Development

Hybrid AI could reshape robotics and defence

Investors and researchers are increasingly arguing that the future of AI lies beyond large language models. In London and across Europe, startups are developing so-called world models designed to simulate physical reality rather than simply predict text.

Unlike LLMs, which rely on static datasets, world models aim to build internal representations of cause and effect. Advocates say these systems are better suited to autonomous vehicles, robotics, defence and industrial simulation.

London based Stanhope AI is among companies pursuing this approach, claiming its systems learn by inference and continuously update their internal maps. The company is reportedly working with European governments and aerospace firms on AI drone applications.

Supporters argue that safety and explainability must be embedded from the outset, particularly under frameworks such as the EU AI Act. Investors suggest that hybrid systems combining LLMs with physics aware models could unlock large commercial markets across Europe.

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

Researchers tackle LLM regression with on policy training

Researchers at MIT, the Improbable AI Lab and ETH Zurich have proposed a fine tuning method to address catastrophic forgetting in large language models. The issue often causes models to lose earlier skills when trained on new tasks.

The technique, called self distillation fine tuning, allows a model to act as both teacher and student during training. In Cambridge and Zurich experiments, the approach preserved prior capabilities while improving accuracy on new tasks.

Enterprise teams often manage separate model variants to prevent regression, increasing operational complexity. The researchers argue that their method could reduce fragmentation and support continual learning, useful for AI, within a single production model.

However, the method requires around 2.5 times more computing power than standard supervised fine tuning. Analysts note that real world deployment will depend on governance controls, training costs and suitability for regulated industries.

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

Latam-GPT signals new AI ambition in Latin America

Chile has introduced Latam-GPT to strengthen Latin America’s presence in global AI.

The project, developed by the National Centre for Artificial Intelligence with support across South America, aims to correct long-standing biases by training systems on the region’s own data instead of material drawn mainly from the US or Europe.

President Gabriel Boric said the model will help maintain cultural identity and allow the region to take a more active role in technological development.

Latam-GPT is not designed as a conversational tool but rather as a vast dataset that serves as the foundation for future applications. More than eight terabytes of information have been collected, mainly in Spanish and Portuguese, with plans to add indigenous languages as the project expands.

The first version has been trained on Amazon Web Services. At the same time, future work will run on a new supercomputer at the University of Tarapacá, supported by millions of dollars in regional funding.

The model reflects growing interest among countries outside the major AI hubs of the US, China and Europe in developing their own technology instead of relying on foreign systems.

Researchers in Chile argue that global models often include Latin American data in tiny proportions, which can limit accurate representation. Despite questions about resources and scale, supporters believe Latam-GPT can deliver practical benefits tailored to local needs.

Early adoption is already underway, with the Chilean firm Digevo preparing customer service tools based on the model.

These systems will operate in regional languages and recognise local expressions, offering a more natural experience than products trained on data from other parts of the world.

Developers say the approach could reduce bias and promote more inclusive AI across the continent.

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

AI safety leader quits Anthropic with global risk warning

A prominent AI safety researcher has resigned from Anthropic, issuing a stark warning about global technological and societal risks. Mrinank Sharma announced his departure in a public letter, citing concerns spanning AI development, bioweapons, and broader geopolitical instability.

Sharma led AI safeguards research, including model alignment, bioterrorism risks, and human-AI behavioural dynamics. Despite praising his tenure, he said ethical tensions and pressures hindered the pursuit of long-term safety priorities.

His exit comes amid wider turbulence across the AI sector. Another researcher recently left OpenAI, raising concerns over the integration of advertising into chatbot environments and the psychological implications of increasingly human-like AI interactions.

Anthropic, founded by former OpenAI staff, balances commercial AI deployment with safety and risk mitigation. Sharma plans to return to the UK to study poetry, stepping back from AI research amid global uncertainty.

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

ChatGPT starts limited advertising rollout in the US

OpenAI has begun rolling out advertising inside ChatGPT, marking a shift for a service that has largely operated without traditional ads since its launch in 2022.

OpenAI said it is testing ads for logged-in Free and Go users in the United States, while paid tiers remain ad-free. The company said the test aims to fund broader access to advanced AI tools.

Ads appear outside ChatGPT responses and are clearly labelled as sponsored content, with no influence on answers. Placement is based on broad topics, with restrictions around sensitive areas such as health or politics.

Free users can opt out of ads by upgrading to a paid plan or by accepting fewer daily free messages in exchange for an ad-free experience. Users who allow ads can also opt out of ad personalisation, prevent past chats from being used for ad selection, and delete all ad-related history and data.

The rollout follows months of speculation after screenshots suggested that ads appeared in ChatGPT responses, which OpenAI described as suggestions. Rivals, including Anthropic, have contrasted their approach, promoting Claude as free from in-chat advertising.

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

Saudi Arabia recasts Vision 2030 with new priorities

The new phase of Vision 2030 is being steered toward technology, digital infrastructure and advanced industry by Saudi Arabia instead of relying on large urban construction schemes.

Officials highlight the need to support sectors that can accelerate innovation, strengthen data capabilities and expand the kingdom’s role in global tech development.

The move aligns with ongoing efforts to diversify the economy and build long-term competitiveness in areas such as smart manufacturing, logistics technology and clean energy systems.

