Morning Digest — November 14, 2026

Today’s strongest signal is convergence: wireless networks are becoming sensors, satellites are becoming production payloads, and AI systems are becoming governed infrastructure rather than isolated models.

SpaceX launched Starship Flight 15 and completed deployment of all 26 Starlink V3 satellites. The upper stage then entered a planned coast of roughly 21 hours, with SpaceX reporting healthy systems several hours into the mission.

The extended flight is intended to culminate in a return-to-launch-site catch attempt at Starbase. Beyond the recovery milestone, placing a full batch of V3 satellites into orbit makes this a direct demonstration of Starship’s role in scaling future LEO-network capacity.

Source: SpaceX mission update and Flight 15 page

Nokia wins a Pentagon 5G/6G sensing prototype contract

Nokia Federal Solutions received a RAPID contract from the U.S. Department of War to prototype integrated sensing and communications infrastructure. The program will use 5G and future 6G systems to detect and track objects such as drones rather than treating the radio network only as a communications pipe.

The award is a concrete step for ISAC, one of 6G’s defining research directions. A successful prototype would show how cellular infrastructure can contribute persistent multi-domain awareness while reusing radios, spectrum, timing, and edge-compute resources already needed for connectivity.

Source: Nokia

DeepMind opens the datasets behind WeatherNext 3

Google DeepMind released live feeds and historical archives for PARDIG and SatWeave, two foundational datasets used by WeatherNext 3. SatWeave combines six geostationary satellites into an hourly global mosaic spanning 17 visible, water-vapor, and infrared channels, with more than a decade of archived observations.

PARDIG produces global precipitation maps every 15 minutes at roughly 10-kilometer resolution. Evaluations against radar networks and more than 36,000 weather stations found that it improved light-to-moderate hourly precipitation detection by more than 30% over NASA’s IMERG-Final baseline, with applications ranging from flood warnings to wildfire and solar-power monitoring.

Source: Google DeepMind

Anthropic publishes a playbook for stronger AI safety evaluations

Anthropic published ten recommendations for designing, running, and reporting safety evaluations. The guidance calls for multidisciplinary teams, sandboxed and monitored test environments, designs matched to causal or observational questions, better construct validity, and statistically defensible comparisons.

The larger point is that model-safety claims are only as credible as the evals behind them. Anthropic also recommends preregistering analyses, documenting changes and limitations, and publishing enough detail for independent reproduction—practices borrowed partly from psychology’s replication reforms.

Source: Anthropic

Anthropic details its IPO and November 19 trading target

Anthropic has launched the roadshow for its initial public offering and expects, subject to regulatory clearance, to set the final price on November 18 and begin trading on Nasdaq under the ticker “ANT” on November 19. Goldman Sachs and Morgan Stanley are among the firms distributing the preliminary prospectus.

The company’s IPO site reports $17 billion in Q3 2026 revenue, more than 11,900 customers generating at least $100,000 in trailing-twelve-month revenue, and a 15% adjusted operating margin. It also reports a $23.3 billion net loss, largely driven by a $21.9 billion non-cash valuation charge tied to convertible instruments.

Source: Anthropic IPO

NVIDIA contributes its OpenShell agent runtime to CNCF

NVIDIA is contributing OpenShell to the Cloud Native Computing Foundation as a Sandbox project. The open runtime applies deterministic policies while autonomous agents use tools, access inference services, and execute workloads, with integrations for credential storage, monitoring, and enterprise compute.

NVIDIA is also supporting Google-originated Agent Substrate, which suspends idle long-running agent environments and resumes them when work returns. New DSX components—NVIDIA Kubernetes Engine, Kubernetes Recipes, and Snapshot—extend the same effort from agent governance into cluster lifecycle management and faster model-replica restoration.

Source: NVIDIA

NVIDIA’s GR00T 2 transfers robot skills across bodies

NVIDIA introduced GR00T 2, an open world-action model built on Cosmos and trained with diverse robot data plus 100,000 hours of open egocentric human video. The model is designed to learn dexterous manipulation from demonstrations and transfer skills across different robot bodies.

GR00T 2 completed 40 of 50 evaluated tasks at least once without task-specific demonstrations. NVIDIA reports 71% average success with one demonstration per task and 95% with five, while an agent-facing API lets higher-level systems sequence skills and recover from errors. Foxconn, Siemens, and other partners are evaluating the model for industrial work.

Source: NVIDIA

Qualcomm open-sources an LLM behavior engine for game NPCs

Qualcomm released the MindSand AI Behavior Engine as open source, giving game developers a way to combine LLM-driven interactions and dynamic behaviors with established game-AI systems. The design aims to add adaptive reasoning without discarding deterministic behavior trees and other production controls.

Qualcomm’s initial case study applies MindSand to smarter teammates in Starship Troopers: Extermination. The release is a useful edge-AI signal: generative behavior is moving into latency-sensitive interactive systems where cost, consistency, and local execution matter as much as conversational quality.

Source: Qualcomm

Research Radar

Context Graph: Explainable Multi-Agent Decision Chains in AI-RAN

Xijun Wang, Zhaoyang Liu, Xiang Chen, Howard H. Yang, and Chenyuan Feng propose a heterogeneous graph that links policies, measurements, decisions, and alerts across AI-RAN xApps and rApps. Decision-event nodes act as causal junctions, and counterfactual verification upgrades downstream edges from correlation toward causality.

The proof of concept operates over existing E2, A1, and O1 data flows and resolves compliance and diagnosis queries at sub-millisecond latency on graphs with tens of thousands of nodes. That directly targets a difficult operational problem: explaining which autonomous controller caused a degradation and how other agents amplified it.

Source: arXiv:2611.16879

Agentic Orchestration for SLA-Constrained Network Slicing in Programmable Mobile Data Planes

Rodrigo Moreira, Larissa Ferreira Rodrigues Moreira, Joberto S. B. Martins, Tereza C. Carvalho, and Flávio de Oliveira Silva combine UPF telemetry with a locally hosted LLM that selects among admissible maximum-bit-rate settings for overloaded 5G slices. Deterministic logic validates each choice before producing PFCP updates.

The evaluated runs preserved the highest-priority slice’s configured rate and produced neither oscillations nor fallback events, though delivered per-class QoS remains unmeasured. The manuscript has been accepted for IEEE CCNC 2027.

Source: arXiv:2611.16801

Physics-Aware Transformer-Based Beamforming Optimization for Near-Field E-MIMO ISAC Systems

Mohsin Ali, Abdulahi Abiodun Badrudeen, Azharul Islam, and Sunwoo Kim propose PATBO for near-field extreme-massive-MIMO integrated sensing and communications. Physics-derived communication and sensing directions reduce a 1,024-antenna beamformer from 12,288 real-valued outputs to six learned stream-power variables.

The complete hybrid system retained 99.89% of the reference rate, reached 100% constraint satisfaction on the retained test set, and required optimizer fallback for 12.21% of scenarios. Direct inference took 1.11 milliseconds—roughly 800 times faster than the evaluated sequential-quadratic-programming baselines.

Source: arXiv:2611.16305

Source Notes

IEEE Xplore and ACM Digital Library returned no newly indexed eligible papers. Research Radar therefore uses verified arXiv records from the last seven days, including one manuscript accepted for IEEE CCNC 2027.