Install.md LLM standard, agents supervise AIs, and bypassing model safety. Plus: McKinsey's AI agents.
Over the weekend, Apple and Google announced a partnership that will see Gemini power Apple's foundation models and likely Siri. Wikipedia also revealed multiple deals with AI giants to license its content. Elsewhere, a new standard for LLM-executable installations, Install.md, has emerged, and researchers are using AI agents to supervise other AIs.
Moxie Marlinspike (Signal creator) is launching Confer, an open-source AI assistant built for privacy. It uses strong encryption and trusted execution environments (TEEs) to keep conversations unreadable by anyone but the user, with encryption keys stored on devices.
Read full story→Wikimedia Foundation signed deals allowing Microsoft, Google, Amazon, and Meta to use Wikipedia content for LLM training. This signals a shift where high-quality, verifiable data is a paid asset, not just a scraped resource, impacting data acquisition strategies for builders.
Anthropic's "Cowork" research preview for Claude Max subscribers on macOS lets Claude access and edit files in a user-designated local folder. Claude can autonomously complete tasks like organizing files, generating spreadsheets, or drafting reports, with users retaining full control over its actions.
McKinsey CEO Bob Sternfels announced that 25,000 of the firm's 60,000 employees are now AI agents.
Apple announced a multi-year partnership with Google, integrating Google's Gemini and cloud technology into future Apple Foundation Models. This collaboration will power upcoming Apple Intelligence features, including a more personalized Siri.
OpenAI committed over $10 billion to deploy 750 megawatts of Cerebras wafer-scale accelerators through 2028. The deal will speed up ChatGPT's inference services, leveraging Cerebras' SRAM-heavy architecture for faster real-time AI responses. Cerebras will also build and lease the necessary data centers for OpenAI.
Constitutional AI (CAI) trains models to be harmless without human labeling. It uses an AI to self-critique responses against principles, then applies 'AI feedback' (RLAIF) for harmlessness evaluations. This cuts human supervision for alignment and trains models to engage with sensitive topics non-evasively.
Omitting `apply_chat_template()` bypasses safety guardrails in models like Gemma and Qwen, letting them generate harmful content. Without the expected chat template tokens, instruction-tuned models revert to unaligned next-token prediction. Safety guarantees for instruction-tuned models are non-existent without strictly sanitized inputs and enforced templates.
GitHub Copilot is implementing an agentic memory system that uses 'just-in-time verification' to store and validate memories with code citations in real-time. Agents can now learn and share knowledge across tasks, leading to a 7% increase in pull request merge rates and a 2% increase in positive feedback for code review agents.
Amazon's AMET Payments team developed SAARAM, a multi-agent AI system, cutting test case generation time from one week to hours. Using Amazon Bedrock with Claude Sonnet and Strands Agents SDK, the system mirrored human QA patterns and delivered a 40% improvement in test coverage and reduced hallucinations.
`install.md` defines a structured Markdown format for software installation instructions, letting AI agents autonomously set up software. This standard makes agent-driven setup more reliable and secure than traditional scripts, with Mintlify already integrating it for users.