Bedrock Knowledge Base adds single-shot and retrieval-only RAG
One AgenticRetrieveStream param now sets the mode per query — plus NVIDIA's open Isaac Lab 3.0, Google Cloud Pathways GA, and a no-retrain MoE pruner.

Copy markdown
One param, three retrieval modes
Amazon's Managed Knowledge Base now exposes retrieval-only (maxAgentIteration=0), single-shot RAG (=1), and full agentic retrieval (2–5) on one AgenticRetrieveStream call. Simple lookups skip query planning for lower latency and fewer model calls — no second API, no separate code path.
NVIDIA opens Isaac Lab 3.0 and GR00T 2
At CoRL, NVIDIA shipped Isaac Lab 3.0 — now switchable between the PhysX and Newton physics solvers — and GR00T 2, which hit a 71% average success rate across 50 manipulation tasks from a single demo each. Both are free to train robot policies in simulation.
Cloud Pathways hits GA on GKE
Google's distributed runtime for large accelerator fleets is now generally available on GKE, after customers logged 130M+ TPU chip-hours in 90 days. The new TPU 8tx pairs TPU silicon with AMD EPYC hosts for confidential inference at up to 2x lower latency.
Prune MoE experts without retraining
Amazon's HOPE scores pairwise expert interactions to drop whole experts from MoE models with no retraining, and it stacks with quantization. It beats first-order pruning by up to 6.1% on agentic coding benchmarks, tested on Qwen3.5-122B — cheaper self-hosting for big mixtures.
vLLM's scheduler: 2.91x over batching
A fresh teardown pins vLLM's continuous batching at a 2.91x throughput gain over static and dynamic batching, and 3.86x end-to-end versus naive HuggingFace transformers. Worth a config pass if you're self-serving models under load.