November 20, 2026
Quick hits from across the AI world in the last 24 hours.
- ⚡ Quick hits
- Industry Former Intel CEO Pat Gelsinger said AI needs a ten-thousand-fold efficiency leap that incremental chip gains cannot provide. IT之家 ↗
- Agents An InfoQ overview argues that reliable agent systems still depend on planning, memory, tool use, evaluation, and governance working together. InfoQ中文 ↗
- AIGC Meshy demonstrated its N1 Texture workflow, which cleans projected-texture seams around a 3D garment with one press. X @MeshyAI ↗
- Agents TaoDesign described a knowledge-engineering workflow that turns design assets into continuously improving agent capabilities. InfoQ中文 ↗
- Agents Entire introduced a Git Forge concept that redesigns collaboration around coding agents rather than human-authored patches alone. InfoQ中文 ↗
- Benchmark OpenAI’s GPT-6.2 Astra reportedly doubled success rates across 16 embodied-AI tasks versus GPT-6 Astra. 量子位 ↗
- Models Tencent Hunyuan team members said on Reddit that a smaller Hy model is already on their roadmap. 量子位 ↗
- Research A Latent Space report said the Riemann Hypothesis was claimed solved twice independently with AI assistance. Latent Space ↗
- AIGC Meshy announced the jury for its Z-Axis Awards, including 3D artists, researchers, and industry figures. X @MeshyAI ↗
- AIGC Meshy said major brands are supporting its Z-Axis Awards with prizes, credits, and tools for 3D creators. X @MeshyAI ↗
- Open Source Together AI advised developers to tune prompts, inference settings, context, and harnesses before replacing an open model that misses an evaluation. X @togethercompute ↗
- AIGC Kling said it joined Adobe MAX 2026 in Miami to discuss AI video with global brands, agencies, and creators. X @Kling_ai ↗
- AIGC Cohere shared a realistic game clip that appears to showcase AI-generated or AI-assisted interactive content. X @cohere ↗
- Agents AWS Quick added conversational flow building, longer-running workflows, multi-agent orchestration, and new Meta Ads and Amplitude connectors. AWS ↗
- Agents AWS introduced a distributed KV-caching agent skill that reduces self-hosted LLM latency and cost by sharing reusable key-value states across nodes. AWS Machine Learning ↗
- Agents Replit Desktop lets its Agent work with local files and tools, extending cloud-based coding assistance into users’ existing desktop workflows. X @Replit ↗
- Agents StepFun previewed Step 5 at a San Francisco agent event, focusing on agents that run longer, fail less often, and cost less. X @StepFun_ai ↗
- Agents Perplexity Computer now operates inside Chromium through an extension, allowing users to hand off browser tasks from a persistent side panel. X @perplexity_ai ↗
- Open Source Ollama added Thinking Machines’ Inkling 2 Mini to a pre-release build, enabling the 26B model to run locally on Macs with at least 32GB memory. X @ollama ↗
- Agents Runway said agents have written nearly every pull request merged into Runway App Studio since October, according to its engineering update. X @runwayml ↗
- Models Baseten released bison-1, which it says ranks first among 19 decision models on Artificial Analysis Decision Index v2.1 with a score of 76. X @baseten ↗
- Industry Microsoft said Ignite launched MAI Code 2, MAI Code 2 Flash, MAI Cyber 2 Flash, MAI Image 2.7, and MAI Realtime 1. X @MicrosoftAI ↗
- Safety OpenAI published research on combating capability distillation, warning that powerful behaviors can be reproduced without the original model’s safeguards. X @OpenAI ↗
- AIGC Dreamina announced Seedream 5.5 with more realistic portraits, more precise image editing, and substantially improved text rendering. X @dreamina_ai ↗
- Policy Vercel became ISO/IEC 42001:2023 certified, covering how organizations build, manage, and govern artificial-intelligence systems. X @vercel ↗
- Agents Google AI Studio shipped improvements for building Android apps, expanding its role from prompt experimentation toward app-development workflows. X @GoogleAIStudio ↗
- AIGC Pika added Seedance 2.5 Flash for high-volume video creation, supporting clips up to 30 seconds for advertising and rapid iteration. X @pika_labs ↗
- AIGC Together AI made Qwen3 TTS available with nine voices, instruction-controlled speaking styles, and voice consistency across ten languages. X @togethercompute ↗