Morning Digest — October 27, 2026

Today’s strongest signal is that AI is becoming part of operational infrastructure rather than remaining a separate application layer. OpenAI is driving down the cost of computer-using agents; O2 and Nokia have put predictive optimization into a live 5G RAN; and Finland’s 6G-ROCKET program is designing future radio networks around AI, sensing, and digital twins from the start. Meanwhile, Starlink is expanding both direct-to-phone and aviation connectivity, Google DeepMind is entering peer review, and medical AI is moving into pre-surgery risk assessment.

OpenAI launches GPT-6.1 Terra for faster, cheaper computer use

OpenAI introduced GPT-6.1 Terra as a cost-efficient model for visual reasoning, writing, coding, and computer use. On OpenAI’s Agents’ Last Exam benchmark, Terra at medium reasoning reportedly scores 15 percentage points above GPT-5.6 Terra at about one-third the cost and matches GPT-6 Astra at one-tenth the price per token. OpenAI also says it completes tasks more than three times faster than Astra and GPT-6.1 Sol. On Terminal-Bench 4.0, Terra reaches 50.3% at maximum reasoning, more than double GPT-5.6 Terra’s published result. The model is available in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users, and through the API at $1 per million input tokens and $5 per million output tokens. The strategic shift is less about one benchmark lead than about making capable computer-use automation economical at high volume.

Source: https://openai.com/index/introducing-gpt-6-1-terra/

O2 and Nokia put predictive AI into a live 5G RAN

O2 and Nokia completed what they describe as the UK’s first live deployment of an AI-powered radio access network feature. The trial ran at O2’s Giga Site in London Paddington while serving commercial customers. Instead of relying only on static radio-resource algorithms, the system continuously analyzed conditions, predicted shifts in demand, and adjusted spectrum and capacity allocation. O2 reports an approximately 20% improvement in spectral efficiency, with higher throughput during busy periods and stable accessibility. That result is meaningful because it extracts more capacity from existing spectrum and infrastructure—the exact economic pressure that will shape the transition from 5G Advanced toward AI-native 6G.

Source: https://news.virginmediao2.co.uk/o2-and-nokia-deploy-ai-technology-to-improve-5g-network-performance-in-uk-first/

Finland launches the €1.4 million 6G-ROCKET AI-RAN program

The University of Oulu’s 6G Flagship launched 6G-ROCKET, a two-year research program supported by Business Finland and joined by Nokia, NVIDIA, DNA, Keysight, MediaTek, Northeastern University, and the University of Tokyo. The consortium will build on a GPU-accelerated AI-RAN platform connected to Oulu’s 5G/6G Test Network. Its agenda includes AI-native RAN design, real-time digital twins, multi-sensor perception, intent-based autonomous control, and an AI-RAN Perception Lab. The partner mix covers the full experimental chain—network systems, accelerated compute, operator requirements, measurement, channel validation, chipsets, and user equipment—giving the program a credible path from algorithms to repeatable hardware demonstrations.

Source: https://www.6gflagship.com/news/6g-rocket-ai-ran/

Starlink says its direct-to-phone service is now live nationwide in Brazil through Vivo. The announcement puts 50 million Vivo customers within reach of apps and messaging in terrestrial dead zones and says seven million phones are already connected. This is a material new development beyond recent Brazilian satellite-to-phone agreements: the service has moved from partnership formation into nationwide operation. Starlink also previewed Mobile V2 as the step intended to provide high-speed mobile internet, which would expand direct-to-device service well beyond today’s lower-bandwidth messaging use cases.

Source: https://x.com/Starlink/status/2114747196226859108

Google DeepMind and Springer Nature build an AI peer-review assistant

Springer Nature and Google DeepMind are co-designing an AI assistant to help reviewers navigate manuscripts and identify sections that merit closer attention. The prototype uses Gemini models and Google’s open-source Agent Development Kit to analyze text, charts, and figures, while the partners emphasize that expert reviewers retain scientific judgment. Confidentiality is a central architectural requirement: the system runs inside Springer Nature’s private Gemini Enterprise environment, and manuscript content remains there during development and testing. If successful, the tool could reduce the mechanical burden of review without delegating acceptance decisions, but its real value will depend on transparent evaluation of missed errors, false alarms, bias, and reviewer overreliance.

