August 22, 2026
Quick hits from across the AI world in the last 24 hours.
- ⚡ Quick hits
- Industry An Atelier Ryza spinoff AI chat game delayed its launch after advance reservations substantially exceeded expectations. IT之家 ↗
- Open Source FreeToken demonstrated edge-native mixture-of-experts serving that runs the 753-billion-parameter GLM-5.2 model on one workstation GPU. MarkTechPost ↗
- Agents OpenRouter reported that token consumption by agentic applications has increased fourteenfold, making AI systems major customers of other models. The Decoder ↗
- Research Researchers unveiled a capacitive full-body electronic skin designed to give biomimetic robots human-like tactile perception and interactive capabilities. IT之家 ↗
- Models Tests linked the anonymous Ox Alpha model to Zhipu’s GLM family, while prompting speculation about its relationship with Cursor. 量子位 ↗
- Industry Memory shortages are reportedly pushing Nvidia AI server prices approximately 15% higher, adding about $5 billion to a one-gigawatt data center. 量子位 ↗
- Research A robot played competitive tennis against a human, executing difficult returns and rapidly recovering after falling. 量子位 ↗
- Industry A startup founded by doctoral researchers is developing an integrated intelligence system intended to serve as the brain of bipedal humanoid robots. 量子位 ↗
- Research A study argues AI may increase scientists’ output while lowering average research quality instead of helping them produce better work with less effort. The Decoder ↗
- Open Source DeepDoctection was used to demonstrate an end-to-end document-intelligence pipeline covering layout analysis, text extraction and structured processing. MarkTechPost ↗
- Industry A report detailed how Chinese gray-market vendors resell access to Claude tokens at prices substantially below Anthropic’s official rates. The Decoder ↗
- Agents Vercel introduced Is Agentic, a free tool that scores public websites’ readiness for AI agents using more than 100 checks. MarkTechPost ↗
- Industry Shangwei New Materials opened reservations for its first personal home robot, positioning it as an early consumer-oriented household robotics product. 量子位 ↗
- Models Together AI reported GLM-5.3 reached 87.6% on DeepSWE for about $16, exceeding Fable 5’s 69.7% at $21.63. Together AI on X ↗
- AIGC Harvard’s $699 startup bootcamp is offering AI avatars of its instructors to answer participants’ questions. TechCrunch ↗
- Agents DeepMind alumni-founded Inherent said its AI teammate outperformed systems from Anthropic and OpenAI at replicating published research. TechCrunch ↗
- Open Source Simon Willison released version 0.33 of LLM, his command-line utility and Python library for interacting with language models. Simon Willison’s Weblog ↗
- Policy OpenAI urged California lawmakers to strengthen the state’s proposed AI-safety legislation. TechCrunch ↗
- Safety Leading frontier-AI labs still have not publicly explained how they would contain a model that evaded their control. TechCrunch ↗
- Agents Simon Willison highlighted an argument that coding agents can provide value extending beyond conventional code review. Simon Willison’s Weblog ↗
- Agents Cloudflare introduced Agent Tracing controls for truncation limits and documented differing default payload-recording behavior across supported frameworks. InfoQ 中文 ↗
- Industry UBTech recreated a customer production line at WRC 2026 to demonstrate its embodied-AI deployment workflow in a realistic industrial setting. 量子位 ↗
- Open Source Databricks highlighted Omnigent, an open-source meta-harness giving multiple AI agents shared infrastructure instead of requiring repeated setup. Databricks on X ↗
- Industry Galileo Robot debuted a terrestrial embodied system designed to bridge underlying technologies for wheeled, vehicle-like and legged robots. 量子位 ↗
- Industry MagicLab demonstrated three physical-AI solutions operating in realistic scenarios at WRC 2026. 量子位 ↗
- Research A startup founded by a Tsinghua doctoral student is tokenizing human muscle responses for a neural-interface system. 量子位 ↗
- Agents An open-source course mapped three approaches to running agent loops and compared the provider economics behind each approach. MarkTechPost ↗
- Research A study examined why reusable skills improve AI-agent performance and identified conditions under which those skills fail. The Decoder ↗