AI models

Google’s Private AI Compute is a new cloud-based system aiming to balance powerful AI with user privacy by replicating on-device data security within a cloud environment. Similar to Apple’s approach, it addresses the challenge of providing computationally intensive AI while protecting data confidentiality. The system uses Google’s infrastructure, including TPUs and TIEs, encrypted connections, and Zero Access assurance to secure data processing. It enhances features like Magic Cue and Recorder, offering faster, more personalized results. Google intends Private AI Compute to unlock a new generation of privacy-centric AI tools.

Despite Meta’s planned $14.3B investment interpreted by some as an acquihire and subsequent paused collaborations with OpenAI, Google, and xAI, Scale AI’s CFO Dennis Cinelli insists the company is thriving. He claims recent significant deals and growth in both its data and applications businesses, including contracts with the U.S. Department of Defense, have resulted in revenue “well into the nine figures,” approaching $1 billion. Scale AI is expanding and hiring, aiming to solidify its position in the AI development landscape.

OpenAI has signed a $38 billion deal with Amazon Web Services (AWS) for cloud capacity, marking its first major agreement with AWS and a move towards independence from Microsoft. The deal involves deploying workloads across AWS infrastructure, leveraging Nvidia GPUs, and expanding capacity. This collaboration boosted Amazon’s stock. OpenAI’s CEO, Sam Altman, emphasizes the need for massive compute power. While committed to Microsoft Azure, OpenAI’s AWS partnership signifies a strategic diversification. Amazon is also heavily invested in Anthropic, highlighting its commitment to AI. The agreement supports both AI model training and inference.

OpenAI has released open-weight AI safety models designed for developers to identify and mitigate risks like bias and toxicity. This shift towards transparency aims to foster collaboration and accelerate innovation in AI safety. By providing accessible tools, OpenAI encourages a broader community to contribute to and improve AI safety best practices. This move addresses increasing pressure for transparency and allows for external audits, while also potentially building a larger community. The success will depend on data quality and developer proficiency, but signifies a commitment to a more responsible AI future.

Anthropic is aggressively expanding its global enterprise AI footprint, driven by surging international demand for its Claude models. The company plans to triple its international workforce and quintuple its applied AI team in 2025. Nearly 80% of Claude’s activity originates outside the US, with adoption rates exceeding US levels in countries like South Korea. Anthropic is establishing its first Asia office in Tokyo and scaling operations throughout Europe, focusing on industry-specific solutions and data sovereignty. This expansion intensifies competition with OpenAI, Microsoft, and Google.

OpenAI and Nvidia are reportedly discussing a potential $100 billion partnership, with Nvidia supplying at least 10 gigawatts of hardware and investing significantly in OpenAI. This collaboration aims to bolster OpenAI’s AI infrastructure for advanced model training, utilizing Nvidia’s Vera Rubin platform starting in 2026. The deal raises concerns about competition, potentially solidifying Nvidia and OpenAI’s dominance. OpenAI seeks to secure computational resources crucial for AI development, while also exploring custom chip solutions. The partnership is under scrutiny for potential circular funding and antitrust implications.

Cambricon, a Chinese AI chip company, has received approval from the Shanghai Stock Exchange for a RMB 4 billion private placement. The funds will be used for AI chip and software platform development, particularly for large-scale AI models, and to supplement working capital. Cambricon’s Q1 2025 revenue surged by 4230.22% year-on-year, with net profit also significantly increasing. The company focuses on cloud, edge products, and IP licensing, aiming to build a comprehensive AI ecosystem.

At the 2025 World Robot Conference, Unitree Robotics CEO Wang Xingxing emphasized the need for robust and unified “embodied intelligence robot foundation models,” citing them as the industry’s core bottleneck. He believes current robotic model architectures are lacking, and the industry needs a “ChatGPT moment.” Unitree is focused on developing general-purpose humanoid robots for diverse tasks and sees unified models, affordable hardware, high-volume manufacturing, and accessible computational power as key priorities for the next 2-5 years.

Seven leading AI models were tested on the challenging 2025 Beijing Junior High School Entrance Examination, focusing on mathematics, Chinese essay, and English essay. While models excelled in essay writing, particularly in Chinese, and showed strong mathematical reasoning, particularly with LaTeX input, challenges remained in image interpretation for math problems. iFlytek Spark and Baidu Wenxin Yiyan performed particularly well across subjects, demonstrating AI’s growing capabilities in academic assessments.

Google launched the AI Edge Gallery, an app enabling on-device AI on smartphones. Sourced from Hugging Face, it allows users to run AI models offline with features like image-based question answering and text manipulation via a “Prompt Lab.” Performance varies based on device hardware and model size. The app is an experimental Alpha version, open-source under Apache 2.0, and encourages user feedback.