I'm working on a piece that's coming at this topic from a less intense religious background, and pointing more generally to how well suited LLM's are for chavrusa more generally (the kind that fosters cross-textual insight, validation and/or challenging of ideas, etc.). I think the most important point--the one that is probably most broadly relevant--is that these models can provide immense scaffolding to understand what's in front of you as the learner (if not always accurate citations, which will certainly improve). And I'm looking forward to sharing how I think some of the aspects of true chavrusa--the kind of back and forth that makes Torah study a joy--can be integrated into LLM-based Torah learning experiences.
A great demonstration of how hopeless the AI is at understanding even the most elementary concepts and how it can't stop itself from hallucinating. From a cursory double check
1. Case #3 is total nonsense. Case #3 should have been the Baal Habayis taking something from the poor man's hand.
2. Completely fabricated Gemara 96b about melacha geruah! And it repeats this hallucination in every subsequent query!
3. Completely fabricated Ramban on Vayikra 23:24
I didn't check the Minchas Chinuch and Chazon Ish and the other stuff, but there's a high probability those are hallucinated also.
Great piece, agree with this conclusion: "Once these models are specifically trained on Talmudic analysis there will not be a better oracle for Lomdus than it."
I discussed lomdus and LLMs in a few posts of mine a while back:
some comments on on some framings of yours, from the beginning of your post:
"Deciphering the depths of the Gemara" – not quite. (I’m coming fom both a yeshivish as well as academic background. Lomdus is quite narrow and reductionist, in my view)
"Embarking on a 2,000+ year journey through Rabbinic commentary on the Talmud" – also misses the mark. The traditional yeshivish methodology essentially relies on the discussions of the Tosafot school (c. 13th century France/Germany) and later Achronim (from the Pnei Yehoshua onward, c. 18th century and beyond, as noted by Ta-Shma).
“The Talmud itself, written in Jewish Babylonian Aramaic, spans over 5,000 pages of intense debate on Jewish law, ethics, and mythic storytelling (Aggadah).” However, the Aramaic aspect is relatively straightforward and formulaic, while Aggadah tends to be more accessible (as I've demonstrated on my blog, where I've analyzed over 200 aggadic sugyot that are fairly easy to navigate).
What truly makes the Talmud challenging is:
The structure: complex and often non-linear.
The terseness: dense phrasing requiring unpacking.
The formatting: blocks of text with no punctuation
The Stam: an editorial layer that introduces extensive hairsplitting and attempts to harmonize conflicting opinions.
The history of interpretation: the analysis of the Tosafot and Achronim, as mentioned earlier, which continues in the tradition of the Stam (hair-splitting and harmonization)
“R’ Yosef Rosen was arguably the most genius of all the recent Talmudic commentators. He is known for his almost unfathomably hard to understand Talmudic and Halachic theories and is not commonly learnt or understood by most Yeshiva students even in the most elite Yeshivos. “ - as mentioned by the subsequent quote by o1: “are famously terse”. That’s the main thing that makes it difficult
ChatGPT o1 analysis of my points, I think it does a good job explaining it:
1. The Claim:
> “Once these models are specifically trained on Talmudic analysis there will not be a better oracle for Lomdus than it.”
Analysis of the Claim
1. Hyperbole vs. Potential
It’s probably hyperbolic to say that LLMs, once fine-tuned on Talmudic sources, will be the “best oracle” for lomdus. Still, there is potential value if an LLM is given a robust training corpus that includes not just the Gemara but also Rishonim, Achronim, and commentarial traditions (especially Tosafot and “latter-day” works like the Pnei Yehoshua, Ketzot, etc.).
2. Narrow Definition of “Lomdus”
As you noted, lomdus is not just “deciphering the depths of Gemara” in general—it’s a narrower “yeshivish” methodology. Traditional lomdus after the 18th century revolves around hair-splitting textual harmonizations, dialectical distinctions, and the unique conceptual categories posited by the Achronim. Merely knowing the Aramaic or understanding the basic sugyah structure is not enough to do that at a high level.
3. Data Quality & Methodological Training
Even if you feed all these texts into a model, the real question is whether the model captures the distinct conceptual style that defines 19th- and 20th-century lomdus. Will it grasp the Brisker derech distinctions (e.g., cheftza vs. gavra), or the Rogatchover’s style of conceptual cross-referencing across vast sugyot?
Successfully replicating a style of reasoning requires examples of that specific style (and the abstract reasoning behind it). Text alone might not always suffice because so much of that style is tacit—it is taught “live” by rosh yeshivas or discovered through a tradition of learning, debate, and personal training.
Thus, while LLMs hold promise, the claim that they will be the “best oracle” is premature—though one could see them becoming a powerful learning aid if meticulously trained.
