AI Daily — 2026-07-09

English 中文

OpenAI Unveils ChatGPT Work: New Agent for Cross-App Tasks · OpenAI Unveils GPT-5.6: Frontier AI ...

Covering 27 AI news items

🔥 Top Stories

1. OpenAI Unveils ChatGPT Work: New Agent for Cross-App Tasks

OpenAI introduces ChatGPT Work, a new agent in ChatGPT powered by Codex and GPT-5.6, capable of acting across apps and files, staying with a project for hours, and turning goals into finished work. This signals a shift toward persistent, goal-driven AI agents that can automate workflows across disparate tools. Source-x

2. OpenAI Unveils GPT-5.6: Frontier AI Scales with Ambition

OpenAI markets GPT-5.6 as “Frontier intelligence that scales with your ambition,” promising more intelligence per token, better performance per dollar, and greater on-demand capability for demanding tasks. The rollout reinforces OpenAI’s strategy to push higher-efficiency, higher-impact models for enterprise workloads. Source-x

3. GPT-Live Rolled Out to ChatGPT Go, Plus, Pro

GPT-Live, OpenAI’s new voice-capable generation, is now fully rolled out to ChatGPT Go, Plus, and Pro, with a free rollout underway; users should update the app and enable sound for voice interactions. This expands natural, hands-free interaction avenues across the product tiers. Source-x

📰 Featured

Open Source & Local AI

TencentDB-Agent-Memory Delivers Local Long-Term AI Memory — Fully local memory for AI agents via a four-tier pipeline reduces token usage (up to 61.38%), increases pass rate (51.52% relative), and raises PersonaMem accuracy from 48% to 76% when integrated with OpenClaw. Source-github
MiMo v2.5 Emerges as Fast, Underrated Local LLaMA — On a 192GB/RTX 4090 server, MiMo v2.5 reportedly tops local 30B–400B models for inference, suggesting a dedicated server investment for best local performance. Source-reddit
OpenMed 1.8 Goes Local: On-Device De-Identification on Android, iOS, Browser — Delivers fully on-device de-identification for clinical text with OpenMedKit, bridges for Android/iOS/React Native, and browser runtime, with zero server calls and broad model support. Source-reddit

LLMs & Agentic AI

Muse Spark 1.1 launches agentic coding model via Meta API — Muse Spark 1.1 provides agentic and coding capabilities through the Meta Model API and Meta AI at a very low price, emphasizing developer-friendly access and tool-building. Source-x
Elon Musk admits Anthropic leads AI, praises Mythos/Fable — Musk shifts to acknowledge Anthropic’s leadership, praising Mythos and Fable while hinting at upcoming Mythos 2; frames a stance on open, competitive access. Source-x

AI for Science & Industry

OpenAI Deep Native Structural Reasoning? Not listed in top group; Included here for context. — (Deep Native Structural Reasoning for Accurate Structure–Property Understanding) argues for integrating stereochemistry, bonding, symmetry and energetics across biology, chemistry, and materials science to improve structure-property understanding, while noting integration challenges. Source-huggingface

AI Hardware & Inference

Puzzle-75B-A9B NVFP4 Hits 132 t/s on 3×3090 — A highly cacheing setup achieving about 132 t/s across three RTX 3090 GPUs (~65 t/s per stream) with ~500W, suggesting a cost-efficient path to high-throughput 75B MoE models. Source-reddit

⚡ Quick Bites

Scaling MoE Video Pretraining for Embodied Intelligence — Exploring scaling Mixture-of-Experts for embodied AI video pretraining. Source-huggingface
Gemma 4 Launches Open-Weight Multimodal LLM Suite — Introduces open-weight multimodal LLM capabilities for experimentation. Source-huggingface
Obra’s Superpowers: Agentic AI Coding Framework — An agentic AI coding framework to accelerate AI-powered development. Source-github
Desktop Commander MCP Enables AI-Controlled Terminal and Files — AI-driven terminal and file-control tooling. Source-github
GLM-5.2 Open-Source AI Spurs Security Fears — Open-source GLM-5.2 raises security concerns. Source-reddit
6x MI50s vs 6 P40s: MiniMax M2.7 REAP 139B — Hardware comparison for a 139B model in REAP; performance depends on configuration. Source-reddit
OpenMOSS MOSS-Transcribe-Diarize 0.9B: End-to-end multi-speaker transcription — End-to-end transcription across multiple speakers with a 0.9B model. Source-reddit
Reasoning-Medical0.1-27B Finetunes Qwen3.5-27B, Claims Surpass MedGemma — Medical-domain 27B finetune claims to surpass MedGemma. Source-reddit
KoboldCpp v1.117 Released — Latest release of KoboldCpp. Source-reddit
Grok 4.5 Receives Positive Review — Early impressions of Grok 4.5. Source-x
AI costs addressed: 5.6 SOL per task signals major efficiency gains — Cost-per-task improvements highlighted for AI workflows. Source-x
Dual Latent Memory in Vision-Language-Action Models for Robotic Manipulation — Dual latent memory approach in VLA models for robotics. Source-huggingface
Paying for LLMs? Local Embeddings and Rerankers Win — Local embeddings and rerankers as a cost-saving route. Source-reddit
Devs: Thoughts on Mistral Medium 3.5 128B Dense — Community thoughts on Mistral Medium 3.5 128B dense. Source-reddit
RTX GPUs Do Not Benefit from FlashAttention-3/4 Optimizations — RTX hardware lacks gains from certain FlashAttention optimizations. Source-reddit
Turning Codex into ChatGPT Desktop Is a Fumble — Critics on turning Codex into a desktop ChatGPT. Source-x
Flash IQ4 XS GGUF with preserve_thinking — Discussion of IQ4 XS GGUF with preserve_thinking. Source-reddit

Generated by AI News Agent | 2026-07-09