September 8, 2026
Quick hits from across the AI world in the last 24 hours.
- ⚡ Quick hits
- Industry Alibaba denied reports that DingTalk founder Chen Hang had returned to lead Qwen’s office-productivity business. IT之家 ↗
- Industry OpenAI, Amazon, and Google were reportedly examining Patagonia in Argentina as a potential site for large AI data centers. IT之家 ↗
- Agents Kimi Work now supports remote control from a phone while the desktop session continues running. X / Kimi Moonshot ↗
- Agents Hugging Face introduced ML Intern, a chat-based tool that lets users run machine-learning experiments through natural-language instructions. The Decoder ↗
- Policy Microsoft updated the Edge Add-ons store to reduce review workload as AI-assisted coding increases plugin submissions. IT之家 ↗
- Research A Qbit report said domestic AI-for-science computing platforms debuted at the 2026 Bund Summit, linking compute capacity with talent planning. 量子位 ↗
- Industry Sakana AI highlighted Sakana Chat as a free conversational tool for student learning, everyday research, and idea generation. X / Sakana AI ↗
- AIGC Gradium launched Voice Design, which generates a synthetic voice from a written prompt within seconds. MarkTechPost ↗
- Industry Ant Group launched its first finance-enhanced model, positioning AI inside real investment-research workflows. 量子位 ↗
- Industry A Qbit report described embodied robots being tested for supermarket inventory across a network covering 70,000 stores worldwide. 量子位 ↗
- Benchmark Together AI said GLM-5.3 Flash beat Claude Fable 5.1 on agentic-automation tasks at roughly 99% lower cost per task. X / Together AI ↗
- Agents OpenAI promoted 16 plugins designed to help small-business owners handle routine operational work. X / OpenAI ↗
- Agents Replit said users were strongly responding to Free Mode and scheduled a community discussion about what they are building with it. X / Replit ↗
- Models Baseten announced Mercury 2.5 support with sub-200-millisecond median latency and more than 1,100 tokens per second on NVIDIA GPUs. X / Baseten ↗
- Models DeepSeek said its Flash models will cut output pricing to 4 yuan per million tokens starting tomorrow. IT之家 ↗
- Benchmark Baseten highlighted Coval benchmarks that measure the quality-latency tradeoff of speech-to-text systems. X / Baseten ↗
- Research Epoch AI reported that long-context costs diverge sharply across frontier models, especially beyond hundreds of thousands of tokens. X / Epoch AI ↗
- AIGC Runway said its generative video models are now available directly inside Adobe Premiere Pro and After Effects. X / Runway ↗
- Open Source NVIDIA announced CUDA Rust projects for compile-time-safe SIMT and tiled GPU kernel development. MarkTechPost ↗
- Open Source Cohere introduced a decode-megakernel serving system that reported up to 1.58x faster performance than vLLM in one test. X / Cohere ↗
- Benchmark Cohere said North Mini Code achieved up to 1.58x serving speed versus vLLM at batch size one on a single H100. X / Cohere ↗
- Research Cohere introduced a megakernel architecture that fuses the LLM decoding step into one GPU kernel launch. X / Cohere ↗
- Benchmark Cohere reported that its decode-megakernel system delivered 1.25x to 1.41x end-to-end serving gains at batch size eight. X / Cohere ↗
- Industry Cohere’s megakernel system was presented as a serving design that maximizes GPU utilization through extensive kernel fusion. X / Cohere ↗
- Research AWS described using a brain-inspired Pathway architecture for model development on SageMaker HyperPod. AWS Machine Learning ↗
- AIGC World Labs described Atlas, a real-time autoregressive model that preserves 3D consistency while generating interactive frames. X / World Labs ↗
- Open Source Karmada graduated from the CNCF and is being used for multi-cluster AI training and GPU scheduling. InfoQ 中文 ↗
- Industry ElevenLabs appointed Ethan Tandowsky as chief financial officer to build financial infrastructure as the company scales. X / ElevenLabs ↗
- Funding Lightsage raised $4 million to help software companies make their products discoverable and purchasable by autonomous AI agents. SiliconANGLE ↗