Key Internet developments

Vietnam expands 5G and strengthens digital trust

Vietnam’s 5G network now reaches approximately 59 percent of the population, slightly over one year after commercial services launched in October 2024.

Government data presented at Internet Day 2025 show that Vietnam ranks 10th globally for fixed broadband speed and 15th for mobile broadband, reflecting rapid improvements in national connectivity.

Officials described the Internet as a second living space for citizens, with nearly 80 million users spending an average of seven hours online each day for work, education and social interaction.

Authorities highlighted that expanded 5G coverage supports the development of a digital economy, e-government services and a more connected digital society.

Alongside infrastructure growth, policymakers stressed the need for stronger digital trust.

Vietnam is shifting towards clearer legal frameworks instead of reliance on voluntary self-regulation, while prioritising cybersecurity, data governance and protection against online fraud, deepfakes and AI-driven deception to sustain long-term digital transformation.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

Study warns that LLMs are vulnerable to minimal tampering

Researchers from Anthropic, the UK AI Security Institute and the Alan Turing Institute have shown that only a few hundred crafted samples can poison LLM models. The tests revealed that around 250 malicious entries could embed a backdoor that triggers gibberish responses when a specific phrase appears.

Models ranging from 600 million to 13 billion parameters (such as Pythia) were affected, highlighting the scale-independent nature of the weakness. A planted phrase such as ‘sudo’ caused output collapse, raising concerns about targeted disruption and the ease of manipulating widely trained systems.

Security specialists note that denial-of-service effects are worrying, yet deceptive outputs pose far greater risk. Prior studies already demonstrated that medical and safety-critical models can be destabilised by tiny quantities of misleading data, heightening the urgency for robust dataset controls.

Researchers warn that open ecosystems and scraped corpora make silent data poisoning increasingly feasible. Developers are urged to adopt stronger provenance checks and continuous auditing, as reliance on LLMs continues to expand for AI purposes across technical and everyday applications.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot! 

Tiiny AI unveils the Pocket Lab supercomputer

Tiiny AI has revealed the Pocket Lab, a palm-sized device recognised as the world’s smallest personal AI supercomputer. Guinness World Records confirmed the title, noting its ability to run models with up to 120 billion parameters.

The Pocket Lab uses an ARM v9.2 CPU, a discrete NPU delivering 190 TOPS and 80GB of LPDDR5X memory. Popular open-source models such as GPT-OSS, Llama, Qwen, Mistral, DeepSeek and Phi are supported. Tiiny AI says its hardware makes large-scale reasoning possible in a handheld format.

Two in-house technologies enhance efficiency by distributing workloads and reducing unnecessary activations. TurboSparse manages sparse neuron activity to preserve capability while improving speed, and PowerInfer splits computation across the CPU and NPU.

Tiiny AI plans a full showcase at CES 2026, with pricing and release information still pending. Analysts want to see how the device performs in real-world tasks compared with much larger systems. The company believes the Pocket Lab will shift expectations for personal AI hardware.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot! 

AI use grows among EU enterprises in 2025

In 2025, one in five EU enterprises with at least ten employees reported using AI technologies, marking a significant rise from 13.5% in 2024. AI adoption has more than doubled since 2021, showing its increasing use in business across the EU.

Nordic countries led the way, with Denmark at 42%, Finland at 37.8%, and Sweden at 35%. In contrast, Romania, Poland, and Bulgaria had the lowest adoption rates, ranging from 5.2% to 8.5%.

Almost all EU member states recorded increases compared with the previous year, with Denmark, Finland, and Lithuania showing the most significant gains.

Enterprises mainly used AI to analyse text, generate multimedia, produce language, and convert speech into machine-readable formats. Analysing written language saw the most significant growth in 2025, followed by content generation, highlighting AI’s expanding role in communication and data processing.

