Apple negotiates agreements with media outlets to integrate journalistic content into AI Siri

Researchers conducted one of the first independent stress tests for GlobalBuildingAtlas (GBA)—a large-scale new digital dataset aimed at mapping every building on the planet. The study, focusing on India's rapidly growing urban landscapes, found that while GBA demonstrates high accuracy in determining building locations on the ground, it systematically mismeasures their height, often underestimating the actual height of skyscrapers by more than half.

This finding is critically important for urban planning specialists and climatologists who rely on three-dimensional models to predict parameters such as energy consumption or floodwater movement in a city.

The work was carried out by a team from the National Remote Sensing Centre (NRSC) at ISRO. Researchers examined various urban areas, including Mumbai, Bangalore, and Hyderabad. As India experiences unprecedented urbanization, having a digital map of its cities has become a necessity for effective resource management.

GlobalBuildingAtlas was recently released as an open-access resource containing over 2.75 billion building markers worldwide. It is the first such resource to offer three-dimensional data at this scale, providing not only building outlines but also approximate heights and simplified 3D shapes. However, since these maps are created using artificial intelligence and satellite imagery, their accuracy can vary significantly depending on local conditions.

To verify GBA, researchers used the NASA ICESat-2 satellite as a space standard. This spacecraft employs a laser altimeter that sends trillions of photons to the Earth's surface, measuring their return time to determine height with centimeter accuracy. The team compared these laser measurements with GBA's height estimates for 20 representative structures, ranging from low-rise industrial warehouses to tall residential towers.

To assess horizontal accuracy, or outline accuracy, the team used high-resolution imagery from ISRO's Bhuvan portal, a remote sensing data visualization platform. These images were processed by the AI model, Segment Anything Model 2, to provide a ground truth.

The results for horizontal mapping were impressive. In planned residential communities across five major Indian cities, GlobalBuildingAtlas was nearly perfect, achieving a completeness coefficient of over 99%. In cities like Bangalore and Calcutta, the digital map matched the actual building. This indicates that GBA is an exceptionally reliable tool for understanding neighborhood density or the scale of urban sprawl.

However, the vertical results told a different story. Researchers discovered a systematic negative bias, meaning GBA almost always underestimated building heights. Errors became more significant as building heights increased. For example, for a 266-meter skyscraper in Mumbai, GBA estimated the height at only 113 meters, representing a discrepancy of over 150 meters. For mid-rise buildings, errors often exceeded 40%.

GBA uses monocular optical imaging, meaning it estimates height from two-dimensional photographs rather than using direct three-dimensional measurements. Although AI can analyze shadows and perspective to estimate height, it often gets confused due to complex roof shapes, nearby trees, or the close proximity of buildings in densely populated Indian cities.

Previous global datasets, such as those released by Microsoft and Google, primarily focused on two-dimensional outlines. GBA represents an ambitious attempt to add the third dimension—height—which is crucial for modern sustainable development goals. However, the ISRO study highlights a serious limitation. The AI models used to create GBA were largely trained on data from Europe and North America. Since architectural styles, building materials, and urban density in the Global South, especially in India, differ greatly, the AI struggles to correctly interpret the scenes.

Furthermore, GBA uses a simplified three-dimensional representation known as Level of Detail 1, which treats each building as a flat block, ignoring complex spires, pitched roofs, and helipads that characterize many Indian landmarks. The study concludes that while GlobalBuildingAtlas is a valuable resource for understanding the horizontal layout of our cities, it should be used with extreme caution when vertical accuracy is required. This work serves as a vital warning to society, as when designing smart cities, planning disaster response, or calculating rooftop solar energy, it is necessary to ensure that the data reflects reality. By identifying these height errors, the ISRO team provided a roadmap for future improvements, suggesting that integrating more laser data from satellites like ICESat-2 could eventually lead to a digital world as tall and complex as our own.

