Capacity development

Z.ai unveils cheaper, advanced AI model GLM-4.5

Chinese AI startup Z.ai, formerly Zhipu, is increasing pressure on global competitors with its latest model, GLM-4.5. The company has adopted an aggressive open-source strategy to attract developers. Anyone can download and use the model without licensing fees or platform restrictions.

GLM-4.5 is designed with agentic AI, breaking tasks into smaller components for improved performance. By approaching problems step by step, the model delivers more accurate and efficient outcomes. Z.ai aims to stand out through both technical sophistication and affordability.

CEO Zhang Peng says the model runs on only eight Nvidia H20 chips, while DeepSeek’s model needs sixteen. Nvidia developed the H20 to comply with US export controls aimed at China. Reducing chip demand significantly lowers the model’s operational footprint.

Zhang said the company has enough computing power and is not seeking further hardware now. Z.ai plans to charge 11 cents per million input tokens, undercutting DeepSeek R1’s 14 cents. Output tokens will cost 28 cents per million, compared to DeepSeek’s 2.19 dollars.

Such pricing could reshape large language model deployment expectations, especially in resource-limited environments. High costs have long been a barrier to broader AI adoption. Z.ai appears to be positioning itself as a more accessible alternative.

Founded in 2019, Z.ai has raised more than 1.5 billion dollars from investors including Alibaba, Tencent, and Qiming Venture Partners. It has grown quickly from a research-focused lab to one of China’s most prominent AI contenders. A public listing in Greater China is reportedly being prepared.

OpenAI recently named Zhipu among the Chinese firms it considers strategically significant in global AI development. US authorities responded by restricting American companies from working with Z.ai. The startup has nonetheless continued to expand its model lineup and partnerships.

Chinese firms increasingly invest in open-source models, often with domestic hardware compatibility in mind. Moonshot, another Alibaba-backed company, released the Kimi K2 model. Kimi K2 has received praise for its performance in coding and mathematical tasks.

Tencent has joined the race with its HunyuanWorld-1.0 model, which is built to generate immersive 3D environments. The HunyuanWorld-1.0 can accelerate game development, virtual reality design, and simulation work. Cutting-edge features are being paired with highly efficient architectures.

Alibaba also introduced its Qwen3-Coder model to assist in code generation and debugging. Such AI tools are seeing increasing use in software engineering and education. Chinese developers are positioning themselves to compete with Western offerings such as OpenAI’s Codex and Anthropic’s Claude.

The momentum within China’s AI sector is accelerating despite geopolitical and trade restrictions. A clear shift is underway from imitation to innovation, with local startups advancing independent research. Many models are trained on China-specific datasets to optimise relevance and performance.

Z.ai’s strategy combines cost reduction, efficient chip use, and broad availability. The company can build community trust and encourage ecosystem growth by open-sourcing its tools. At the same time, pricing undercuts major rivals and could disrupt the market.

Global AI development is increasingly decentralised, with Chinese firms no longer just playing catch-up. Large-scale funding and state support are helping to close gaps in hardware and training infrastructure. Z.ai is one of several firms pushing toward greater technological autonomy.

Open-source AI development is also helping Chinese companies win favour with developers outside their borders. Many international teams are experimenting with Chinese models to diversify risk and reduce reliance on US tech. Z.ai’s GLM-4.5 is among the models gaining traction globally.

By offering a powerful, lightweight, and affordable model, Z.ai is setting a new benchmark in the industry. The combination of technical refinement and strategic pricing draws attention from investors and users. A new era of AI competition is emerging.

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

Robot artist Ai-Da explores human self-perception

The world’s first ultra-realistic robot artist, Ai-Da, has been prompting profound questions about human-robot interactions, according to her creator.

Designed in Oxford by Aidan Meller, a modern and contemporary art specialist, and built in the UK by Engineered Arts, Ai-Da is a humanoid robot specifically engineered for artistic creation. She recently unveiled a portrait of King Charles III, adding to her notable portfolio.

Aidan Meller, Ai-Da’s creator, stated that working with the robot has evoked ‘lots of questions about our relationship with ourselves.’ He highlighted how Ai-Da’s artwork ‘drills into some of our time’s biggest concerns and thoughts.’

Ai-Da uses cameras in her eyes to capture images, which are then processed by AI algorithms and converted into real-time coordinates for her robotic arm, enabling her to paint and draw.

Mr Meller explained, ‘You can meet her, talk to her using her language model, and she can then paint and draw you from sight.’

He also observed that people’s preconceptions about robots are often outdated: ‘It’s not until you look a robot in the eye and they say your name that the reality of this new sci-fi world that we are now in takes hold.’