Rising AI adoption is also linked to efficiency gains and innovation across EU businesses. Companies report using AI to streamline operations, support decision-making, and enhance customer engagement, signalling broader economic and technological impacts.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

OpenAI launches GPT‑5.2 for professional knowledge work

OpenAI has introduced GPT‑5.2, its most advanced model series to date, designed to enhance professional knowledge work. Users report significant time savings, with daily reductions of 40-60 minutes and more than 10 hours per week for heavy users.

The new model excels at generating spreadsheets, presentations, and code, while also handling complex, multi-step projects with improved speed and accuracy.

Performance benchmarks show GPT‑5.2 surpasses industry professionals on GDPval tasks across 44 occupations, producing outputs over eleven times faster and at a fraction of the cost.

Coding abilities have also reached a new standard, encompassing debugging, refactoring, front-end UI work, and multi-language software engineering tasks, providing engineers with a more reliable daily assistant.

GPT‑5.2 Thinking improves long-context reasoning, vision, and tool-calling capabilities. It accurately interprets long documents, charts, and graphical interfaces while coordinating multi-agent workflows.

The model also demonstrates enhanced factual accuracy and fewer hallucinations, making it more dependable for research, analysis, and decision-making.

The rollout includes ChatGPT Instant, Thinking, and Pro plans, as well as API access for developers. Early tests show GPT‑5.2 accelerates research, solves complex problems, and improves professional workflows, setting a new benchmark for real-world AI tasks.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

US approaches universal 5G as global adoption surges

New data from Omdia and 5G Americas showed rapid global growth in wireless connectivity during the third quarter of 2025, with nearly three billion 5G connections worldwide.

North America remained the most advanced region in terms of adoption, reaching penetration levels that almost match its population.

The US alone recorded 341 million 5G connections, marking one of the highest per capita adoption rates in the world, compared to the global average, which remains far lower.

Analysts noted that strong device availability and sustained investment continue to reinforce the region’s leadership. Enhanced features such as improved uplink performance and integrated sensing are expected to accelerate the shift towards early 5G-Advanced capabilities.

Growth in cellular IoT also remained robust. North America supported more than 270 million connected devices and is forecast to reach nearly half a billion by 2030 as sectors such as manufacturing and utilities expand their use of connected systems.

AI is becoming central to these deployments by managing traffic, automating operations and enabling more innovative industrial applications.

Future adoption is set to intensify as regional 5G connections are projected to surpass 8.6 billion by 2030.

Rising interest in fixed wireless access is driving multi-device usage, offering high-speed connectivity for households and small firms instead of relying solely on fibre networks that remain patchy in many areas.

Globally, the sector has reached more than 78 million connections, with strong annual growth. Analysts believe that expanding infrastructure will support demand for low-latency connectivity, and the addition of satellite-based systems is expected to extend coverage to remote locations.

By mid-November 2025, operators had launched 379 commercial 5G networks worldwide, including seventeen in North America. A similar number of LTE networks operated across the region.

Industry observers said that expanding terrestrial and non-terrestrial networks will form a layered architecture that strengthens resilience, supports emergency response and improves service continuity across land, sea and air.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

India expands 5G coverage nationwide

The 5G footprint of India has expanded across all states and union territories, with services now reaching 99.9% of districts.

Telecom service providers have installed more than 5 lakh 5G base stations, contributing to a broader network of over 31 lakh sites nationwide. The government has emphasised the importance of reliable coverage in both rural and urban regions.

Efforts to strengthen connectivity in underserved areas continue through BharatNet, new mobile services in Left Wing Extremism (LWE)-affected regions, support for Aspirational Districts and the 4G Saturation Scheme, which aims to bring coverage to every uncovered village.

Streamlined Right of Way rules and faster approval for using street furniture have created an environment where operators can deploy small cells more efficiently, rather than facing long administrative delays.

As 5G coverage expands nationwide, operators are preparing for a future driven by AI workloads rather than traditional network demands. Private and state providers are both expanding infrastructure, often sharing facilities when it is technically and commercially viable.

