What the interns have wrought, special jumbo 2026 edition (blog.janestreet.com)

Jane Street recently highlighted the impactful projects developed by its interns during a summer program, notably expanding the scope to include contributions from ML research and engineering, as well as software and systems engineering. Among the standout initiatives, intern Arsh Koneru introduced an activation checkpointing scheme that optimizes memory usage beyond PyTorch’s conventional limits, enhancing the efficiency of model training. This innovation addresses a critical challenge in deep learning, balancing compute and memory constraints effectively. Additionally, ML interns like Kavish and Monte explored generative modeling of market data and methods to reduce memorization in large language models, respectively, demonstrating the potential for significant advancements in synthetic data generation and predictive accuracy. The significance of these projects lies in their direct implications for both AI/ML advancements and operational efficiencies within trading environments. By refining techniques like event-level modeling and memory optimization, the interns not only contribute to the academic discourse around AI but also prepare Jane Street’s infrastructure for more resilient operations in high-stakes markets. Projects in Linux engineering, such as Jacob Root’s kernel log reporter and Kian Kasad’s high-performance FUSE library, also reflect a commitment to enhancing system reliability and data accessibility, ensuring that technological foundations support innovative trading strategies.