Ai-Da’s contributions to the art world continue to grow. She had produced and showcased her work at the AI for Good Global Summit 2024 in Geneva, Switzerland, an event under the auspices of the UN. That same year, her triptych of Enigma code-breaker Alan Turing sold for over £1 million at auction.

Her focus this year shifted to King Charles III, chosen because, as Mr Meller noted, ‘With extraordinary strides that are taking place in technology and again, always questioning our relationship to the environment, we felt that King Charles was an excellent subject.’

Buckingham Palace authorised the display of Ai-Da’s portrait of the King, despite the robot not meeting him. Ai-Da, connected to the internet, uses extensive data to inform her choice of subjects, with Mr Meller revealing, ‘Uncannily, and rather nerve-rackingly, we just ask her.’

The conversations generated inform the artwork. Ai-Da also painted a portrait of King Charles’s mother, Queen Elizabeth II, in 2023. Mr Meller shared that the most significant realisation from six years of working with Ai-Da was ‘not so much about how human she is but actually how robotic we are.’

He concluded, ‘We hope Ai-Da’s artwork can be a provocation for that discussion.’

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

Viasat launches global IoT satellite service

Viasat has unveiled a new global connectivity service designed to improve satellite-powered internet of things (IoT) communication, even in remote environments. The new offering, IoT Nano, supports industries like agriculture, mining, transport with reliable, low-data and low-power two-way messaging.

The service builds on Orbcomm’s upgraded OGx platform, delivering faster message speeds, greater data capacity and improved energy efficiency. It maintains compatibility with older systems while allowing for advanced use cases through larger messages and reduced power needs.

Executives at Viasat and Orbcomm believe IoT Nano opens up new opportunities by combining wider satellite coverage with smarter, more frequent data delivery. The service is part of Viasat’s broader effort to expand its scalable and energy-efficient satellite IoT portfolio.

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

Meta forms AI powerhouse by appointing Shengjia Zhao as chief scientist

Meta has appointed former OpenAI researcher Shengjia Zhao as Chief Scientist of its newly formed AI division, Meta Superintelligence Labs (MSL).

Zhao, known for his pivotal role in developing ChatGPT, GPT-4, and OpenAI’s first reasoning model, o1, will lead MSL’s research agenda under Alexandr Wang, the former CEO of Scale AI.

Mark Zuckerberg confirmed Zhao’s appointment, saying he had been leading scientific efforts from the start and co-founded the lab.

Meta has aggressively recruited top AI talent to build out MSL, including senior researchers from OpenAI, DeepMind, Apple, Anthropic, and its FAIR lab. Zhao’s presence helps balance the leadership team, as Wang lacks a formal research background.

Meta has reportedly offered massive compensation packages to lure experts, with Zuckerberg even contacting candidates personally and hosting them at his Lake Tahoe estate. MSL will focus on frontier AI, especially reasoning models, in which Meta currently trails competitors.

By 2026, MSL will gain access to Meta’s massive 1-gigawatt Prometheus cloud cluster in Ohio, designed to power large-scale AI training.

The investment and Meta’s parallel FAIR lab, led by Yann LeCun, signal the company’s multi-pronged strategy to catch up with OpenAI and Google in advanced AI research.

The collaboration dynamics between MSL, FAIR, and Meta’s generative AI unit remain unclear, but the company now boasts one of the strongest AI research teams in the industry.

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

UN urges global rules for AI to prevent inequality

According to Doreen Bogdan-Martin, head of the UN’s International Telecommunications Union, the world must urgently adopt a unified approach to AI regulation.

She warned that fragmented national strategies could deepen global inequalities and risk leaving billions excluded from the AI revolution.

Bogdan-Martin stressed that only a global framework can ensure AI benefits all of humanity instead of worsening digital divides.

With 85% of countries lacking national AI strategies and 2.6 billion people still offline, she argued that a coordinated effort is essential to bridge access gaps and prevent AI from becoming a tool that advances inequality rather than opportunity.

ITU chief highlighted the growing divide between regulatory models — from the EU’s strict governance and China’s centralised control to the US’s new deregulatory push under Donald Trump.

She avoided direct criticism of the US strategy but called for dialogue between all regions instead of fragmented policymaking.

Despite the rapid advances of AI in sectors like healthcare, agriculture and education, Bogdan-Martin warned that progress must be inclusive. She also urged more substantial efforts to bring women into AI and tech leadership, pointing to the continued gender imbalance in the sector.

As the first woman to lead ITU, she said her role was not just about achievement but setting a precedent for future generations.

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

New AI startup enables context across thousands of hours of video

Samsung Next has invested in Memories.ai, a startup specialising in long-duration video analysis capable of processing up to 10 million hours of footage.

