We read ~50 engineering blogs so you don't have to. Every Saturday you get the dozen posts worth your time — inference, training, RL, agents — summarized in one email and an EPUB for your Kindle.
October 17, 2026
~50 engineering blogs and Hacker News, scanned every week.
Benchmarks, kernels, training recipes. No launches, no fluff, no repeats.
In your inbox Saturday morning, or as an EPUB on your Kindle.
This week focused on unified megakernel decode paths for large MoE models (Kimi K3 and Reflection Beam 501B) across NVIDIA Blackwell GB300 and AMD Instinct MI355X GPUs, custom VM filesystems for high-concurrency agent sandboxes, and elastically deterministic streaming data loaders in PyTorch TorchTitan.
For latency-critical MoE decode on Blackwell and MI355X, fuse cross-operator communication into a persistent decode megakernel and profile P90 interactivity at 50 tok/s/user rather than unthrottled peak throughput.
Pair 500B+ sparse MoE models (20B–25B active params) with expert parallelism and adaptive draft verification (DFlash2) to keep decode memory-bandwidth utilization high without regressing prefill TTFT.
When hand-tuning HIP/ROCm GEMM kernels on RDNA GPUs, track VGPR count and LDS bank conflicts after every tiling change, as Wave32 occupancy drops sharply once per-thread register thresholds are crossed.
Adopt elastically deterministic streaming data loaders in TorchTitan so data mixture ablations and node-count rescaling never require offline corpus re-materialization or sacrifice training reproducibility.
Move fuzzy deduplication and model-based quality scoring onto GPU pipelines early in pretraining corpus construction to iterate on synthetic data mixtures at multi-trillion-token scale.
Profile filesystem metadata and container fork latency in your RL/agent sandbox fleet; content-addressed copy-on-write caching and high-core-count host CPUs prevent GPU starvation during tool execution.
Exploit rack-scale NVLink domains (NVL72) to share prefix KV caches across forked sub-agent trajectories rather than recomputing repository context per agent worker.