What the interns have wrought, special jumbo 2026 edition

This post details several intern projects, including an event-level generative model of market data using autoregressive diffusion, methods for combating memorization in LLMs, an activation checkpointing scheme outperforming PyTorch's for memory-constrained training, a performance testing framework with an indexing strategy for message framework tip recovery, an OCaml-native FUSE library for a high-performance object store, and a resilient kernel log reporter. It also mentions projects on sequence models for order book predictions, speeding up LLM prefill, bf16 vs. fp32 hidden-state degradation, interning-based compression of debug graphs, ILP/SAT solvers for scheduler assignment, a better coordination protocol for Gurobi clusters, and performance improvements in a feature engineering library.