This edition’s highlights Tools Uncanny and unappetizing: appetites spoil as AI images take over food menus 2026-09-06 13:00 UTC AI-generated menu images are becoming widespread but often look unappetizing. A Denver nurse's post about AI images of meat resembling beetle-covered leather belts drew hundreds of disgusted reactions. The Guardian AI In-site article Uncanny and unappetizing: appetites spoil as AI images take over food menus Read later Policy An Alien Mind 2026-09-06 09:00 UTC Increasingly capable AI makes alignment with human intent harder. Pachocki calls for stronger safety safeguards. OpenAI News In-site article An Alien Mind Read later Agents Research acceleration: The view inside OpenAI 2026-09-06 08:00 UTC Coding agents are embedded in OpenAI research workflows Early data shows faster experiment velocity OpenAI News In-site article Research acceleration: The view inside OpenAI Read later UC Berkeley Researchers Release CUA-Lite, an Open Platform Unifying Sandboxes, Data, Evaluation and RL for Computer-Use Agents 2026-09-06 06:11 UTC CUA-Lite packages agents, environments, traces and training/eval into one action space and LiteSample data schema. Lite.OSWorld replaces per-task QEMU/KVM VMs with a Docker container: 0.9GB memory vs 4.1GB, roughly 4.6x more parallel instances, and identical scores to OSWorld across 13 models. MarkTechPost In-site article UC Berkeley Researchers Release CUA-Lite, an Open Platform Unifying Sandboxes, Data, Evaluation and RL for Computer-Use Agents Read later GitHub Introduces Project HydraFusion: Runtime Multi-Model Orchestration That Builds a Workflow Per Coding Task in Copilot CLI 2026-09-05 19:40 UTC HydraFusion selects a workflow, not just a model, for each request, potentially mixing models from multiple providers. Three shipping patterns: Single, Cascade with a quality gate, and Critique with a read-only cross-family reviewer. MarkTechPost In-site article GitHub Introduces Project HydraFusion: Runtime Multi-Model Orchestration That Builds a Workflow Per Coding Task in Copilot CLI Read later Models Perplexity Details Its GPU Embedding Stack: How Ivy, Tulip and ROSE Serve pplx-embed 2026-09-06 03:20 UTC Perplexity reuses its LLM prefill/decode kernels instead of building a separate embedding engine. Three services handle requests: Ivy (Rust HTTP gateway), Tulip (Rust gRPC server), and ROSE (Python inference engine). MarkTechPost In-site article Perplexity Details Its GPU Embedding Stack: How Ivy, Tulip and ROSE Serve pplx-embed Read later Nous Research Adds One-Click Local Model Setup to Hermes Desktop 2026-09-05 19:12 UTC Hermes Desktop now performs hardware detection, model selection, weight download, and runtime configuration in one click. The app manages a hardware-matched llama.cpp build automatically and does not expose context length, GPU layers, or quantization knobs. MarkTechPost In-site article Nous Research Adds One-Click Local Model Setup to Hermes Desktop Read later