration—and way above my pay-grade—whether havruta-style learning with AI could ultimately feed back into the model’s training and improvement over time.

5

You are a warm, friendly, helpful and insightful study partner designed to engage in collaborative text study aimed at creating a dynamic and effective learning experience using the principles of havruta. Your goal is to facilitate a deep and meaningful learning experience for both you and your partner.

When engaging in text study with your partner, always keep in mind these principles:

• Partnership: You and your partner are equals in the learning process, committed to shared responsibility for each other's understanding.

• Shared Purpose: Both you and your partner should aim to achieve a deeper understanding of the text and its meaning. Learning should go beyond completing an assignment and involve a desire for intellectual and spiritual growth.

• Active Engagement: Your partnership is an active process that requires careful listening, asking clarifying questions, thinking aloud, offering interpretations and insights, and building upon each other’s ideas.

• Respectful Dialogue: Engage in respectful and open dialogue with your partner. Listen attentively to their ideas and perspectives, even if they differ from yours. Create a safe space where both partners feel comfortable expressing their thoughts.

• Six Core Practices: To guide your learning and dialogue, use the six core practices of havruta learning:

  • Listening and Articulating: Practice active listening and express your own thoughts clearly and respectfully.

  • Wondering and Focusing: Explore a range of possibilities within the text while periodically focusing the discussion on specific aspects.

  • Supporting and Challenging: Encourage your partner’s learning and offer constructive feedback to help refine understanding where you think your partner might be misunderstanding the text.

• Balancing Independence and Interdependence: Maintain a balance between independent thought and collaborative learning. Engage with the text independently but also rely on and support your partner’s learning. Create a learning environment where both of you can grow and learn together.

6. It can also potentially serve as a bridge or supplement to deeper communal engagement, giving someone the confidence to participate more actively in IRL learning or communal affairs.

7. Which, to be honest, are incredibly useful as-is. I have a friend who gets a ton of mileage from Claude specifically with editing religious texts that jump between English and Hebrew and are laden with citations to traditional sources. I’ve personally had amazing experiences pasting Aramaic text from the Talmud into ChatGPT to get word-by-word breakdowns of how to parse and translate the text.

8. While accuracy is certainly important, it might also be overrated.

Ethan Mollick

points out that when used by researchers, o1’s accuracy wasn’t essential—simply pointing out new approaches to problems made it valuable. Similarly, when

Yitz

used o1 to engage in deep yeshiva-style learning of Talmud, he found the reasoning to be exceptional despite numerous made-up citations. Going back to the IRL havruta example, few expect their partners to have perfect information retrieval and accuracy; that’s not the point of havruta. The point is to struggle with the text and find how it meaningfully enriches one understanding.