AI Daily — 2026-09-23
Realtime-Venus enables full-duplex audio-visual interaction; OmniEdu open-sources K-12 foundation models; GameHorizon adds multi-horizon gameplay eval
Covering 11 AI news items
🔥 Top Stories
1. Google Open-Sources AX Agentic Orchestration Runtime
Google has introduced AX, an open-source agentic orchestration runtime for declaring tasks with workspaces and gateway specifications, built on Agent Substrate for sandboxed execution. The project aims to support billions of autonomous agent workloads per cluster, though it remains under active development with possible breaking changes. This signals growing consolidation around secure, high-scale agent infrastructure for practitioners. Source-github
2. OmniEdu Releases Open Foundation Models for K-12 Learning and Teaching
OmniEdu has launched an open family of educational foundation models for K-12 learning and teaching, trained with a corpus combining over 100 educational resources and general instruction data. The models target curriculum understanding, learner-difficulty diagnosis, problem solving, and instructional support. If validated, they could accelerate domain-specific open models in education while raising new evaluation and data-quality questions. Source-huggingface
3. Realtime-Venus: Full-Duplex Audio-Visual Interaction with Async Delegation
Researchers present Realtime-Venus, a proactive full-duplex interaction system built from two separately trained 9B models: one for audio-visual dialogue and one for spoken dialogue. The design targets continuous perception and timely, visually grounded responses in digital and physical environments. This pushes multimodal assistants closer to real-time, embodied interaction, though latency and deployment cost remain practical constraints. Source-huggingface
📰 Featured
Benchmarks & Evaluation
- GameHorizon Suite: Multi-Horizon Data and Evaluation for AI Gameplay — A unified data and evaluation suite measures AI gameplay across visual understanding, instruction decomposition, goal planning, and action control, addressing narrow or language-poor game benchmarks. Source-huggingface
- Paper Introduces Taste Benchmark for Long-Horizon LLM Agents — A new benchmark proposal measures “taste” in LLM agents by judging intermediate decision quality rather than only end-to-end task success. Source-huggingface
Robotics & Embodied AI
- Grounded Action Models Use 3D Grounding for Robot Foundation Models — Grounded Action Models explicitly build 3D grounding into robot foundation models so manipulation policies know which objects matter and where they are. Source-huggingface
Agent Infrastructure & Safety
- Agent Substrate: Secure AI Agent Runtime with High-Density Sandboxing — An open-source runtime targets millions of sandboxes with microVMs and gVisor, sub-500ms resume, and zero-trust kernel and network isolation. Source-github
Tools & Developer Workflows
- Browser Use Releases video-use: Edit Videos with Claude Code Agents — An open-source tool lets users chat with coding agents to generate a final video, with filler-word removal, color grading, subtitles, animation overlays, self-evaluation, and persistent session memory. Source-github
- Univer Launches Open-Source Office SDK for AI Agents — Univer provides embeddable spreadsheet, document, and presentation building blocks with a plugin architecture, Canvas rendering, formula engine, and Facade API for browser and Node.js. Source-github
- Open-Source CLI Offers Templates for Configuring Anthropic Claude Code — A CLI and web interface offers ready-to-use agents, custom commands, settings, hooks, MCP integrations, and project templates for Claude Code. Source-github
⚡ Quick Bites
- Treg: OpenRouter for AI Agent Tools with 3,000+ Endpoints — Treg is pitched as an OpenRouter-style gateway for AI agent tools, exposing more than 3,000 endpoints for discovery and integration. Source-github
Generated by AI News Agent | 2026-09-23