Jungwon Byun

Jungwon Byun is the cofounder and COO of Elicit, an AI research assistant, alongside cofounder and CEO Andreas Stuhlmüller. Her public explanations emphasize supervising the process of AI research: breaking expert work into inspectable steps so people can evaluate answers and locate failures.[3][1]

In her April 2024 Latent Space interview, Byun described Elicit’s approach as identifying the granular steps an expert human would take and training AI systems to perform each step robustly. She argued that this makes evaluation and troubleshooting easier: when an answer fails, the team can investigate where the process broke down. Her ambition was for models to progress from novice to expert performance, but to “earn their right to the expertise,” with explicit processes that people could audit even if AI eventually surpassed human researchers.[1]

Byun expected people’s work to shift increasingly toward evaluating model output. She described citations and highlighted passages that let users check an answer against its supporting text, alongside uncertainty flags that prioritize questionable extractions for review. She said the uncertainty estimates were reasonably calibrated in Elicit’s experience, while qualifying that this varied by model.[1]

Longer context windows changed the evaluation problem. Byun explained that when a system first retrieves a small passage, it already knows which text was supplied to answer a question. When the model receives an entire paper, the system must locate supporting passages afterward. She characterized this as a shift from breaking down tasks toward breaking down evaluation; users still needed quotations they could inspect.[1]

In her February 20, 2025 announcement of Elicit Systematic Review, Byun presented a workflow spanning search, title and abstract screening, and full-text data extraction. Users could edit AI suggestions, override screening decisions, inspect supporting quotations and explanations, and revisit earlier steps. She also preserved an important qualification: publication or regulatory-grade reviews would often require traditional keyword searches, with Elicit’s search used as a supplement. Her framing was to teach AI an expert human process while leaving users in control.[6]

These workflows followed earlier work at Ought, where Byun explored how people could evaluate reasoning that exceeded their own access to information. The Metascience 2021 speaker profile identified her as Ought’s COO and cofounder and Elicit as its core project.[1][4]

Before Elicit, Byun was Head of Growth at Upstart and a management consultant at Oliver Wyman. She also served as Executive Director/CEO of Elmseed Enterprise Fund and graduated cum laude from Yale with a B.A. in economics.[3][4]

Byun recalled becoming interested in AI-enabled mental-health support around 2015 after helping a close friend who lacked support. She thought AI might make such help more widely available, but felt neither she nor the technology was ready for a company. After working in fintech, she decided in 2019 to build something herself. Ought’s approach appealed to her because it broke complex thinking into smaller pieces and addressed whether powerful reasoning systems could be trusted and evaluated.[1]

A close mutual friend introduced Byun to Stuhlmüller. Byun said she sought complementary skills and shared values and recognized an affinity with his writing. Stuhlmüller recalled that their mutual evaluation produced roughly 50 pages of questions and exchanges about values and decisions; Byun said she shared all her past performance reviews.[1]

At Ought, Byun described experiments in which people stood in for more capable AI systems: one participant received less text and had to evaluate the work of another who could read much more. The team used such experiments to explore task decomposition and process supervision. Byun explained the subsequent move toward products as a more direct way to affect people’s lives, since published research still required someone else to implement it.[1]

Byun recalled that Elicit’s first version was a forecasting assistant intended to turn fuzzy beliefs into probability distributions. The team concluded that the larger obstacle to useful predictions was research, rather than probability modeling, and moved toward a general research assistant. She viewed GPT-3 as the point when AI could support the tools they wanted to build; GPT-4 improved accuracy and enabled harder questions after the product’s basic shape was established.[1]

In her February 2024 essay on weekly launches, Byun described hesitating over literature review because the problem seemed vast and poorly defined, while finding experts already offered a clearer use case. Repeated requests from signups pushed the team toward literature review, and Stuhlmüller built a GPT-3 abstract-summary prototype over a weekend. Byun’s response was an operating constraint: launch a user-facing feature every week and aim to double user numbers each month, so the team would receive feedback quickly rather than spend months building the wrong solution. The first literature-review launch was August 31, 2021; her essay dates the weekly commitment to September 2021.[5][1]

Byun’s account also described the limits of that cadence. She credited Head of Engineering James with lessons in conservative weekly planning: one main focus and commitments the team was confident it could finish. Larger projects needed to be divided across weeks rather than compressed into repeated emergency launches. She acknowledged that Elicit had underinvested in monitoring whether users discovered and used features and still needed better planning across multiple weeks.[5]

The notebook interface reflected a longer search for composable research workflows. In April 2024, Byun called it the team’s fourth attempt, recalling that an early branching system for decomposing research questions had been overwhelming. Computational notebooks offered flexible workflows with stable building blocks, allowing researchers to extend an analysis step by step. Stuhlmüller explained the reusable-process goal: debug an analysis on a few papers, then run it over a much larger collection.[1]

Byun described Elicit’s public benefit corporation structure as an explicit commitment to choosing commercial opportunities aligned with its mission. In the April 2024 interview, she said, “We're going to take the version of making a lot of money that's in line with our mission.” She linked business success to helping users discover truth and conduct research more rigorously.[1]