The tool uses AI to transform massive video archives into searchable, structured datasets, even across multiple videos spanning hours or days.

The solution employs a layered pipeline: it filters noise, compresses critical segments, indexes content for natural-language queries, segments into meaningful units, and aggregates those insights into digestible reports. This structure enables users to search and analyse complex visual datasets seamlessly.

Memories.ai’s co-founders, Dr Shawn Shen and Enmin (Ben) Zhou, bring backgrounds from Meta’s Reality Labs and machine learning engineering.

The company raised $8 million in seed funding, surpassing its $4 million goal, led by Susa Ventures, including Samsung Next, Fusion Fund, Crane Ventures, Seedcamp, and Creator Ventures.

Samsung is banking on Memories.ai’s edge computing strengths, particularly to enable privacy-conscious applications such as home security analytics without cloud dependency. Its startup focus includes security firms and marketers needing scalable tools to sift through extensive video content.

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

Google launches AI feature to reshape how search results appear

Google has introduced a new experimental feature named Web Guide, aimed at reorganising search results by using AI to group information based on the query’s different aspects.

Available through Search Labs, the tool helps users explore topics in a more structured way instead of relying on the standard, linear results page.

Powered by Google’s Gemini AI, Web Guide works particularly well for open-ended or complex queries. For example, searches such as ‘how to solo travel in Japan’ would return results neatly arranged into guides, safety advice, or personal experiences instead of a simple list.

The feature handles multi-sentence questions, offering relevant answers broken into themed sections.

Users who opt in can access Web Guide via the Web tab and toggle it off without exiting the entire experiment. While it works only on that tab, Google plans to expand it to the broader ‘All’ tab in time.

The move follows Google’s broader push to incorporate Gemini into tools like AI Mode, Flow, and other experimental products.

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

Finnish researchers extend quantum coherence beyond one millisecond

Aalto University researchers have achieved a milestone in quantum computing by extending the coherence time of a superconducting transmon qubit beyond one millisecond. The breakthrough significantly improves how long quantum states remain stable, enabling more reliable operations.

The Finnish team used ultra-pure materials and precision engineering techniques to fabricate the qubit, achieving a median coherence time of 541 microseconds and a peak of 1057 microseconds.

Unlike earlier coherence records set by more exotic qubit types, Aalto’s success came within the widely used transmon framework, making it easier for others to replicate. The researchers published detailed fabrication methods to help advance consistency across labs globally.

Crossing the millisecond mark opens new possibilities for scalable quantum systems, reducing the burden of error correction. Finland’s growing leadership in the field is further solidified as other research teams explore Aalto’s reproducible approach.

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

Microsoft replaces the blue screen of death with a sleek black version in Windows 11

Microsoft has officially removed the infamous Blue Screen of Death (BSOD) from Windows 11 and replaced it with a sleeker, black version.

As part of the update KB5062660, the Black Screen of Death now appears briefly—around two seconds—before a restart, showing only a short error message without the sad face or QR code that became symbolic of Windows crashes.

The update, which brings systems to Build 26100.4770, is optional and must be installed manually through Windows Update or the Microsoft Update Catalogue.

It is available for both x64 and arm64 platforms. Microsoft plans to roll out the update more broadly in August 2025 as part of its Windows 11 24H2 feature preview.

In addition to the screen change, the update introduces ‘Recall’ for EU users, a tool designed to operate locally and allow users to block or turn off tracking across apps and websites. The feature aims to comply with European privacy rules while enhancing user control.

Also included is Quick Machine Recovery, which can identify and fix system-wide failures using the Windows Recovery Environment. If a device becomes unbootable, it can download a repair patch automatically to restore functionality instead of requiring manual intervention.

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

Meta boosts teen safety as it removes hundreds of thousands of harmful accounts

Meta has rolled out new safety tools to protect teenagers on Instagram and Facebook, including alerts about suspicious messages and a one-tap option to block or report harmful accounts.

The company said it is increasing efforts to prevent inappropriate contact from adults and has removed over 635,000 accounts that sexualised or targeted children under 13.

Of those accounts, 135,000 were caught posting sexualised comments, while another 500,000 were flagged for inappropriate interactions.

Meta said teen users blocked over one million accounts and reported another million after receiving in-app warnings encouraging them to stay cautious in private messages.

The company also uses AI to detect users lying about their age on Instagram. If flagged, those accounts are automatically converted to teen accounts with stronger privacy settings and messaging restrictions. Since 2024, all teen accounts are set to private by default.

Meta’s move comes as it faces mounting legal pressure from dozens of US states accusing the company of contributing to the youth mental health crisis by designing addictive features on Instagram and Facebook. Critics argue that more must be done to ensure safety instead of relying on user action alone.

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