Recent adjustments involve scaling back or rescheduling some giga projects so that investment can be channelled toward initiatives with strong digital and technological potential.

Elements of the NEOM programme have been revised, while funding attention is shifting to areas that enable automation, renewable technologies and high-value services.

Saudi Arabia aims to position Riyadh as a regional hub for research, emerging technologies and advanced industries. Officials stress that Vision 2030 remains active, yet its next stage will focus on projects that can accelerate technological adoption and strengthen economic resilience.

The Public Investment Fund continues to guide investment toward ecosystems that support innovation, including clean energy, digital infrastructure and international technology partnerships.

An approach that reflects earlier recommendations to match economic planning with evolving skills, future labour market needs and opportunities in fast-growing sectors.

Analysts note that the revised direction prioritises sustainable growth by expanding the kingdom’s participation in global technological development instead of relying mainly on construction-driven momentum.

Social and regulatory reforms connected to digital transformation also remain part of the Vision 2030 agenda. Investments in training, digital literacy and workforce development are intended to ensure that young people can participate fully in the technology sectors the kingdom is prioritising.

With such a shift, the government seeks to balance long-term economic diversification with practical technological goals that reinforce innovation and strengthen the country’s competitive position.

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

New AI system forecasts mobility after joint replacement

AI is being deployed to forecast how well patients regain mobility after hip replacement surgery, offering new precision in orthopaedic recovery planning.

Researchers at the Karlsruhe Institute of Technology developed a model capable of analysing complex gait biomechanics to assess post-operative walking outcomes.

Hip osteoarthritis remains one of the leading drivers of joint replacement procedures, with around 200,000 artificial hips implanted in Germany in 2024 alone. Recovery varies widely, driving research into tools predicting post-surgery mobility and pain relief.

Movement data collected before and after operations were analysed using AI as part of a joint project with the Universitätsmedizin Frankfurt.

The system examined biomechanical indicators, including joint angles and loading patterns, enabling researchers to classify patients into three distinct gait recovery groups.

Results show the model can predict who regains near-normal walking and who needs intensive rehabilitation. Researchers say the framework could guide personalised therapy and expand to other joints and musculoskeletal disorders.

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

AI model achieves accurate detection of placenta accreta spectrum in high-risk pregnancies

A new AI model has shown strong potential for detecting placenta accreta spectrum, a dangerous condition that often goes undiagnosed during pregnancy.

Researchers presented the findings at the annual meeting of the Society for Maternal-Fetal Medicine, highlighting that traditional screening identifies only about half of all cases.

Placenta accreta spectrum arises when the placenta attaches abnormally to the uterine wall, often after previous surgical procedures such as caesarean delivery.

The condition can trigger severe haemorrhage, organ failure, and death, yet many pregnancies with elevated risk receive inconclusive or incorrect assessments through standard ultrasound examinations.

A study that involved a retrospective review by specialists at the Baylor College of Medicine, who analysed 2D obstetric ultrasound images from 113 high-risk pregnancies managed at the Texas Children’s Hospital between 2018 and 2025.

The AI system detected every confirmed case of placenta accreta spectrum, produced two false positives, and generated no false negatives.

Researchers believe such technology could significantly improve early identification and clinical preparation.

They argue that AI screening, when used in addition to current methods, may reduce maternal complications and support safer outcomes for patients facing this increasingly common condition.

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

Enterprise AI adoption stalls despite heavy investment

AI has moved from experimentation to expectation, yet many enterprise AI rollouts continue to stall. Boards demand returns, leaders approve tools and governance, but day-to-day workarounds spread, risk grows, and promised value fails to materialise.

The problem rarely lies with the technology itself. Adoption breaks down when AI is treated as an IT deployment rather than an internal product, leaving employees with approved tools but no clear value proposition, limited capacity, and governance that prioritises control over learning.

A global B2B services firm experienced this pattern during an eight-month enterprise AI rollout across commercial teams. Usage dashboards showed activity, but approved platforms failed to align with actual workflows, leading teams to comply superficially or rely on external tools under delivery pressure.

The experience exposed what some leaders describe as the ‘mandate trap’, where adoption is ordered from the top while usability problems fall with middle managers. Hesitation reflected workflow friction and risk rather than resistance, revealing an internal product–market fit issue.

Progress followed when leaders paused broad deployment and refocused on outcomes, workflow redesign, and protected learning time. Narrow pilots and employee-led enterprise AI testing helped scale only tools that reduced friction and earned trust.

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

LegalOn launches agentic AI for in-house legal teams

LegalOn Technologies has introduced five agentic AI tools aimed at transforming in-house legal operations. The company says the agents complete specialised contract and workflow tasks in seconds within its secure platform.

Unlike conventional AI assistants that respond to prompts, the new system is designed to plan and execute multi-step workflows independently, tailoring outputs to each organisation’s templates and standards while keeping lawyers informed of every action.

The suite includes tools for generating playbooks, processing legal intake requests and translating contracts across dozens of languages. Additional agents triage high-volume agreements and produce review-ready drafts from clause libraries and deal inputs.

Founded by two corporate lawyers in Japan, LegalOn now operates across Asia, Europe and North America. Backed by $200m in funding, it serves more than 8,000 clients globally, including Fortune 500 companies.

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