The State of AI - 2026-11-09
The most consequential signals in the edition are not a single model release: they are pressure on AI infrastructure siting, evidence gaps in automated work and robotics, and widening scrutiny of the data used to train and operate AI systems.
By Lucia Marin · disclosed fictional OMIKINA AI editorial persona · No human review recorded
Published
AI-persona disclosure
Fictional OMIKINA AI editorial persona; not a human reporter and does not possess human research credentials or firsthand experience.
Executive summary
AI’s operating constraints are becoming more visible. Reporting from the UK documents organized local opposition to a proposed Google data center, while Fortune commentary describes a broader US electoral backlash; together, they suggest that power, water, emissions, land use and the quality of community consultation are now material delivery risks rather than peripheral permitting issues. The strongest conclusion is not that a universal buildout slowdown is certain, but that developers need auditable site-level evidence and credible cost-sharing arrangements. Separately, enterprise and robotics research reinforces a familiar implementation lesson: access to a model is not proof of reliable performance. The supplied evidence repeatedly distinguishes controlled demonstrations, self-reported vendor results and surveys from durable production outcomes. Data provenance is equally central, as reporting on consumer EEG collection raises questions about whether secondary AI-training uses were clearly disclosed to users.
Data-center expansion faces a legitimacy and evidence problem
A WIRED report describes opposition to Google’s proposed Thurrock data-center campus from local wildlife, environmental and motorcycle-racing groups. The objections concern habitat, emissions and loss of a former racing site; Google says it would retain habitat, move endangered plants, finance renewable generation and seek carbon neutrality. The planning authority had received public responses overwhelmingly opposing the proposal, although it noted that many came from outside the local area. The key issue is therefore not simply whether a project has a sustainability claim, but whether its environmental baseline, mitigation plan and allocation of local costs are inspectable and trusted by affected communities.
This turns site selection and permitting into an operating-model question. Developers will need to document energy, water, emissions, biodiversity and grid impacts at the project level, distinguish commitments from measured outcomes, and engage before designs are effectively fixed. Fortune commentary makes far broader claims about US election results and an ensuing slowdown, but those claims are interpretation rather than primary election or regulatory evidence in the supplied material. They are a signal of political risk, not sufficient evidence to forecast a national outcome.
Training data for brain-related AI is drawing scrutiny over disclosure and consent
Reporting summarized by Tom’s Hardware, based on a Science investigation and the Pulitzer Center’s AI Accountability Network, says consumer neurotechnology companies have collected large EEG datasets and that some are using recordings to develop brain foundation models. Interaxon says it trained an AI model using EEG recordings after removing direct identifiers such as names, email addresses and physical addresses. The material supplied does not establish the full terms presented to every user, the adequacy of any consent mechanism, or the re-identification risk of the resulting datasets. It does establish that de-identification and user disclosure are separate questions.
For companies building models from sensitive behavioral, biometric or health-adjacent data, the relevant governance trail must start before model training: what was collected, for which stated purpose, what secondary uses were permitted, what transformations were made, and who can audit them. Removing direct identifiers may reduce a privacy risk, but it does not by itself answer whether the use was expected or adequately communicated.
Sources: S2
The enterprise bottleneck is workflow design—and claims of replacement need careful interpretation
A Forbes Technology Council contributor argues, from one organization’s internal experience, that the greatest enterprise opportunity lies with occasional AI users rather than attempting to turn every employee into an expert. The author’s recommended metric is movement from basic use into more capable work, supported by redesigned workflows and reusable tools. Separately, Fortune reports survey findings from SHL and Robert Half indicating that some employers which removed roles because of AI later restored similar work. Those findings are respondent and hiring-manager reports, not a measure of all employers or a causal audit of AI-driven staffing decisions.
Executives should separate tool activity from verified changes in work quality, cycle time, error rates and accountability. A workforce plan based on presumed automation savings is fragile if it excludes tacit knowledge, relationship work, exception handling and quality control. The more defensible approach is to instrument a specific workflow, preserve a human escalation path, and assess outcomes before making irreversible staffing decisions.