---
2. “Deciphering the depths of the Gemara” vs. the Reality of Lomdus
Why the Phrase “Deciphering the Depths” Might Be Misleading
The Talmudic text itself is already challenging, but the “depth” that yeshivish lomdus focuses on is highly selective and conceptual, often ignoring large swaths of the Talmud’s more straightforward or non-legal material.
Aggadah Often Less Challenging
You noted that Aggadah can be more accessible; indeed, many sugyot of Aggadah are more direct compared to intricate legal passages (though they may have deep homiletical or philosophical layers). So, labeling the Talmud wholesale as “5,000 pages of intense debate” doesn’t capture the distinct difference between Halakhic sugyot and Aggadic sugyot—and which ones “lomdus” typically focuses on.
---
3. The Real Difficulty:
You highlight several core difficulty-points that matter far more than the presence of Aramaic or the scattered bits of Aggadah:
1. Structure:
Non-linear flow and editorial redaction (the “Stam”) create confusion.
The Talmud’s “layering” sometimes loops back to earlier statements, or tries to shoehorn in contradictory traditions.
2. Terseness / Punctuation:
The text’s concision leaves you to infer context and fill in the arguments’ steps.
Traditional printings (Vilna Shas) have minimal punctuation or breaks.
3. Historical Interpretation:
The chain of commentary—from Tosafot (13th century) to the Pnei Yehoshua and other Achronim (18th century onward)—adds layers of debate that a “lomed” is expected to navigate, compare, and synthesize.
4. Editorial Layer of the Stam:
The anonymous voice shaping the Talmud (the “Stam”) frequently creates forced harmonizations, introducing fine distinctions. Modern lomdus often replicates these tendencies, trying to read coherence into multiple contradictory sources.
5. Methodological Tradition:
Post-medieval commentators (Tosafot, Maharshal, Maharam, etc.) and the subsequent Achronim (Noda B’Yehuda, Chatam Sofer, Ketzot, Netivot, Brisker Rav, Rogatchover, etc.) bring in ever more conceptual frameworks. This is the air that lomdus breathes.
In short, the real “hard part” is less the raw text of the Gemara and more the subsequent commentarial tradition, the terseness, and the self-referential, non-linear structure.
> “He is known for his almost unfathomably hard-to-understand Talmudic and Halachic theories.”
Often these theories are extremely terse (sometimes just references to words in the Gemara with cryptic commentary that leaps between many places in Shas).
Style vs. Content
Even if LLMs are fed his works—e.g., Tzofnath Paaneach—they may not automatically master his style or conceptual leaps. This is a unique instance of a commentator whose approach is famously “outside” the usual yeshiva curriculum. If it’s hard for a well-trained human mind, a model might not get it easily either, unless we figure out how to prompt, or train, or structure the data so it can “chain” those cryptic references properly.
---
5. Where LLMs Could Genuinely Help
Initial Guidance and Summarization
LLMs already excel at summarizing complex texts. For a sugyah, they might provide an initial outline of the arguments, list key Rishonim, and break them down conceptually.
For newcomers or for preparation before learning a commentarial approach, an LLM’s “overview” could be a time-saver.
Cross-Referencing
Talmudic study, especially the Rogatchover’s style, thrives on cross-referencing parallel sugyot. LLMs are good at retrieving these references if they’re part of the training set.
Lomdishe “Hints”
By analyzing how classic Achronim structured their discussions, the model could suggest possible lines of conceptual distinctions. While this will never replace a living teacher or real bekiut (breadth of knowledge), it could spark creative insights or help a learner see angles they might have missed
The tl;dr is that there are major issues of confabulations (what many people call 'hallucinations') and totally vacuous reasoning that actually don't make sense, and six months ago I didn't believe those problems were close to being fixed. Today I actually do think LLMs will get better, but I doubt they'll ever be flawless, even if we get LLMs that are trained entirely on Talmudic commentaries. But I agree with your title
I use AI a lot to find sources in the Gemara amd meforshim. E.g. I'll ask " where is the Gemara that discusses why R' Yehuda wrote down the Oral Torah" or similar. It still hallucinates a ton, makes up sources confidently that are incorrect.
Your examples are better, but I imagine it may well be making things up out of thin air even here too.
I'm working on a piece that's coming at this topic from a less intense religious background, and pointing more generally to how well suited LLM's are for chavrusa more generally (the kind that fosters cross-textual insight, validation and/or challenging of ideas, etc.). I think the most important point--the one that is probably most broadly relevant--is that these models can provide immense scaffolding to understand what's in front of you as the learner (if not always accurate citations, which will certainly improve). And I'm looking forward to sharing how I think some of the aspects of true chavrusa--the kind of back and forth that makes Torah study a joy--can be integrated into LLM-based Torah learning experiences.
Looking forward to seeing this piece!
