Frontier Notes / Daily Signal Report
Today
Hermes agent's 'Ministry of Experts' architecture uses a panel of models orchestrated by a lead model with caching to reduce costs while improving answer quality, enabling computer control and skill learning.
Editor's Notes
This week's releases span agentic orchestration (Hermes' Ministry of Experts), specialized diffusion models for drug discovery (Pearl from Genesis Molecular AI), and API expansions (Google's Gemini Omni Flash for video editing). Together, they signal a shift toward modular, task-specific AI systems that prioritize cost efficiency, domain expertise, and new interaction modalities like video and GUI automation.
Key Takeaways
Anthropic released Claude Sonnet 5 alongside a Linux desktop app beta. The model offers improved agentic coding and a new 'ultra code' effort level, but testing showed it is slow and produced mixed results on complex tasks like 3D game generation. The reviewer expressed frustration, finding the model competent but not impressive given its expected pricing post-August 31st.
The most innovative diffusion research is now happening in drug discovery, specifically for 3D protein-small molecule structure prediction, not image generation. Genesis Molecular AI's Pearl model achieves sub-angstrom accuracy for protein-ligand poses, enabling more reliable downstream predictions and drug design. The company combines physics-based synthetic data, diffusion models, and inference-time scaling to create foundation models that generalize to novel, challenging drug targets.
Hermes agent's latest update introduces a Ministry of Experts feature that uses a panel of models orchestrated by a lead model, leveraging caching to reduce costs while improving answer quality. Other major updates include computer control for GUI automation, a /learn skill that turns interactions into reusable skills, voice support via whisper and 11 Labs, and new integration options like iMessage and WhatsApp. The update also brings memory reliability improvements, background agents, native desktop app, and automation blueprints.
Anthropic's Claude Design 2.0 introduces major upgrades including unified token usage with the Claude subscription, MCP connectors for third-party tools like Hickfield, canvas editing with a Figma-like sidebar, improved design system creation and porting from Claude Code, and the ability to export projects to other platforms. These changes make Claude Design more efficient, flexible, and integrated into the development workflow.
Google has released the Gemini Omni Flash API with capabilities beyond VEO, focusing on four key areas: conversational video editing, multimodal input references, world model simulations, and in-video text/logos. The API uses a new Interactions API for multi-turn video generation and editing, supporting text-to-video, image-to-video, and video-to-video edits with audio, though currently limited to 10-second clips and no deepfake-style features.