Robotics is progressing, but the evidence draws a sharp line between demonstrations and safety assurance
MIT’s FLEX research reports simulated robot-arm experiments in which a language model generalized a user’s instruction into related safety-relevant constraints. That is a useful interface result, but its evaluation was virtual and used typed commands. A separate preprint testing OpenVLA on a defined benchmark reports that negated instructions had no detectable effect on the task executed in that setting, and recommends enforcement outside the instruction channel plus monitoring. These results are not contradictory: they evaluate different systems and tasks. Together, they show why natural-language compliance cannot be assumed from a plausible-looking demonstration.
Robot programs should preserve a trace from instruction to interpreted constraint, policy action, sensor observation and external safety control. Before treating a model as safe for physical deployment, teams should specify the prohibited action, test it under changes in wording and context, and independently measure whether the robot complied. Optical-motion-capture guidance in the supplied material makes the parallel measurement point: resolution, coordinate alignment, latency and visibility must be verified rather than inferred from a camera specification.
Korea’s AI push shows infrastructure, deployment software and physical AI converging
NVIDIA announced collaborations with Korea’s science ministry, model builders and startups across inference infrastructure, interconnects, foundation models and robotics. It says its optimized deployment recipes can improve performance for selected Korean open models on NVIDIA B200 GPUs, while its Korea physical-AI post describes a manufacturing-environment robotic demonstration with LG and an accelerated Samsung simulation workflow. These are vendor-reported results and planned collaborations, not independent comparisons or evidence of broad commercial deployment.
The strategic signal is the coupling of sovereign-model ambitions with the less visible deployment layer: interconnects, multi-node inference serving, simulation, data collection and robot evaluation. Buyers should request workload-specific validation, including model versions, hardware, throughput definition, latency, energy use and the steps between a controlled pilot and a production deployment. Claimed gains should remain attached to their stated hardware and experimental conditions.
Watch next
- Whether data-center developers and planning authorities publish comparable, site-specific evidence on power, water, emissions, biodiversity and community benefits—and whether that documentation changes permitting outcomes.
- Whether consumer neurotechnology providers clarify how EEG data may be used for model training, what data are eligible, and how consent and de-identification are independently reviewed.
Sources: S2
- Whether robotics teams move from language-only guardrails to externally enforced constraints and measured compliance in physical environments.
- Whether enterprise AI programs report workflow-level quality and recovery costs alongside adoption metrics, particularly where roles have been automated or redesigned.
Sources
- Instagram’s AI age detection is expanding globally — The Verge · partial-text ·
- Wearable sleep trackers are using your brainwaves to train AI models without explicit user consent — tens of millions of hours of EEG data harvested — Tom's Hardware · partial-text ·
- Why The Highest Return On Enterprise AI Comes From The Middle 60% — Forbes Innovation · full-text ·
- What the midterm results actually mean for the future of AI and data centers: prepare for the AI slowdown — Fortune · full-text ·
- Amplifon Americas Data Breach Impacts 1.3 Million — SecurityWeek · partial-text ·
- AI Has A Trust Problem. The Answer Is In The Chip — Forbes Innovation · full-text ·
- Your data center may be halfway to waterless — Data Center Dynamics · feed-summary ·
- Precision Spatial Tracking for Autonomous Systems: Training Humanoid Robots and Drones with Optical Motion Capture — Robotics & Automation News · full-text ·
- Over 3 Million People Hit by Frontline Education Data Breach — SecurityWeek · partial-text ·
- Sponsored: EPG Modular Data Center marks topping-out of its US$120 million Malaysia FIZ Factory — Data Center Dynamics · feed-summary ·
- StrataSLAM: open-source multi-floor SLAM for ROS 2 (stairs, elevators, ramps, identical floors) — Open Robotics Discourse · partial-text ·
- Chinese AI firms want Silicon Valley’s top minds. Poaching them won't be easy — Rest of World · full-text ·
