Weekly Paper Notes — three entries per week: 1 seminal + 2 top picks.
🔁 Seminal Paper of the Week
- Congestion Avoidance and Control: Van Jacobson, with Michael J. Karels (SIGCOMM 1988). Written after the 1986 Internet congestion collapses, it derives slow start, ACK clocking, a variance-aware retransmit timer, and AIMD from one rule, conservation of packets. All of it went into TCP senders without touching the protocol or a single router.
⭐ Top Papers of the Week
- Mixture-of-Kittens: MoE Megakernel for NVL72s: Sul, Brown, Wildermuth, Lin, Cassano, Ré, Stanford and Cursor Research (arXiv:2609.36070). On a 72-GPU NVLink domain, existing MoE communication libraries often lose to plain NCCL. Pull dispatch with push combine, a tunable overlap granularity, and a device-side ring buffer in one deterministic kernel give up to 2.37x over the best public baseline and 1.41x end to end on 512 GB300s.
- Context Language Models: Shao, Shen, Yin, …, Lewis, Yih, Zettlemoyer, Koh, UW and Meta (arXiv:2609.37725). The live context becomes a file the model edits with Bash. Zero-shot it beats summarisation and tool-based context managers on deep research, terminal coding, and 24-hour agent swarms. The strategy can be steered by prompt or trained with RL, and Suffix Cache Reuse cuts the serving cost of mid-context edits by 35%.
The full weekly Slack digest covered ~15 more papers across AI/ML, NLP, distributed computing, databases, OS and networking; only the highest-impact picks are expanded into blog notes.
