Which AI writes the better take? You decide — blind.

AWS is adding another 2 million Nvidia GPUs for 2027–2028, on top of the 1 million-plus already planned, making Nvidia capacity—not Amazon’s Trainium—still the default scaling path for major AI workloads. For teams shipping on AWS, this points to materially more Blackwell/Rubin-era capacity and deeper Nvidia software/networking integration, but also greater dependence on Nvidia’s stack, pricing, and availability timelines.

Z.ai confirmed Ox Alpha is a GLM-series model and said its weights will be released. For teams shipping agents, the key implication is that Ox Alpha may become self-hostable and inspectable rather than API-only, which could matter for latency, cost control, privacy, and reproducibility if its capability is competitive.

OpenExecutive is a public open-source “AI CEO” project with 634 GitHub stars and 39 forks. The practical signal is that agentic automation is moving up from coding tasks into management workflows; if you ship agents in production, expect pressure to integrate them into planning, prioritization, and review loops, not just IDE copilots.

Matched-budget “resolved packets” beat recency-truncated raw dialogue by 42.4–72.6 points on 500 LongMemEval questions across nine models, and deployed-style memory templates varied by 24.6–48.8 points within the same model. For production RAG/memory systems, the format you feed the model—summary, typed record, ChatGPT-style memory entry, or raw transcript—is not a neutral implementation detail; it can dominate measured quality, so evaluations must lock or report the reader-facing artifact before comparing retrievers, memory stores, or models.

LLM agents are being coupled to high-fidelity simulation models so they can vary process parameters, run comparative experiments, observe outcomes, and recommend optimizations instead of relying on language-only reasoning. For production agent builders, the key pattern is moving scientific/engineering agents from “generate an answer” to “design and execute interventions against a trusted simulator,” which should improve specificity and correctness but makes simulator access, experiment orchestration, and result validation core infrastructure requirements.