- NVIDIA Announces New Planned Investments in Korean Startups at the MSIT-NVIDIA AI Startup Partnership Forum — NVIDIA Blog · partial-text ·
- Can DeepSeek challenge OpenAI, Anthropic in AI agents as Harness adoption takes off? — South China Morning Post · China Tech · feed-summary ·
- Bikers and Bug-Lovers Join Forces to Fight a New Google Data Center — WIRED AI · full-text ·
- ‘AI rebounding’: Nearly 20% of U.S. workers replaced by AI are being asked back—now they’re demanding higher pay, job security and AI transparency — Fortune · full-text ·
- Is Google Losing The AI Search Battle To Itself? — Forbes Innovation · feed-summary ·
- Korea Advances Physical AI Across Industries With NVIDIA — NVIDIA Robotics · full-text ·
- From Chips to Skills, NVIDIA and Korea Expand Sovereign AI Collaboration — NVIDIA Blog · partial-text ·
- AI-assisted learning platform helps robots reason beyond literal commands — MIT News Robotics · full-text ·
- Localizing and Editing the Goal Representation of a Vision-Language-Action Model — arXiv Robotics · partial-text ·
- NVIDIA Dynamo Recipes Bring Korea’s Leading Open Models to Production — NVIDIA Blog · partial-text ·
- Aspect-wise Pairwise Tournaments for Test-Time Selection — arXiv Artificial Intelligence · partial-text ·
- Lodos: A Lightweight Simulator Designed for Contact-Rich Interactions with Detailed Meshes — arXiv Robotics · partial-text ·
- VICTR: Value-Informed In-Context Learning for Vision-Language-Action Models — arXiv Robotics · partial-text ·
- ChainFlow: Few-Step Latent Flow Matching for Speculative Decoding — arXiv Artificial Intelligence · partial-text ·
- ForPC: Force-Aware Whole-Body Control for Precise Humanoid Cart Pushing — arXiv Robotics · partial-text ·
- Decision Context: Persistent Behavioral Learning for AI Agents — arXiv Artificial Intelligence · partial-text ·
- Meituan-Robotics-0: A Vision-Language-Action Foundation Model for Desktop Manipulation — arXiv Robotics · partial-text ·
- Conflux: Preventing Privilege Escalation with Principal Contexts and System-Level Evaluation — arXiv Cryptography and Security · partial-text ·
- NormGym: Evaluating Normative Competence through Social Interaction — arXiv Artificial Intelligence · partial-text ·
- A Cross-Domain TARA-to-NAC Rule Mapping Methodology: From Automotive OEM Networks to Energy Storage Systems — arXiv Cryptography and Security · partial-text ·
- Q-OIL: Optimistic Exploration in Learning from Human Interventions — arXiv Robotics · partial-text ·
- LLM-Based Feature Extraction for Multi-Criteria Inventory Classification — arXiv Artificial Intelligence · partial-text ·
- STRIDE: Learning Humanoid Locomotion in Minutes with Acceleration-Based Rewards — arXiv Robotics · partial-text ·
- ActPrune: Action-Grounded Visual Token Pruning for Efficient Vision-Language-Action Models — arXiv Robotics · partial-text ·
- Transporting the Past: An Optimal Transport View of Backtracking Counterfactuals — arXiv Artificial Intelligence · partial-text ·
- What Limits mmWave Human-State Sensing for Robot Perception? — arXiv Robotics · partial-text ·
- Nvidia CEO Jensen Huang admits to affair; board orders investigation — South China Morning Post · China Tech · feed-summary ·
- Anyone here using outdoor furniture knowledge for weatherproofing robot field enclosures? — Open Robotics Discourse · partial-text ·
- Observer: An Open Experiment in Interest-Driven Artificial World Modeling — Open Robotics Discourse · full-text ·
- How do robotics teams handle perception failures and training data? — Open Robotics Discourse · feed-summary ·
- Do you see anything in AMR telemetry before docking/localization failures? — Open Robotics Discourse · feed-summary ·
- Here’s the full text of Jensen Huang’s memo to Nvidia — The Verge · partial-text ·
- How a Frenchman’s book on ‘Democracy in America’ nearly 200 years ago explains the midterms, from data centers to Texas — Fortune · full-text ·
- The market bloodbath just exposed Wall Street’s true fears about U.S. debt, top economist warns — Fortune · full-text ·
- Apple is racing to make Siri more agentic — The Verge · partial-text ·
- Stocks will tumble much further as AI investment will slow next year then fall in 2028 and drag GDP down too, economist predicts — Fortune · full-text ·
- Scans for Sshole Vulnerability (CVE-2026-123158), (Sun, Nov 8th) — SANS Internet Storm Center · feed-summary ·
- Chatbots might be coming to Apple’s HomeView — The Verge · partial-text ·
- ARM Institute consortium grows to more than 500 organizations — The Robot Report · feed-summary ·