A great demonstration of how hopeless the AI is at understanding even the most elementary concepts and how it can't stop itself from hallucinating. From a cursory double check
1. Case #3 is total nonsense. Case #3 should have been the Baal Habayis taking something from the poor man's hand.
2. Completely fabricated Gemara 96b about melacha geruah! And it repeats this hallucination in every subsequent query!
3. Completely fabricated Ramban on Vayikra 23:24
I didn't check the Minchas Chinuch and Chazon Ish and the other stuff, but there's a high probability those are hallucinated also.
Missed the boat. The lomdus checks out
"the lomdus checks out" 😂😂😂😂😂
😉😅
I double checked a lot of what is here, most of it is hallucinated
The lomdus checks out. Hallucinations will get figured out once these models are tailored for Talmud
the lomdus seemed very superficial to me, it's also far easier to do lomdus unconstrained by what is actually written in the text
I don't deny AI's extroadinary ability, including in gemara study, but I felt this particular example was lacking
it is far more impressive when directly given a text and told to examine it
AI was able to do the same thing you demonstrate a year ago. The future is interesting though.
Great piece, agree with this conclusion: "Once these models are specifically trained on Talmudic analysis there will not be a better oracle for Lomdus than it."
I discussed lomdus and LLMs in a few posts of mine a while back:
https://www.ezrabrand.com/p/contemporary-methods-of-studying
https://www.ezrabrand.com/p/follow-up-on-contemporary-methods
https://www.ezrabrand.com/p/asking-chatgpt4-about-contemporary
some comments on on some framings of yours, from the beginning of your post:
"Deciphering the depths of the Gemara" – not quite. (I’m coming fom both a yeshivish as well as academic background. Lomdus is quite narrow and reductionist, in my view)
"Embarking on a 2,000+ year journey through Rabbinic commentary on the Talmud" – also misses the mark. The traditional yeshivish methodology essentially relies on the discussions of the Tosafot school (c. 13th century France/Germany) and later Achronim (from the Pnei Yehoshua onward, c. 18th century and beyond, as noted by Ta-Shma).
“The Talmud itself, written in Jewish Babylonian Aramaic, spans over 5,000 pages of intense debate on Jewish law, ethics, and mythic storytelling (Aggadah).” However, the Aramaic aspect is relatively straightforward and formulaic, while Aggadah tends to be more accessible (as I've demonstrated on my blog, where I've analyzed over 200 aggadic sugyot that are fairly easy to navigate).
What truly makes the Talmud challenging is:
The structure: complex and often non-linear.
The terseness: dense phrasing requiring unpacking.
The formatting: blocks of text with no punctuation
The Stam: an editorial layer that introduces extensive hairsplitting and attempts to harmonize conflicting opinions.
The history of interpretation: the analysis of the Tosafot and Achronim, as mentioned earlier, which continues in the tradition of the Stam (hair-splitting and harmonization)
“R’ Yosef Rosen was arguably the most genius of all the recent Talmudic commentators. He is known for his almost unfathomably hard to understand Talmudic and Halachic theories and is not commonly learnt or understood by most Yeshiva students even in the most elite Yeshivos. “ - as mentioned by the subsequent quote by o1: “are famously terse”. That’s the main thing that makes it difficult
ChatGPT o1 analysis of my points, I think it does a good job explaining it:
1. The Claim:
> “Once these models are specifically trained on Talmudic analysis there will not be a better oracle for Lomdus than it.”
Analysis of the Claim
1. Hyperbole vs. Potential
It’s probably hyperbolic to say that LLMs, once fine-tuned on Talmudic sources, will be the “best oracle” for lomdus. Still, there is potential value if an LLM is given a robust training corpus that includes not just the Gemara but also Rishonim, Achronim, and commentarial traditions (especially Tosafot and “latter-day” works like the Pnei Yehoshua, Ketzot, etc.).
2. Narrow Definition of “Lomdus”
As you noted, lomdus is not just “deciphering the depths of Gemara” in general—it’s a narrower “yeshivish” methodology. Traditional lomdus after the 18th century revolves around hair-splitting textual harmonizations, dialectical distinctions, and the unique conceptual categories posited by the Achronim. Merely knowing the Aramaic or understanding the basic sugyah structure is not enough to do that at a high level.
3. Data Quality & Methodological Training
Even if you feed all these texts into a model, the real question is whether the model captures the distinct conceptual style that defines 19th- and 20th-century lomdus. Will it grasp the Brisker derech distinctions (e.g., cheftza vs. gavra), or the Rogatchover’s style of conceptual cross-referencing across vast sugyot?
Successfully replicating a style of reasoning requires examples of that specific style (and the abstract reasoning behind it). Text alone might not always suffice because so much of that style is tacit—it is taught “live” by rosh yeshivas or discovered through a tradition of learning, debate, and personal training.