Google has implemented an extensive restructuring of its artificial intelligence division, relocating parts of the Google DeepMind teams to the corporate structure and concentrating strategic decisions under the leadership of Koray Kavukcuoglu. This information was reported by Reuters on Wednesday, the 12th.

This change occurs amid intense competition, where Google seeks to regain lost ground against competitors such as Anthropic and OpenAI. Furthermore, the move reduces the historical autonomy of the lab, established after its acquisition in 2014, aligning Gemini's development more closely with the company's commercial objectives.

In parallel, Sergey Brin resumed his influence on model training efforts, advocating for greater acceleration to reach technological frontiers. However, Google faces challenges, including delays in the launch of the next major version of Gemini and restrictions on available computational capacity.

The transformation gained public visibility on August 5th, the date Google announced changes in the management of Google DeepMind. Demis Hassabis stepped down from the executive leadership of the lab and assumed the presidency, while his successor, Koray Kavukcuoglu, took on the central role in the unit's direction.

This new setup formalized a power transition that was already underway. Kavukcuoglu had held the position of Google's principal artificial intelligence architect since 2025, a role created specifically to expand Gemini's presence in the company's products. Sources close to the operation informed Reuters that his influence over the model's direction has grown significantly since then.

Kavukcuoglu's position also brought him closer to the divisions responsible for AI revenue generation. Even before moving from London to Mountain View, he was seen as the main link between DeepMind and Google Cloud, a sector considered vital to Google's AI business.

The new arrangement impacted other prominent names. Jeff Dean and Oriol Vinyals, who were part of the original Gemini technical team, left the company to join a new startup. The departure of these executives raised questions about the future of groups that had greater freedom to conduct research outside of direct model training.

During a general meeting held on August 6th, employees were notified that certain non-technical teams would be removed from the DeepMind structure, reporting directly to the corporate organization of Google. This measure reinforced among lab members the feeling that their independence was diminishing.

Management attempted to calm these concerns, assuring that the lab's daily activities would remain essentially unchanged. Hassabis would continue to focus on long-term issues related to the science and social impacts of artificial intelligence, while Kavukcuoglu would maintain control over Gemini.

However, this change comes amidst growing pressure for results. According to reports gathered by Reuters, internal tests forced Google to postpone the arrival of the next major version of Gemini by two months due to lower results compared to those achieved by competitors in areas such as programming.

Recent history contributed to this pressure. Gemini briefly led among competing models after an update released in November 2025, but subsequently lost ground against advances made by rival companies.

Sources involved in development attributed part of this difficulty to structural problems. Limitations in computational capacity availability and disagreements among project leaders hindered the concentration of resources in crucial areas, such as programming. Internal bureaucracy was also cited as a factor contributing to a slower release pace than that of more agile competitors.

In this scenario, Brin informally intervened with the teams responsible for model training. According to someone familiar with his involvement, the co-founder spent months pressuring employees for a faster approach to state-of-the-art systems.

Brin's influence also extended to resource allocation. Among the topics that received his attention was the concept of recursive self-improvement, linked to the possibility of AI systems improving their own capabilities without human intervention.

Although he no longer holds an executive position, Brin maintains influence due to his status as a co-founder. Reuters mentioned that he also adopted a posture of internal mobilization after the launch of ChatGPT in 2022, when OpenAI's progress generated a strong internal reaction at Google.

For his part, Hassabis had reduced his direct involvement in Gemini's management after assisting with the model's launch in 2023. Sources cited by Reuters indicated that development responsibilities were progressively transferred to his subordinates.

The concentration of authority in Kavukcuoglu may also influence the dynamic between DeepMind and Google Cloud. A report obtained by Reuters suggests that cloud computing executives expect the change to minimize conflicts related to the distribution of artificial intelligence chips, known as TPUs.

Thus, the new design positions model strategy, Gemini's integration into products, and commercial requirements within a structure closer to Google headquarters. A company spokesperson confirmed to Reuters that Kavukcuoglu will have the final say on important decisions involving DeepMind.