October 5, 2026
Quick hits from across the AI world in the last 24 hours.
- ⚡ Quick hits
- Safety Google Research outlined open problems in agentic privacy and security, emphasizing context-sensitive risks rather than isolated model behavior. Google Research ↗
- Open Source Reflection AI introduced Beam, a 501-billion-parameter open-weight MoE model with 23 billion active parameters for coding and agentic workloads. MarkTechPost ↗
- Open Source NVIDIA released Nemotron-Labs-3-Competitive-Coding on Hugging Face, a competitive-programming specialist based on Nemotron-3-Ultra. X @NVIDIAAI ↗
- AIGC Figma announced that text animation and audio are now available in Figma Motion, expanding the product’s timeline-based creative toolkit. X @figma ↗
- Agents LlamaIndex argued that agentic OCR can preserve tables, charts, and multi-column structure that conventional one-pass OCR often loses. X @llama_index ↗
- Agents Cognition introduced Dreaming, which lets Devin build a cross-session memory graph and improve it overnight by removing stale records. X @cognition ↗
- AIGC HeyGen launched HyperFrames as an app where an agent creates a first video cut from a description and accepts visual, frame-level direction. X @HeyGen ↗
- Agents Cursor added steerable SDK agents, returning background-subagent results to parent runs and exposing MCP annotations for safer tool selection. X @cursor_ai ↗
- Agents Together launched Link, which routes open models into coding harnesses while tracking spending and automatically selecting lower-cost models when appropriate. X @togethercompute ↗
- AIGC Krea added an After Effects integration allowing its agent to create motion graphics, resize videos, and handle sound and images inside the editor. X @krea_ai ↗
- Policy OpenAI will add invisible text watermarks to eligible ChatGPT and Codex outputs in the European Union while keeping API watermarking opt-in globally. OpenAI ↗
- Models Reka introduced Rho-1, a single omni-model that handles text, images, video, and robot control, trained from scratch on 320 H100 GPUs. X @RekaAILabs ↗
- AIGC Manus Video can review nearly two hours of travel footage, identify highlights, and assemble a finished edit from one prompt. X @ManusAI ↗
- Agents Cohere unveiled North 2 with reusable agents, multi-agent orchestration, enterprise connectors, security certifications, and token-spending caps. SiliconANGLE ↗