Nvidia has introduced its personal AI supercomputer, Project Digits, at CES. This device aims to provide AI researchers, data scientists, and students worldwide with access to the power of the Grace-Blackwell platform. Project Digits uses the new Nvidia GB10 Grace Blackwell Superchip, developed in collaboration with MediaTek. It offers a petaflop of AI computing power for prototyping, fine-tuning, and executing large AI models with up to 200 billion parameters. Users can develop models on their own desktop systems and then seamlessly deploy them on accelerated cloud or data center infrastructure. The system will be available from May at Nvidia and partners, starting at $3,000.
Nvidia is also accelerating AI agent development with pre-made blueprints. These blueprints are prefabricated instructions with all necessary components and configurations for developing AI agents that can independently plan and execute complex tasks. Five AI companies have already developed their own specialized blueprints for applications like automatic code documentation, language AI agents, structured reporting, research assistance, and virtual AI assistants. Nvidia offers blueprints for converting PDFs into podcasts and for video analysis. The technical foundation is the new Nemotron models for text processing and image analysis in three sizes for different use cases. Implementation is done through Nvidia AI Enterprise Software, which can be used in data centers and cloud services.
Additionally, Nvidia has introduced video-based “World Models” on the Cosmos platform. These models aim to generate photorealistic training data for robots and autonomous vehicles without expensive real-world tests. The World Foundation Models were trained with 9,000 trillion tokens from 20 million hours of video material. They are designed to create physics-based videos from various inputs such as text, image, video, and robot sensor or motion data. Initial users include ride-hailing company Uber and robotics companies and developers of autonomous vehicles. However, there is criticism of the concept of video generators as world models, as current models reportedly do not understand universal physical laws.
Microsoft plans to invest $80 billion in AI data centers this year. Microsoft’s Vice President Brad Smith has outlined a multi-year plan for building data centers for artificial intelligence. More than half of the AI budget, totaling $80 billion, will be invested in facilities in the USA. Smith emphasizes that such large investments affect not only the high-tech industry but also a variety of sectors such as construction, energy supply, and other infrastructure providers. Microsoft aims to differentiate its “American AI” from China. In addition to data centers, the development of applications with AI and export to “allies and friends” are important. Smith does not specify concrete locations for new data centers. He also does not address the increasingly controversial energy demand for AI in his blog post.
OpenAI CEO Sam Altman reports unexpected losses with the premium service ChatGPT Pro, which costs $200 per month. The main reason is the unexpectedly intense use of the service, which provides access to the Sora video generator, among other things. OpenAI’s financial challenges are multifaceted: high development costs for new AI models, hardware, personnel, and energy costs strain the budget. Training GPT-4 alone is said to have cost around $80 million in 2023. The company offers various subscription tiers: free access with GPT-4o mini, a $20 subscription with extended access, and the $200 Pro subscription with unlimited access to all models and 500 Sora videos monthly. In a blog post, Altman also reflects on the challenges of building the company while reaffirming the vision of superintelligence.
In Berlin, Justice Senator Felor Badenberg believes that the use of artificial intelligence in courts is necessary. Two projects are currently being tested, including an AI-supported research tool for asylum procedures, which is to be tested from 2025. The program aims to compile all relevant information about the situation in the respective country of origin, which is necessary for a decision in the asylum procedure. Until now, courts have compiled this information themselves, which is very time-consuming. Badenberg is convinced that Berlin courts will manage the transition to electronic files by early 2026. The transition is considered challenging for criminal proceedings. By the deadline of January 1, 2026, a total of twelve courts with more than 3,000 employees are to be converted in Berlin.
Apple’s in-house AI solution “Apple Intelligence” has seen an increase in memory requirements from the original 4 GB to now 7 GB. This development is particularly relevant as Apple operates its AI models directly on devices rather than in the cloud—a strategy the company justifies with advantages in privacy and independence. The launch in Germany and other EU countries is planned for April 2025, while usage in English is already possible with a workaround. With iOS 18.2, new features were introduced, including the “Image Playground” app for comic-style image generation and “Genmojis” for custom emojis. Further improvements, especially for Siri, are still pending and are expected with iOS 18.4. A complete overhaul of the Siri system based on language models is planned for iOS 19 in the summer.
Researchers provide another reason for skepticism about AI benchmarks. An independent investigation shows that OpenAI’s latest language model o1 performs significantly worse on programming tasks than stated. In the “SWE-Bench Verified” benchmark, it solves only about 30% of the tasks set—significantly less than the nearly 49% claimed by the company. The large discrepancy between test results is due to different methodologies: OpenAI used a program that gives the AI very tight guidelines, while the researcher used one that gives the AI more freedom in problem-solving. The case once again shows that the results of AI benchmarks depend on many factors, and it is difficult for outsiders to assess the actual performance of an AI system, let alone its practical utility.