The government of India has emphasised that coordinated planning will be crucial for managing the increasing digital traffic.

Seven working groups under the Bharat 6G Alliance have outlined progress on India’s next-generation roadmap. The communications minister has emphasised that spectrum policy, device readiness, applications and sustainability need to align so innovation can mature.

Monthly joint reviews will be conducted to ensure that breakthroughs in one domain can lead to practical outcomes in other areas, supporting India’s long-term 6G strategy.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

India moves toward mandatory AI royalty regime

India is weighing a sweeping copyright framework that would require AI companies to pay royalties for training on copyrighted works under a mandatory blanket licence branded as the hybrid ‘One Nation, One Licence, One Payment’ model.

A new Copyright Royalties Collective for AI Training, or CRCAT, would collect payments from developers and distribute money to creators. AI firms would have to rely only on lawfully accessed material and file detailed summaries of training datasets, including data types and sources.

The panel is expected to favour flat, revenue-linked percentages on global earnings from commercial AI systems, reviewed roughly every three years and open to legal challenge in court.

Obligations would apply retroactively to AI developers that have already trained profitable models on copyright-protected material, framed by Indian policymakers as a corrective measure for the creative ecosystem.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot! 

Multimodal AI reveals new immune patterns across cancer types

A recent study examined the capabilities of GigaTIME, a multimodal AI framework that models the tumour immune microenvironment by converting routine H and E slides into virtual multiplex immunofluorescence images.

Researchers aimed to solve long-standing challenges in profiling tumour ecosystems by using a scalable and inexpensive technique instead of laboratory methods that require multiple samples and extensive resources.

The study focused on how large image datasets could reveal patterns of protein activity that shape cancer progression and therapeutic response.

GigaTIME was trained on millions of matched cells and applied to more than fourteen thousand slides drawn from a wide clinical network. The system generated nearly 300.000 virtual images and uncovered over 1000 associations between protein channels and clinical biomarkers.

Spatial features such as sharpness, entropy and signal variability were often more informative than density alone, revealing immune interactions that differ strongly across cancer types.

When tested on external tumour collections, the framework maintained strong performance and consistently exceeded the results of comparator models.

The study reported that GigaTIME could identify patterns linked to tumour invasion, survival and stage. Protein combinations offered a clearer view of immune behaviour than single markers, and the virtual signatures aligned with known and emerging genomic alterations.

Certain proteins were easier to infer than others, which reflected structural differences at the cellular level rather than model limitations. The research also suggested that immune evasion mechanisms may shift during advanced disease, altering how proteins such as PD-L1 contribute to tumour progression.

The authors argued that virtual multiplex imaging could expand access to spatial proteomics for both research and clinical practice.

Wider demographic representation and broader protein coverage are necessary for future development, yet the approach demonstrated clear potential to support large population studies instead of the restricted datasets produced through traditional staining methods.

Continued work seeks to build a comprehensive atlas and refine cell-level segmentation to deepen understanding of immune and tumour interactions.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!

AI model predicts prediabetes with high accuracy

Researchers have developed an AI model that enhances prediabetes prediction by integrating oxidative stress markers with traditional clinical indicators. The Pattern Neural Network model achieved 98.3% accuracy in Indian adults, outperforming other machine learning methods.

Total antioxidant status emerged as a key predictor, with lower antioxidant capacity observed in individuals with prediabetes. Waist circumference and BMI were also highly informative, alongside glucose markers such as HbA1c and OGTT.

The inclusion of oxidative stress measures provides a deeper understanding of the mechanisms underlying metabolic risk.

The study used clinical and biochemical data from 199 adults, with the PNN trained on 14 features, including demographic and biochemical variables. High accuracy across all sets indicates strong potential for quick, low-cost screening and personalised early interventions.

While the results are promising, the single-centre design and limited sample size indicate that external validation is needed. Future studies should test the model in larger, multi-site cohorts and integrate longitudinal data to enhance its real-world applicability and public health impact.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!