How to Preserve Context in Long-Running Codex Projects

AWS published a guide on using Amazon Nova 2 Lite for content moderation via prompting, demonstrating how to apply the MLCommons AILuminate Assessment Standard's 12-category hazard taxonomy without requiring model fine-tuning. The approach allows organizations to update moderation policies by editing prompts rather than retraining models, and includes benchmarks comparing Nova 2 Lite against other foundation models on public datasets. The technique works with both the AILuminate taxonomy and custom moderation policies, making it adaptable to different organizational needs.

Microsoft is expanding Edge's Copilot AI chatbot to access information across all open browser tabs, allowing users to ask questions about tab content, compare products, and summarize articles. The feature includes user controls to opt in or out of specific experiences. Microsoft is simultaneously retiring Copilot Mode, which offered similar tab-based context plus agentic capabilities like booking reservations, consolidating these functions into the new unified Copilot interface.

SpaceX has struck a deal with Cursor, an AI code-writing startup, that gives the rocket and satellite company an option to acquire Cursor for $60 billion later this year, or alternatively pay $10 billion for collaborative work. The deal comes as SpaceX prepares for what could be one of the largest IPOs ever, likely in June. SpaceX framed the potential acquisition as enabling it to build more useful AI models, continuing CEO Elon Musk's push to position the company as an AI and data center operator alongside its core space business. Cursor, founded in 2022 and valued at over $3 billion in prior funding, has faced competitive pressure from code-writing tools launched by OpenAI and Anthropic.

Anthropic released Claude Opus 4.7 on April 16, positioning it as a focused upgrade for teams building agents, long-running coding workflows, and vision-based tools rather than a broad leaderboard sweep. The model handles images up to 2,576 pixels (3.75 megapixels), reads cluttered screens at 98.5% accuracy versus 54.5% for Opus 4.6, and shows measurable gains in tool use, error recovery, and long-context coherence. Key improvements target production failure modes like infinite loops, tool failures, and hallucination on missing data, with notable jumps on coding benchmarks like SWE-bench Pro (53.4% to 64.3%).