Jacob Kahn

Timeline

1. 2015–2018

   Computer science at Penn

   Raised outside Chicago, he studies computer science at Penn Engineering and statistics and operations research at the Wharton School through the University of Pennsylvania's M&T Program. While at Penn, he runs PennApps, then the largest collegiate hackathon in the world. From 2015 to 2018, he works as a teaching assistant in computer science and serves as head TA for operating systems and algorithms courses in 2017 and 2018.

2. December 2018

   The wav2letter++ framework

   His early research at Facebook AI Research (FAIR) focuses on speech recognition and learning from limited supervision. In December 2018, he coauthors the paper introducing wav2letter++, an open-source speech recognition framework written entirely in C++ that in some cases trains end-to-end models more than twice as fast as other optimized frameworks.

3. December 2019

   The Libri-Light benchmark

   In December 2019, he leads and first-authors Libri-Light, a benchmark for speech recognition with limited or no supervision. Built from LibriVox audiobooks, its more than 60,000 hours of spoken English form what the team describes as the largest freely available speech corpus.

4. 2022–2023

   Flashlight and Shumai

   In January 2022, he is first author of the paper introducing Flashlight, an open-source C++ machine learning library built to help systems researchers prototype new ideas, and the work is presented at ICML 2022. He leads the teams behind Flashlight, wav2letter and Shumai, a differentiable tensor library for JavaScript and TypeScript, and parts of this software find their way into PyTorch. He presents Shumai at Google DeepMind in 2023.

5. 2024

   Chameleon and Transfusion

   He co-leads Chameleon, a text-and-image model at Meta. In August 2024, he coauthors Transfusion, which trains a single transformer on text and images by combining next-token prediction with diffusion, with models of up to 7 billion parameters.

6. September 2025

   Code World Model

   In September 2025, Meta's FAIR CodeGen team releases Code World Model (CWM), a 32-billion-parameter open-weights LLM mid-trained on observation-action trajectories from Python interpreters and agentic Docker environments. He helps lead its development and presents CWM at the AI Engineer conference in 2025.

7. As of October 2026

   Codegen co-lead at Meta Superintelligence Labs

   He is a research manager and engineer at Meta Superintelligence Labs (MSL), where he co-leads code generation research across FAIR and TBD with a team of scientists and engineers in Menlo Park, New York and Paris. He is also a lecturer in Penn's Department of Computer and Information Science, where he teaches CIS 5690, a course he created on GPU programming for machine learning systems. He serves as an advisor to Hunterbrook Media, an investigative news organization.