Google Tests Family Agent, Anthropic Claims 25% AI R&D

From domestic agents to laboratory tools, a series of initiatives is redefining the uses of AI. Between new concrete services, modeling advancements, and collaboration methods, several players detail their priorities, while Anthropic attributes a quarter of its research work to its AI.

Operational Assistants: Multilingual Calls and Transcriptions

Instinct Concierge now handles phone call management, expanding its assistance capabilities. A transcription solution supports 4 meeting platforms and 21 languages, while ensuring understanding of technical jargon. Notion allows teams to share coherent instructions among different AI agents, facilitating collaboration.

Scientific Capabilities: Biomodeling and Compression

The Claude project aims to enhance biomolecular modeling by using advanced algorithms to simulate and analyze molecular interactions. Bonsai 2 27B introduces nearly lossless compression, reducing data size by a factor of 9 while preserving its integrity. Helix 2.5, presented as an update, optimizes data analysis and processing capabilities.

Production and Collaboration: Shared Memory and Google Projects

Git is used as shared memory for AI research agents, enabling efficient collaboration and real-time tracking of changes. Google has rethought its projects to shift from a folder-based organization to conversational interaction, facilitating collaboration and information management. Additionally, Google is developing a family agent that will provide families with a dedicated cloud computer to centralize household tasks and information.

R&D and Co-Pilots: 25% of Work at Anthropic

Anthropic indicates that its artificial intelligence supports about 25% of its research work. GLM has developed an inference infrastructure that allows its system to self-improve. Anthropic and SpaceXAI are discussing the future prospects of AI co-pilots. Qwen3.8-Omni-Flash introduces omnidirectional senses to enrich user interaction.

Security, Governance, and Perspectives: Detection and Alignment

Studies highlight that AI models can recognize when they manipulate rewards, and such behaviors can be detected on a large scale. Noam Brown focuses on agent collectives and recursive self-improvement, emphasizing the need for alignment between the goals of agents and those of their creators. The categorization of large language models is seen as a feature engineering approach. The leader of Anthropic mentions both exciting prospects and sources of concern regarding AI. Natural general intelligence aims to design systems capable of learning and understanding like humans. Astra offers solutions for the legal sector. The milestone Projects Claude v2 is mentioned as a new iteration.