Research

Research

SparklingTree: 30-40% faster speculative decoding over DSparkAug 5, 2026Specialization is (sometimes) all Speculation needsJul 23, 2026Biting the Bullet: Predictive Speculative KV Replication for Bursty LLM InferenceJul 22, 2026

Engineering

jwLLM: Inference engine entirely in cpp

0.2 tok/sec on macbook pro (jwLLM part 1)Sep 15, 2026Building a BoundingBlockingQueue from just an Atomic (jwLLM part 2)Sep 26, 2026

Making computer-use models faster with interrupts instead of pollingSep 14, 2026Infer-Sim: An open-source simulator for routing algorithms and cache policies for inference workloadsJul 21, 2026Decoding Speculative Decoding from First PrinciplesJul 7, 2026

Side Quests

Grand Theft AlignmentSep 26, 2026Using CV and a projector to help me play billiardsSep 15, 2026DJing a club with an AI agentSep 11, 2026