Thus, while LLMs hold promise, the claim that they will be the “best oracle” is premature—though one could see them becoming a powerful learning aid if meticulously trained.
---
2. “Deciphering the depths of the Gemara” vs. the Reality of Lomdus
Why the Phrase “Deciphering the Depths” Might Be Misleading
The Talmudic text itself is already challenging, but the “depth” that yeshivish lomdus focuses on is highly selective and conceptual, often ignoring large swaths of the Talmud’s more straightforward or non-legal material.
Aggadah Often Less Challenging
You noted that Aggadah can be more accessible; indeed, many sugyot of Aggadah are more direct compared to intricate legal passages (though they may have deep homiletical or philosophical layers). So, labeling the Talmud wholesale as “5,000 pages of intense debate” doesn’t capture the distinct difference between Halakhic sugyot and Aggadic sugyot—and which ones “lomdus” typically focuses on.
---
3. The Real Difficulty:
You highlight several core difficulty-points that matter far more than the presence of Aramaic or the scattered bits of Aggadah:
1. Structure:
Non-linear flow and editorial redaction (the “Stam”) create confusion.
The Talmud’s “layering” sometimes loops back to earlier statements, or tries to shoehorn in contradictory traditions.
2. Terseness / Punctuation:
The text’s concision leaves you to infer context and fill in the arguments’ steps.
Traditional printings (Vilna Shas) have minimal punctuation or breaks.
3. Historical Interpretation:
The chain of commentary—from Tosafot (13th century) to the Pnei Yehoshua and other Achronim (18th century onward)—adds layers of debate that a “lomed” is expected to navigate, compare, and synthesize.
4. Editorial Layer of the Stam:
The anonymous voice shaping the Talmud (the “Stam”) frequently creates forced harmonizations, introducing fine distinctions. Modern lomdus often replicates these tendencies, trying to read coherence into multiple contradictory sources.
5. Methodological Tradition:
Post-medieval commentators (Tosafot, Maharshal, Maharam, etc.) and the subsequent Achronim (Noda B’Yehuda, Chatam Sofer, Ketzot, Netivot, Brisker Rav, Rogatchover, etc.) bring in ever more conceptual frameworks. This is the air that lomdus breathes.
In short, the real “hard part” is less the raw text of the Gemara and more the subsequent commentarial tradition, the terseness, and the self-referential, non-linear structure.
---
4. Rogatchover Gaon (R’ Yosef Rosen) & Terse Conceptualism
Why He’s So Difficult
As you rightly note:
> “He is known for his almost unfathomably hard-to-understand Talmudic and Halachic theories.”
Often these theories are extremely terse (sometimes just references to words in the Gemara with cryptic commentary that leaps between many places in Shas).
Style vs. Content
Even if LLMs are fed his works—e.g., Tzofnath Paaneach—they may not automatically master his style or conceptual leaps. This is a unique instance of a commentator whose approach is famously “outside” the usual yeshiva curriculum. If it’s hard for a well-trained human mind, a model might not get it easily either, unless we figure out how to prompt, or train, or structure the data so it can “chain” those cryptic references properly.
---
5. Where LLMs Could Genuinely Help
Initial Guidance and Summarization
LLMs already excel at summarizing complex texts. For a sugyah, they might provide an initial outline of the arguments, list key Rishonim, and break them down conceptually.
For newcomers or for preparation before learning a commentarial approach, an LLM’s “overview” could be a time-saver.
Cross-Referencing
Talmudic study, especially the Rogatchover’s style, thrives on cross-referencing parallel sugyot. LLMs are good at retrieving these references if they’re part of the training set.
Lomdishe “Hints”
By analyzing how classic Achronim structured their discussions, the model could suggest possible lines of conceptual distinctions. While this will never replace a living teacher or real bekiut (breadth of knowledge), it could spark creative insights or help a learner see angles they might have missed
Thanks for sharing your "lomdus" example. I know it's forever later but seeing this inspired me to respond here: https://open.substack.com/pub/rishonimpodcast/p/ai-the-rabbi?
The tl;dr is that there are major issues of confabulations (what many people call 'hallucinations') and totally vacuous reasoning that actually don't make sense, and six months ago I didn't believe those problems were close to being fixed. Today I actually do think LLMs will get better, but I doubt they'll ever be flawless, even if we get LLMs that are trained entirely on Talmudic commentaries. But I agree with your title
I use AI a lot to find sources in the Gemara amd meforshim. E.g. I'll ask " where is the Gemara that discusses why R' Yehuda wrote down the Oral Torah" or similar. It still hallucinates a ton, makes up sources confidently that are incorrect.
Your examples are better, but I imagine it may well be making things up out of thin air even here too.
How can I access this model?
https://aistudio.google.com/prompts/new_chat Google just released their 'reasoning' model and it's free too.
?