NIST bans FF3, Kafka 4.3 linger.ms, 92°F trick

Kafka 4.3’s default linger.ms rose from 0 to 5 ms, shrinking latency under light load but boosting throughput when traffic spikes. The change lets producers accumulate larger batches, cutting CPU overhead at the cost of a few milliseconds extra latency per record. Teams should tune linger.ms to match per‑partition rates for optimal performance.

The second public draft of NIST SP 800‑38G Rev. 1 removes the FF3 method after cryptanalysis showed weaknesses, leaving FF1 as the sole approved format‑preserving encryption. It also raises the minimum domain size to one million and disallows the inverse AES cipher and floating‑point arithmetic, tightening security for real‑world deployments.

A Fox 8 New Orleans graphic uses a baseline just under 92°F, making a two-degree temperature change appear dramatic. The piece illustrates how baseline choices can distort perception, reminding designers to anchor charts at zero or an appropriate reference point.

Google Research unveiled SensorFM, a foundation model trained on over a trillion minutes of wearable sensor data from five million users. The model delivers a universal physiological representation that transfers across 35 health tasks, enabling label‑efficient adaptation and powering a Personal Health Agent. This could slash data‑label costs and accelerate personalized preventive care.

Aurora 1.5 expands Microsoft’s open Earth‑system foundation model with 22 new weather variables, hourly resolution, and probabilistic ensemble forecasts. The update makes the model ready for real‑world climate, energy, and agriculture use cases, and ships as open‑source code and checkpoints for developers.

Malloyyo provides a thin MCP server and web UI that expose Malloy semantic models to LLMs, guaranteeing consistent SQL generation and shareable query results. It ships with DuckDB support, a CLI for model publishing, and one‑click Vercel deployment, letting teams hand off reliable analytics to AI assistants.

SmolVLM2‑2.2B fits on a 5 GB GPU yet outperforms larger 2 B models on video benchmarks, letting you generate structured JSON summaries, scene descriptions, key moments, action items, entirely locally. The KDnuggets guide wires frame extraction, batch inference, and timestamped output into a single pipeline that runs on an RTX 3060, MacBook M2, or free Colab T4.

The article measures DDP, FSDP and DeepSpeed ZeRO on four A100 80GB GPUs and shows that mismatched NVLink wiring on H200 nodes can slow training by up to 2×. It proves that choosing a parallelism strategy without accounting for inter‑GPU bandwidth can waste hardware and time.

The Hugging Face blog walks you through profiling PyTorch's attention layers with the built‑in Profiler, exposing costly matmuls, scaling, masking, and softmax steps. By comparing naive, in‑place, and SDPA implementations, you can pinpoint bottlenecks and choose the fastest kernel for transformer models, shaving GPU time and cost.

Monte Carlo deployed 19 AI‑driven agents to replace manual security architecture and vulnerability reviews. The agents write findings to a shared, git‑tracked folder, which triggers downstream tasks and yields 5-10 high‑severity tickets per day, without expanding the security team.