Photos by Sandra Singh

Last year, while taking classes on the Mishna, the first written collection of Jewish oral traditions, Antoine Leboyer (MBA 1992) wanted to dive deeper into the material. He found while there was a flood of additional texts and commentaries, there was no user-friendly, searchable way to access them.

Leboyer could have joined a havruta, the rabbinical version of a study group, to help him navigate the corpus, but as managing director of the Venture Lab for Software and AI at the Technical University of Munich, he had a more cutting-edge idea. He would use the potential of generative AI to create a “virtual havruta” and create a test case for this new technology.

Leboyer reached out to Israel’s Sefaria, the world’s largest open-source, digital library of Jewish texts, to discuss how a large language model (LLM) could enhance the collection’s user value. An LLM is a type of artificial intelligence program trained on gigantic amounts of data that can perform natural language processing tasks such as summarizing content and answering questions.

A pilot program was launched in November 2023, fed by 15 million words from three Jewish texts—the Torah, the Mishnah, and the Tanach. It was expanded to 40 million words from Sefaria’s library in December. As of August 2024, there were 222 users from 10 nations posing all kinds of questions about the Jewish faith and teachings. For instance, one user recently asked what the Talmud, the primary source of Jewish law, says about playing sports on Shabbat, the day of rest.

The platform, aptly titled Virtual Havruta, responded that while playing sports on Shabbat is not directly referenced in the Talmud, there are principles that could apply. Drawing on the philosophies of three branches of Judaism—Conservative, Orthodox, and Reform—it cited some writings that said it is permissible so long as it does not involve gambling; other texts took the stricter stance that playing sports defies the spirit of Shabbat. Virtual Havruta also provided links to reference documents for the questioner to consider and study further. The program will never replace rabbis and scholars, Leboyer says, but it can be an effective way to meet people where they are, so to speak, to re-engage lapsed Jews and enable more effective religious training.

“We developed the program so that it would reflect the way good rabbis answer,” said Leboyer. “They rarely give you one single answer for every question. They take you by the hand to make you aware of multiple possibilities, but leave much of the work to you.”

Because LLMs are known to “hallucinate”—that is, to produce inaccurate information—appliedAI, Germany’s largest AI nonprofit, which created the software, turned to advanced retrieval augmented generation (RAG) technology. Leboyer calls the tool “a good librarian.” RAG structures data to allow for the selection of the most relevant documents from a dataset.

Leboyer says RAG “makes us confident we can use such a platform in a wide group, without concerns of poor links and incorrect information. There are two parts of the technology: the retrieval of documents and its ‘logic.’” At the moment, Virtual Havruta increasingly manages to identify and retrieve all the documents one might expect in response to a question, and it is able to extract quality answers from the data.

Like any AI platform, Virtual Havruta is constantly evolving through testing and use. Developers are working to enrich the background knowledge provided to the system—including the identification of document categories, historical eras, and authors—in order to provide more thorough and intelligent responses. The ultimate plan is to incorporate Virtual Havruta into the Sefaria website and app, which currently has nearly 1 million users, offering it worldwide at no cost.

Leboyer says the process of creating Virtual Havruta offers insight into how the technology can be used in business, government, and beyond. “Data is fundamental for AI projects; the big winners of AI will be those with quality data. If it is possible to use AI to ask questions of a rabbi, as one would do in real life, think what it could mean for e-commerce, enterprise resource planning, database consultation, and more.”

Leboyer’s Venture Lab is currently using these lessons to assist Lexet.ai, a non-profit startup working to help Ukraine join the European Union, a tedious process that requires legal analysis of reams of documents. It’s a task hastened by the use of AI, which might conceivably save lives by saving time. By joining the EU, the nation could benefit from mutual assistance in the face of military aggression.

“We are faced with a parallel situation of incoherent unstructured data—just like the Torah and its commentaries,” Leboyer says. “This project started with Torah study and could end up helping Ukraine.”

Post a Comment