Research Center

• On 143 NEJM case conferences, the reasoning model gave the correct or very close diagnosis clearly more often than GPT-4 did.
• In the emergency-department test, the model's advantage was largest at triage, when information was scarcest; blinded reviewers could not reliably tell AI output from physician output.
• The authors stress that the model only read text written by others and call for prospective clinical trials before any deployment.

Hospitals in the Gulf are investing heavily in digital health, and this is the first peer-reviewed evidence in a top journal that a general-purpose reasoning model can rival specialists on real emergency cases. It argues for AI as a second opinion at the moments when clinicians have the least information, while making clear that trials, governance and physician oversight must come first.

• About 11.7% of US wage value, roughly 1.2 trillion dollars, sits in skills that existing AI tools can already perform, five times the visible exposure in technology jobs.
• Exposure is geographically broad: states such as Delaware, South Dakota, Ohio and Tennessee score as high as or higher than California.
• The skill-embedding approach predicted 85% of observed career transitions, giving policymakers a tool to test interventions before they are rolled out.

Gulf governments are planning national workforce strategies around AI, and this report offers a reusable method: measure exposure skill by skill, simulate the labour market, then test policy before acting. It shifts the conversation from fear of job losses to redeploying people from routine administrative tasks toward higher-value work, a priority for Saudi Vision 2030 and the UAE's AI strategy.

• US private AI investment was about 286 billion dollars in 2025, and industry produced more than 90% of the notable frontier models.
• Generative AI reached 53% adoption in three years, and around four in five university students now use it for coursework.
• The US-China gap on leading model benchmarks shrank to roughly 3%, while China leads in the number of public AI supercomputers.

For executives and policymakers in the Gulf, the AI Index is the most-cited neutral baseline for strategy documents, investment cases and regulation. Its 2026 numbers on adoption speed, compute concentration and public trust are directly useful when national AI programmes in the UAE and Saudi Arabia set targets or benchmark themselves against the United States, China and Europe.

• Models that score highly on medical benchmarks did not improve the decisions of 1,298 real users compared with ordinary web search.
• Standard benchmarks and even simulated AI patients failed to predict these failures, so they cannot substitute for testing with humans.
• The paper recommends systematic trials with real users before any public-facing health chatbot is deployed.

Health authorities and insurers in the Gulf are exploring AI symptom checkers in Arabic and English. This study is a caution: the limiting factor is the conversation between a worried person and the model, not the model's medical knowledge. Any deployment needs careful interface design, guidance on what to tell the system, and evaluation with real users rather than benchmark scores alone.

• A single neural network replaced the entire forecasting pipeline and still rivalled operational physics-based systems.
• A forecast took about one second on four GPUs, versus roughly a thousand node-hours for the equivalent conventional workflow.
• Because training is cheap and data-driven, the approach can be retargeted to specific regions or hazards, such as extreme heat or dust storms.

Forecasting at this cost opens weather services to organisations that could never run a supercomputer: utilities, airlines, ports, agriculture and real-estate developers planning for heat and dust. For Gulf countries, where extreme heat, flash floods and dust storms affect energy demand and construction, a locally fine-tuned model of this type is a realistic near-term tool.

• Safety for humanoids is defined as the ability to execute a safe stop with a given fallback policy, not as an abrupt power cut.
• A learned monitor with importance sampling needs fewer simulations and makes fewer false-safe predictions than uniform sampling.
• The monitor transferred from simulation to a real Unitree G1, pointing toward practical runtime safety certification.

Humanoid robots are moving from demonstrations into logistics and manufacturing, and Gulf ports, airports and giga-projects are early candidates. Before regulators and insurers allow them near people, someone must answer when a robot can be stopped safely. This work turns that question into a measurable, certifiable property, the kind of evidence procurement teams will soon require.

• Modelling evolutionary relationships explicitly beat much larger sequence-only models at predicting which genetic variants are harmful.
• Different evolutionary depths answer different questions: deep alignments for rare coding variants, primate alignments for complex traits.
• The model is cheap to train and comes with public genome-wide predictions to guide laboratory experiments.

Genomic medicine programmes in Saudi Arabia, the UAE and Qatar are sequencing large national cohorts. Tools that rank which variants matter, at modest compute cost, make that data actionable for diagnosing rare diseases and for population-specific research. GPN-Star also shows that smart use of scientific knowledge can beat brute-force scale, a lesson for AI investment generally.

• Fully open training data, code and checkpoints can deliver performance close to open-weight models from large companies.
• Compliance was built in from the start: public data only, retroactive opt-outs, filtering of personal data, and transparency aligned with the EU AI Act.
• Strong multilingual coverage shows that smaller languages need not be an afterthought in frontier-scale models.

Gulf states are building sovereign Arabic models such as Falcon, Jais and ALLaM. Apertus is the clearest public blueprint for doing that with transparent data governance, which matters for regulators, public-sector procurement and trust. It also demonstrates that a national supercomputer plus universities can produce a credible model without depending on foreign vendors.

• LLMs can serve as consultant, designer and simulator across the planning cycle rather than as a single tool.
• Agent-based simulations of residents reproduced real mobility hotspots and 15-minute accessibility metrics in two US cities.
• Quality urban design data, compute cost and bias are the main barriers, so the framework is meant to assist planners, not replace them.

Gulf cities are being planned at unprecedented speed, from new districts in Riyadh to the Dubai 2040 Urban Master Plan. A method that lets planners generate alternatives and test them with simulated residents before construction can cut cost and improve liveability. For developers and brokers, agent-based demand simulation is also a new way to assess where facilities, retail and housing will actually be used.

• Separating state (reasoned by an LLM) from appearance (rendered by diffusion) keeps simulations consistent over long horizons.
• Agents that use PAN to simulate outcomes before acting succeed more often than agents acting alone.
• The work places the UAE's MBZUAI among the few groups publishing general-purpose world models at the research frontier.

World models are the next layer under robotics, autonomous vehicles and simulation-based training, all priorities for the UAE's industrial strategy. PAN shows that a Gulf institution can contribute at the research frontier rather than only adopt imported systems, and it gives regional companies a documented architecture to build on for digital twins, logistics simulation and embodied AI.