Zhipu’s GLM-5.3 Agent Optimizes Its Own Inference Stack
Zhipu says a GLM-5.3-powered infrastructure agent helped triple serving performance on a 100,000-accelerator Chinese cluster.
September 16, 2026
All briefings and quick hits published that day.
- ⚡ Quick hits
- Research OpenAI developer tools reportedly helped decode an 83-year-old German military radio message in ten hours. IT之家 ↗
- Industry OpenRouter’s latest token chart shows explosive usage growth, intensifying debate over whether AI demand reflects durable utility or speculative overbuilding. The Decoder ↗
- Research Researchers presented RhymeRL, a reinforcement-learning system that improves performance by learning from historical experience. InfoQ 中文 ↗
- Research OpenAI’s GPT-6 Astra reportedly decrypted a Nazi radio message in ten hours after 83 years of unsuccessful human analysis. The Decoder ↗
- Industry Graphics researcher Xin Tong joined Meshy to lead its push toward playful, consumer-facing AI-generated 3D creation tools. 量子位 ↗
- Agents A Chinese state-owned enterprise’s general-purpose agent reportedly ranked among IDC’s top three in practical evaluations. 量子位 ↗
- ⚡ Quick hits
- AIGC Alibaba’s Wan3.0 demonstrated a single 30-second generated video shot, extending its earlier five- and fifteen-second production milestones. Alibaba Wan ↗
- Funding Icelandic voice-simulation startup Treble raised $18 million to expand its platform for generating realistic synthetic speech data. TechCrunch ↗
- Agents GraphEcho proposes structural redundancy and evidence provenance mechanisms for improving reliability in LLM-based graph agents. arXiv ↗
- AIGC Kling AI previewed Odysseus: The Fall, a feature film directed by Ash Koosha and powered by Kling AI’s video-generation technology. Kling AI ↗
- Agents Hang Ten Systems raised another $53 million to deliver enterprise AI services using agentic code generation and reusable skills libraries. SiliconANGLE ↗
- Funding Open-weight model developer Arcee AI reached a valuation above $1 billion through a new undisclosed Series B financing round. SiliconANGLE ↗
- Open Source Nunchux AI introduced VC-Attention, a training-free low-bit attention kernel designed to accelerate video diffusion transformers. MarkTechPost ↗
- Agents Sakana AI upgraded Sakana Chat with a newer orchestrator model and added memory capabilities. Sakana AI ↗
- Agents Snap introduced Specs Intelligence, an AI assistant for its smart glasses that will also be available on iOS and Mac. The Verge ↗
- Funding Noetive launched with $41 million to build self-improving AI systems for factories and logistics. SiliconANGLE ↗
Cohere and Aleph Alpha Sign Definitive Sovereign AI Merger Agreement
Cohere and Aleph Alpha have signed a definitive merger agreement, creating a Canada-Germany AI group aimed at regulated and sovereign deployments; final financial terms were not disclosed.
- ⚡ Quick hits
- Agents GitHub detailed a new Copilot Inline Suggestions model that unifies completions and next-edit predictions. X / GitHub Copilot ↗
- Funding CADDi raised $114 million at a $1.2 billion valuation to expand its manufacturing-data platform into North America. SiliconANGLE ↗
- Research Stanford researchers released Paper2Agent, which turns research papers into agents that reproduce results and run on new data. MarkTechPost ↗
- Open Source Knowledgator released GLiFormer, a 575-million-parameter encoder reporting 91.10 F1 on nested JSON extraction without token generation. MarkTechPost ↗
- Agents Baseten introduced hosted web search and grounded inference for open models, removing manual orchestration and separate search keys. X / Baseten ↗
- Agents AWS published open-source agent skills for improving AI reasoning in healthcare and life-science workflows. AWS Machine Learning Blog ↗
- Models A new 4-billion-parameter model reportedly produced query plans 81% faster than PostgreSQL in benchmark testing. Rohan Bansal ↗
- AIGC Runway introduced Flux Video Edit for adding, removing, replacing, or restyling objects in existing footage, plus 2K and 4K upscaling. X / Runway ↗
Google Opens Home Devices to Third-Party AI Agents
Google is exposing Google Home controls through an agent interface, allowing outside AI systems to operate compatible smart-home devices.
OpenAI Publishes Framework for Reporting Model Misalignment
OpenAI has released criteria and timelines for disclosing consequential model-misalignment incidents, opening a debate over voluntary safety reporting.
- ⚡ Quick hits
- Research Google DeepMind launched an interdisciplinary institute focused on major scientific and philosophical questions surrounding AGI. The Decoder ↗
- Agents ElevenLabs launched Reception, an AI receptionist platform for small businesses built on its ElevenAgents technology. X / ElevenLabs ↗
Anthropic Merges Claude Chat and Cowork Into One Workspace
Anthropic is combining Claude’s conversational and task-execution products, adding native document, presentation, and design workflows inside one interface.
- ⚡ Quick hits
- Agents Cognition introduced Devin Code Scans for finding slow paths, untested flows, dead code, and accessibility issues across repositories. X / Cognition ↗
- Industry Mozilla’s new Smart Window assistant uses Mistral models to provide AI help inside the browser. The Decoder ↗
- AIGC Tripo launched a ChatGPT plugin that turns prompts into 3D assets for gameplay and animation workflows. X / Tripo ↗
- AIGC OpenAI published a campaign describing how advertisers can use AI to create and personalize marketing materials. OpenAI ↗