Source: https://group.springernature.com/gp/group/media/press-releases/research-partnership-with-googledeepmind-to-support-peer-review/53891572

easyJet will adopt in-flight connectivity for the first time by installing Starlink across its modern Airbus fleet. Deployment is scheduled to begin in the first half of 2027 on A320, A320neo, and A321neo aircraft and finish by early 2028. The airline operates 355 aircraft across 38 countries and intends to support streaming, gaming, and live flight information. For Starlink, the agreement extends an aviation pipeline that SpaceX recently said covers roughly 12,000 aircraft. For the satellite market, it demonstrates that LEO broadband is no longer confined to premium long-haul cabins; it is becoming commercially relevant to high-utilization, short-haul fleets where installation time, antenna economics, and turnaround reliability are especially demanding.

Source: https://www.satellitetoday.com/mobility/2026/10/26/easyjet-to-adopt-ifc-with-starlink-deal/

NVIDIA-backed AI estimates breast-cancer recurrence risk from existing MRIs

NVIDIA highlighted SimBioSys’s FDA-cleared TumorSight Risk system, which estimates five- and ten-year recurrence risk for patients with HR-positive, HER2-negative early-stage breast cancer. The software combines an existing pre-surgery breast MRI with pathology and clinical information, potentially providing additional evidence before treatment decisions are finalized. The important deployment detail is that it works from imaging already collected in the normal care pathway rather than requiring a new specialized scan. As with any clinical risk model, the practical standard is not only discrimination on validation data but calibration across patient populations and evidence that the output improves decisions when used alongside physicians.

Source: https://x.com/nvidia/status/2114827978748383342

Research Radar

From Human Intent to In-Orbit Action: A Self-Learning Large Language Model Agent for 6G Satellite Networks

Zhengqian Xu, Yasheng Zhang, Chenhua Sun, Shan Lu, and Chengcheng Li introduce ALMOND, a framework that translates natural-language service goals into multi-objective reinforcement-learning rewards and then learns constraints from failed policies. Its ground-centric design keeps the LLM off the spacecraft; satellites execute distilled policies below 50,000 parameters with sub-15 ms inference. In a 350-satellite digital twin using real ephemeris data, the system raises polar handover success from 55% to above 95%, converges about three times faster than baseline reinforcement learning, and reduces reward-engineering time from 48 hours to 4.5 hours. The work directly addresses the gap between intent-based networking concepts and resource-constrained in-orbit control.

Source: https://arxiv.org/abs/2610.34387

Satellite Intelligence: Distributed Training, Collaborative Inference, and LLM Adaptation in LEO Constellation

Zheng Lin and colleagues present a survey that jointly covers distributed training, collaborative inference, and on-orbit LLM adaptation in LEO networks. The paper organizes federated and split learning, single- and multi-satellite inference, satellite-ground execution, and resource-aware LLM serving under a shared constraint model shaped by intermittent links and severe size, weight, power, memory, and compute limits. Its value is the synthesis: these topics are often studied separately, even though an operational space-native AI system must decide where every training, inference, synchronization, and adaptation step belongs.

Source: https://arxiv.org/abs/2610.35282

Denial of Service Attacks on Beam Allocation in Large LEO Constellations

Ali Abbasi, Pasha Barahimi, Marco Paolieri, and Leana Golubchik study attacks that inject artificial terminal demand into LEO beam-allocation systems. Their results expose a robustness-performance tradeoff: greedy allocation with demand-proportional sharing performs best before an attack but is also the most vulnerable. Under the tested attacks, demand-proportional sharing increases the fraction of under-served legitimate users by as much as 34%, while equal sharing limits the increase to 13%. Max-min-fair allocation also forces an attacker to target more cells to produce widespread damage. The paper makes beam scheduling a security problem, not merely an optimization problem.

Source: https://arxiv.org/abs/2610.34631

Source note

Bloomberg blocked automated extraction for a Meta financing story, so that candidate was omitted rather than summarized from an inaccessible page. IEEE Xplore and ACM Digital Library searches returned no newly indexed matching papers; the research section therefore uses verified arXiv records from the last seven days.

Takeaway

AI is moving from an add-on into the control plane of work, radio networks, satellites, science, aviation, and medicine—making efficiency, security, and human oversight the defining design constraints.