Samuel Thompson

Paige and Microsoft’s PRISM2 is a revolutionary AI for digital pathology. It interprets whole-slide images and generates natural language answers to diagnostic questions, moving beyond simple pixel identification. Trained on 2.3 million whole-slide images and clinical dialogue from 685,507 pathology reports, PRISM2 consolidates tile embeddings into a single representation. GPT-4o was used to create question-and-answer pairs from this data, ensuring authentic clinical reasoning for intuitive diagnostic insights.

Alibaba launches Qwen3.8-Max, a 2.4 trillion parameter AI model, employing a mixture-of-experts design for efficiency. This move intensifies competition with rivals like DeepSeek, which challenges with aggressive, low-cost pricing for its V4-Flash model. While Qwen3.8-Max boasts multimodal capabilities and a large context window, DeepSeek prioritizes affordability, making cost-per-task a key battleground in the AI market. Both companies also offer open-weight versions, increasing accessibility for developers.

Red Hat launches “asago,” an open-source initiative to automate AI governance from policy to production code. Asago maps policies to risks, assesses them with tailored scenarios, recommends mitigations, and orchestrates deployment, significantly reducing manual effort. It ensures continuous audit trails and infrastructure-agnostic configurations. The project, a collaborative effort with major tech and academic institutions, aims to address the growing need for AI safety and compliance in a scalable manner.

Article 50 of the EU AI Act mandates transparency for generative AI. Providers and deployers must disclose when users interact with AI or when content is AI-generated, particularly for deepfakes and public interest texts. This aims to prevent manipulation and deception. Specific requirements include machine-readable markings for synthetic media and clear notifications for emotion recognition/biometric categorization tools. Enforcement involves national authorities, the AI Office, and the EDPS. A voluntary Code of Practice aids compliance with marking obligations.

GSK is expanding its AI-driven drug discovery efforts through a significant collaboration with Relation Therapeutics. This partnership leverages Relation’s expertise in generating high-volume, sophisticated human cell datasets. These datasets will train advanced AI models, including Relation’s MORGAN platform, to rapidly pinpoint novel drug targets. Building on prior successful collaborations, this agreement integrates cutting-edge data generation with AI model development for more precise therapeutic discovery.

OpenAI is proactively aligning its safety, security, and transparency initiatives with the EU AI Act’s GPAI Code of Practice. The company highlights its rigorous testing, system cards, red-teaming, and internal governance frameworks as evidence of its commitment. OpenAI also addresses AI-generated content provenance through Content Credentials and SynthID watermarking, and manages dual-use risks in cybersecurity via its EU Cyber Action Plan. The company emphasizes the need for flexible regulations to keep pace with evolving AI technology.

Mark Zuckerberg’s op-ed advocates for a democratized future of AI, where “personal superintelligence” empowers individuals, not just institutions. Meta’s philosophy prioritizes individual empowerment, invention, and a balance of power. Zuckerberg criticizes the prevalent “doom” narrative in AI development, arguing that concentrating power is dangerous. He believes open access to AI for cybersecurity is key, while acknowledging the need for institutional coordination on biological risks. The vision emphasizes individual ingenuity driving progress and fostering entrepreneurship.

OpenAI’s report showcases AI coding agents accelerating scientific computing. Eight projects demonstrated reduced runtimes by streamlining packaging, optimizing code, and migrating backends. While AI significantly speeds up development and tackles technical debt, human expertise remains crucial for scientific rigor, validation, and guiding nuanced tasks. These advancements promise to make ambitious projects more feasible.

Google’s AI integration into search is dramatically increasing with AI Overviews appearing in 43% of U.S. searches. This shift towards conversational AI is evident in the substantial growth of AI Mode visits. While AI aggregates information, publishers worry about reduced referral traffic. Google is improving source visibility, and platforms like Cloudflare are exploring ways for publishers to monetize AI access. Citation rates vary, and AI often cites deep-linked content but drives traffic to homepages.

Guardoc Health leverages Amazon Nova models via Bedrock to process over one million clinical documents daily, significantly reducing documentation errors by 46% and audit fines by 70%. This AI integration addresses complex document formats, including handwritten annotations and varied EHR structures. The system employs a cost-tiered retrieval pipeline, utilizing Amazon Textract and Titan Text Embeddings before advanced reasoning from Amazon Nova Pro, to ensure accuracy and efficiency in clinical documentation, ultimately improving patient safety and reducing compliance costs.