August 19, 2026
Quick hits from across the AI world in the last 24 hours.
- ⚡ Quick hits
- Research MT Lab proposed an ICML 2026 method for seamlessly editing multilingual text embedded in complex images. InfoQ 中文 ↗
- Industry Microsoft is recruiting a chief legal engineer combining AI and legal expertise, offering annual compensation approaching $280,000. IT之家 ↗
- Models Anthropic reportedly uses an unreleased system called Model 2 internally while withholding it from customers. The Decoder ↗
- Agents A KDC architecture article argues that retaining interaction history alone does not give long-running agents functional memory. InfoQ 中文 ↗
- Agents A reported Kimi K3 research system coordinates 18 agents, with its harness helping performance approach Claude Opus 5. 量子位 ↗
- Agents Binance opened trading access to AI agents, while placing much of the responsibility for agent controls on users. TechCrunch ↗
- Industry China’s AI sector is developing circular financing arrangements linking chip suppliers, model companies and infrastructure spending. The Decoder ↗
- Safety Terence Tao warned that unreliable AI-generated mathematical results could create the field’s largest foundational crisis since Gödel. The Decoder ↗
- AIGC Alibaba demonstrated Wan 3.0 turning uploaded documents, images or websites into generated creative outputs. X / Alibaba Wan ↗
- Agents OpenAI highlighted companies embedding its open-source Codex harness into internal applications, operations dashboards and customer-facing builders. X / OpenAI Developers ↗
- Industry Vercel’s v0 temporarily cut prices for GPT-5.6 Sol and Fast mode by 50% through September 18. X / v0 ↗
- Open Source Smolmachines introduced smolvm, a WebAssembly-based sandbox for executing untrusted Python and JavaScript generated by AI systems. Simon Willison ↗
- AIGC Pika showcased Seedance 2.5 extrapolating editing style, music and camera language from several visual references at 1080p resolution. X / Pika ↗
- Agents Google Cloud launched its All Things Agentic hackathon for developers building Gemini-powered agents that automate complex work. X / Google Cloud Tech ↗
- Models DeepSeek-V4-Pro High rose to second among open models in Agent Arena, with a reported 6% net improvement and $0.21 median task cost. Agent Arena on X ↗
- Agents Amazon Bedrock AgentCore Web Search added domain allowlists, blocklists and publication-date filters for more controlled agent retrieval. AWS Machine Learning Blog ↗
- Open Source Nvidia said open-source cuOpt became the fastest solver across three optimization classes in Hans Mittelmann’s benchmark suite. NVIDIA AI on X ↗
- Safety Google Cloud open-sourced a zero-trust agent blueprint that assumes model compromise and enforces sensitive controls in infrastructure. Google Cloud Tech on X ↗
- AIGC Pika detailed its Pika SFX model, which generates prompt-aligned, production-ready sound effects in under one second. Pika on X ↗
- Open Source TrueFoundry open-sourced TrueForge, a vendor-neutral agent harness covering context management, MCP tools, subagents, sandboxing, approvals and persistent state. Together AI on X ↗
- Agents Vercel Agent can now join Slack threads with production context, create plans and pull requests, and roll back deployments. Vercel on X ↗
- AIGC Krea introduced Seedance Studio, combining Seedance 2.5 with prompt presets, camera paths and 3D scene controls in one creation workspace. Krea on X ↗
- Agents Google AI Studio added GitHub synchronization, letting Build users import repositories and push or pull changes across environments. Google AI Studio on X ↗
- AIGC ElevenLabs released Eleven v3 Conversational for real-time applications, offering expressive audio controls and support for more than 70 languages. ElevenLabs on X ↗
- Agents Replit introduced a Free Mode designed to make its AI-assisted application creation workflow available without upfront usage costs. Replit on X ↗
- Agents Nvidia introduced SkillEvaluator after benchmarking more than 300 verified skills to measure their effect on agents completing identical real-world tasks. NVIDIA AI on X ↗
- Open Source Unsloth released Dynamic 3.0 GGUFs for Qwen3.8-27B, reporting over 10% better accuracy than competing quantizations on several benchmarks. Unsloth on X ↗
- Agents Mem0 launched a Playground for testing and tuning what application agents remember and how strictly they retrieve stored memories. Mem0 on X ↗