Jacob Kahn is a research manager and engineer at Meta Superintelligence Labs, where he co-leads code-generation research. He helped lead development of Code World Model, an open-weights system that learns how programs execute, giving software agents a foundation for anticipating the consequences of their actions. He also teaches computer science at the University of Pennsylvania, where he created a course on GPU programming for machine learning.
From speech recognition to coding agents
Kahn studied computer science, statistics, operations research, and economics at Penn through its Management and Technology program. His early research at FAIR addressed speech recognition and learning from limited supervision: he contributed to the open-source wav2letter system and led Libri-Light, a benchmark for speech recognition with scarce labeled audio.
He subsequently led teams behind Flashlight, a C++ machine-learning framework designed for efficient experimentation, and Shumai, a differentiable tensor library for JavaScript and TypeScript. His research expanded into multimodal generation through Chameleon, a text-and-image model he co-led, and Transfusion, which combines language modeling with image diffusion.
Code World Model, or CWM, applies that systems and modeling experience to software. The 32-billion-parameter dense transformer learns from repository-scale code and program execution traces that capture changing variables, program states, and executed lines. Its public repository provides inference code, checkpoints, evaluation materials, and neural-debugging demonstrations.
- Execution-aware code generation. Kahn models programs as computational processes whose successive states can be predicted. Pull requests, tests, and continuous-integration runs supply training signals beyond source-code syntax.
- Computational world models. He frames program execution and agent decisions through the same state-action-outcome structure: an autoregressive model can estimate what code or a command will do before interacting with an actual environment.
- Bash-centered software agents. CWM learns to inspect repositories, search files, edit code, and run tests through ordinary terminal workflows. Supervised fine-tuning, reinforcement learning, and failed trajectories improve behavior inside realistic engineering environments.
- Asynchronous reinforcement learning. Queued model checkpoints and sampled trajectories allow environment execution, scoring, and training to proceed concurrently; checkpoints can even change during an ongoing interaction.
- Neural debugging. Execution-aware models can infer intended behavior from incomplete code, estimate intermediate values, and explore expensive computations without always running them. Their approximate reasoning about termination does not imply a solution to the general halting problem.
