August 23, 2026
Quick hits from across the AI world in the last 24 hours.
- ⚡ Quick hits
- Industry Nio autonomous-driving chief Ren Shaoqing reportedly founded an embodied-intelligence startup that will receive strategic investment from Nio. IT之家 ↗
- Open Source Netflix open-sourced an agent workflow designed to conduct causal reasoning rather than merely identify correlations in observational data. InfoQ 中文 ↗
- Industry Cerebras unveiled its CS-4 system, claiming twice the performance of its predecessor while retaining the same wafer-scale processor. The Decoder ↗
- Agents Robert Martin said his experiment trusting AI-generated code revealed unresolved reliability problems despite substantial productivity gains. InfoQ 中文 ↗
- Safety An investigation found AI chatbots frequently directed pregnant users to anti-abortion websites without disclosing those sites’ ideological affiliations. The Decoder ↗
- AIGC Alibaba’s Wan 3.0 became available on Leonardo AI, extending the video model’s distribution to another mainstream creative platform. X / Alibaba Wan ↗
- AIGC Dreamina showcased Seedance 2.5 generating a surreal video whose character becomes trapped inside a CAPTCHA interface. X / Dreamina ↗
- Research MIT Technology Review examined evidence that children infer language rules more effectively than current AI systems, with the underlying reason still unresolved. MIT Technology Review ↗
- Research Alibaba DAMO Academy introduced a liver-cancer AI model reportedly capable of accurately identifying tumors as small as one centimeter. 量子位 ↗
- Industry Zhongcheng Hualong released its HL200 inference chip alongside a supernode architecture designed to scale into clusters containing thousands of accelerators. 量子位 ↗
- AIGC Wan 3.0 launched with native 30-second video generation, synchronized audio, multimodal references and improved temporal consistency. 量子位 ↗
- Research A WRC presentation outlined native omni-modal world models that progress from passively simulating environments toward supporting interactive worlds. 量子位 ↗
- Models Together AI said GLM-5.3 completed over five times more DeepSWE work than Fable 5 under an identical $100 budget. X / Together AI ↗
- Policy Sam Altman warned that concentrating control of advanced AI among too few organizations could create serious societal risks. SiliconANGLE ↗
- Industry Usage data indicates Anthropic’s flagship model trails cheaper alternatives among developers despite its stronger capabilities. Simon Willison ↗
- Models Harvey introduced Tenet, a Kimi K3-based model post-trained with Fireworks for long-horizon legal-agent tasks. MarkTechPost ↗
- Agents Google Cloud published guidance for giving LLM agents session state so they can preserve context and preferences across interactions. Google Cloud Tech on X ↗
- Open Source AWS open-sourced Dogwood to define and enforce safer, more predictable tool-calling rules for AI agents. InfoQ 中文 ↗
- Agents A new organizational-design proposal argues deployed AI agents need explicit roles, reporting lines and escalation paths. SiliconANGLE ↗
- Safety Flock Safety’s chief executive called for compromise as opposition grows toward the company’s AI-assisted camera surveillance network. TechCrunch ↗
- Agents Databricks demonstrated Unity AI Gateway for centrally tracking model, coding-agent and MCP usage, costs and access policies. Databricks on X ↗
- Policy TechCrunch reviewed the unsettled legal standards governing whether copyrighted books may be copied to train AI models. TechCrunch ↗
- Industry Amazon DynamoDB added native capabilities for AI-oriented search, reducing the need to maintain a separate vector database. InfoQ 中文 ↗
- Industry A World Robot Conference exhibit staged an immersive café where robots and people collaboratively handled service tasks. 量子位 ↗
- Agents AI21 Labs trained a lower-cost verifier that it says matches frontier-verifier quality across agentic systems. AI21 Labs on X ↗
- Agents An experimental AI company boss dismissed its first employee only after human supervisors reminded the agent of its own rules. The Decoder ↗
- AIGC Lovart added a branding workflow that turns an uploaded visual identity and prompt into complete commercial campaign materials. Lovart on X ↗
- Industry A survey profiled Hong Kong professors commercializing embodied-AI research through robotics startups and university-linked ventures. 量子位 ↗