AGI

Okta’s Model Context Protocol (MCP) addresses the high cost of AI tool access by reducing token consumption. Instead of providing AI agents with a complete list of available tools, Okta filters tools based on user identity and permissions *before* they reach the AI model. This identity-scoped approach significantly minimizes unnecessary token usage, lowering operational expenses. The solution also enhances security by limiting an agent’s visibility to only authorized tools, reducing the potential attack surface.

Google’s AMIE, an AI for healthcare, achieved human-level ratings in synchronous video consultations with patient actors across key metrics. Its multi-agent architecture, separating dialogue, planning, and perception, allows for low latency and improved interaction. While promising, researchers emphasize real-world patient trials are crucial before clinical adoption.

Novo Nordisk is expanding its strategic partnership with AWS, making AWS its preferred cloud and AI provider. This collaboration will integrate AWS’s AI capabilities across Novo Nordisk’s drug discovery pipeline, establishing a co-innovation hub to accelerate research workflows, from target identification to patient treatment. Novo Nordisk aims to leverage AI agents and advanced models to discover novel drug targets, design therapies, and analyze complex biological data, significantly shortening development timelines.

AI is accelerating vulnerability discovery, but identifying deployed instances remains a challenge, especially in complex container environments. A recent report details the first known instance of AI aiding in zero-day exploit development. While AI can also help generate fixes, effective remediation relies on accurate software inventories and streamlined container rebuilding processes. The true competitive advantage now lies in an organization’s ability to rapidly assess exposure and deploy fixes, rather than just discover vulnerabilities.

Meta’s Superintelligence Labs released Muse Glimmer, a 30B parameter LLM under Apache 2.0, for local AI agents. Available on Hugging Face, it runs on consumer GPUs, enhancing privacy and reducing latency by enabling on-device processing of sensitive data. Muse Glimmer excels in agentic tasks, coding, and general reasoning, often outperforming competitors like Gemma4-31B and Qwen3.6-27B in benchmarks. Its multimodal capabilities and memory-efficient design, utilizing 4-bit quantization, make it suitable for real-time interactions on consumer hardware.

Physics AI, like Siemens’ Simcenter PhysicsAI, can accelerate engineering design exploration by up to 1,000 times using surrogate models. While powerful for rapid iteration and hypothesis testing, it is not suitable for final safety-critical component sign-off. Its accuracy is high, but limited by the data it was trained on, requiring human validation for absolute certainty. Transparency about these limitations is key to building engineer trust.

Stanford researchers used generative AI (Evo 2) to design and synthesize novel bacteriophages. The AI generated nearly 300 potential phage genomes, with 16 showing potent activity against E. coli. This approach, utilizing computational screening and a cocktail strategy, enhances therapeutic potential and offers a new weapon against antibiotic resistance. The Evo 2 model is now open-source to accelerate synthetic biology innovation.

Instagram’s AI curates content, but genuinely human content is increasingly prioritized. AI handles intricate engagement mechanics and automation, freeing creators for strategy and storytelling. While AI chatbots streamline communication and production, authentic human interaction and ingenuity remain paramount. The platform aims to balance AI-driven efficiency with the preservation of human creativity and connection.

Alibaba is reportedly moving towards a revenue-sharing model for its Qwen open-weight AI. Commercial users leveraging the model for revenue generation will need commercial agreements. This marks a shift from traditional licensing, influenced by trends seen with models like Moonshot’s Kimi K3, which also employs revenue-sharing for large-scale commercial applications. The strategy aims to monetize advanced AI technologies beyond basic licensing, acknowledging the significant infrastructure costs associated with deploying large models.

Research shows user expertise critically impacts AI explainability in healthcare. Non-experts, particularly with LLM explanations, tend to defer to AI, increasing errors. Primary care physicians perform best with AI predictions alone, being more critical of explanations. The study stresses tailoring AI interfaces to user expertise to avoid automation bias and ensure safe integration.