Anthropic introduced Claude Design, a research-preview tool powered by Claude Opus 4.7, enabling users to generate prototypes, slides, and one-pagers from natural language. This move positions Anthropic directly into the design tooling market, challenging existing platforms like Figma.
Sources: Latent Space, Simon Willison
Meta launched Muse Spark, its first AI model in a year, featuring multimodal reasoning, tool use, and multi-agent orchestration. It shows strong performance in health and multimodal benchmarks, positioning Meta to compete for business customers, but its proprietary nature marks a shift from Meta's previous open-weights strategy.
Sources: The Batch
Pharmaceutical giant Eli Lilly has agreed to pay Insilico Medicine up to $2.75 billion for exclusive rights to develop and sell AI-discovered drugs, building on previous collaborations. This significant investment highlights the growing confidence in generative AI's potential to accelerate drug discovery, with Insilico already having 28 candidate drugs, some in clinical trials.
Sources: The Batch
Despite efforts from President Trump to discourage state-level AI regulation, U.S. states are proceeding with enacting their own laws. This indicates a fragmented and evolving regulatory landscape for AI within the United States.
Sources: The Batch
Dubai police arrested an airline worker after accessing a private WhatsApp group, explicitly stating they conduct electronic monitoring operations capable of detecting private messages. This incident highlights the use of sophisticated surveillance tools like the Israeli-developed Pegasus spyware, which can infect devices and access encrypted communications, raising global concerns about user privacy and state access to private data.
Sources: Hacker News
NVIDIA has released Nemotron OCR v2, a new production-ready, multilingual OCR model capable of processing 34.7 pages/second on a single A100 GPU. Trained on 12 million synthetic images generated by a novel pipeline, this unified model simultaneously supports six languages, simplifying deployment. NVIDIA also made the 12.2 million sample synthetic dataset publicly available to address data scarcity in multilingual OCR research.
Sources: Hugging Face Blog
Amazon Bedrock has introduced a new Model Distillation feature, allowing users to train smaller, more cost-effective 'student' models by transferring intelligence from larger 'teacher' models like Amazon Nova Premier. This innovation significantly reduces LLM inference costs by over 95% and lowers latency by 50%, as demonstrated in applications like video semantic search. AWS provides comprehensive GitHub resources, including a Jupyter notebook, to help developers implement and experiment with this new capability.
Sources: AWS ML Blog
Amazon Web Services (AWS) has introduced a new, comprehensive video semantic search solution, centered around its newly launched Nova Multimodal Embeddings. This architecture unifies text, images, video, and audio into a shared semantic space, leveraging Anthropic Claude Haiku for intent analysis and dynamic FFmpeg scene detection for optimal video segmentation. The solution aims to significantly enhance retrieval accuracy and cost efficiency for video content by providing finer control over multimodal signal representation.
Sources: AWS ML Blog
Simon Willison successfully applied "agentic engineering" using Claude Code to automate a complex update for his blog-to-newsletter tool. This innovative approach leverages concise prompts, GitHub repositories for contextual understanding, and self-validation mechanisms, enabling AI agents to infer complex logic and significantly streamline software development tasks.
Sources: Simon Willison
Amazon Bedrock now provides automatic, granular cost attribution for AI inference calls to specific IAM principals, enabling organizations to accurately track, optimize, and charge back AI cloud spend. This feature integrates with AWS Billing and supports custom cost allocation tags for detailed financial analysis.
Sources: AWS ML Blog
A key feature of the Amazon Nova Forge SDK is its data mixing technique, which blends domain-specific training data with Amazon-curated datasets. This method is critical for preserving a model's general capabilities and MMLU scores while achieving significant performance improvements on specialized tasks, directly addressing the challenge of catastrophic forgetting.
Sources: AWS ML Blog
AWS Marketing's Technology, AI, and Analytics (TAA) team collaborated with Gradial to launch an agentic AI solution on Amazon Bedrock. This innovative system, utilizing models like Anthropic Claude and Amazon Nova, automates enterprise content publishing and validation, cutting webpage assembly time by over 95%—from four hours to just ten minutes. The solution also integrates a Model Context Protocol (MCP) server for real-time content quality checks, significantly enhancing efficiency for marketing teams.
Sources: AWS ML Blog
A developer has successfully enabled Anthropic's 122B parameter Claude Code model to run entirely locally on Apple Silicon Macs, eliminating API fees and cloud dependencies. This technical achievement allows for faster, private, and free use of the model, potentially impacting the reliance on cloud-based AI inference.
Sources: Keep Bookmarks
Anthropic has released the system prompts for its Claude chat models in Markdown format, allowing developers and researchers to examine the underlying instructions. This transparency facilitates a deeper understanding of how Claude models, including Opus 4.6 and 4.7, have evolved.
Sources: Simon Willison
PyCon US 2024, scheduled for May 13-19 in Long Beach, California, will introduce a new dedicated AI track. This track will delve into practical applications and research, covering topics such as running large language models on consumer hardware, edge inference, building AI agents, and the broader impact of AI on open-source contributions, education, and language accessibility.
Sources: Simon Willison