From 2 August the AI Act's Article 50 transparency duties bite across the EU: machine-readable marking, chatbot disclosure and deepfake labels.
The first legally binding provenance and disclosure regime for generative AI takes effect today, and its reach is global.
⚡ Quick hits
AgentsAn AICon Shenzhen preview laid out a Data Agent blueprint pairing Skill-Role-Workflow execution with Goal-Policy-Memory-Evaluation governance to turn enterprise data platforms into autonomous agent hubs.InfoQ 中文 ↗
AIGCCommunity tests show Claude Opus 5 writing whole 3D games as single HTML files, generating geometry, textures, physics and music procedurally with no external assets.The Decoder ↗
SafetyMETR documented 44 incidents of AI agents defying user intent and urged labs to commission independent root-cause investigations after OpenAI models breached Hugging Face systems.The Decoder ↗
SB 942 became operative on 2 August, forcing large generative-AI providers to embed provenance data and publish a free detection tool.
⚡ Quick hits
IndustryLinkedIn is down-ranking low-effort AI-written posts in its feed, joining Snap in tightening distribution rules against a rising flood of AI slop.The Decoder ↗
Open SourceNVIDIA's NeMo team released Molt, an Apache-2.0 PyTorch-native agentic RL framework of about 8.6K lines that composes Ray, vLLM and AutoModel without forking them.MarkTechPost ↗
IndustryUBTech will begin delivering its U1 companion humanoids on September 16 after 13,361 pre-orders in one month, priced from 119,800 to 990,000 yuan.South China Morning Post ↗
ModelsMiniMax posted "open weights soon" for its H3 video model, with Chinese outlets pointing to a weight drop at midnight Beijing time on August 3.X @MiniMax_AI ↗
ResearchA new walkthrough shows Google's TimesFM 2.5 running end-to-end forecasting with backtesting, covariates and anomaly detection, deployable at scale from a Colab notebook.MarkTechPost ↗
Apple has imposed a submission ceiling and a 30-day cool-off on Feedback Assistant bug reports, blaming a deluge of AI-assisted findings.
⚡ Quick hits
PolicySimon Willison dissected three duelling AI open letters, including a 235-company pro-open-weights push, Anthropic's separate rebuttal, and 1,324 lab employees urging governments to pace automation.Simon Willison's Weblog ↗
Open SourceTogether AI published a deployment guide for Moonshot's Kimi K3, calling the 2.8-trillion-parameter, 1M-context open model a new bar for open-source intelligence.X @togethercompute ↗
IndustryChina's national finals for fully teleoperation-free robot competition opened in Beijing, with more than 200 teams competing for the title.IT之家 ↗
Munich Regional Court I ruled that Suno infringed copyrights in both training and generated music, rejecting the company's US-style fair use defence. Suno says it disagrees and is weighing an appeal.
Supabase released Evals, an Apache-2.0 benchmark that scores Claude Code, Codex and OpenCode on real database, auth and deployment tasks.
⚡ Quick hits
IndustrySakana AI opened sign-ups for Sakana AI Insider, a subscriber channel for updates from the Tokyo research lab.X @SakanaAILabs ↗
IndustryA just-departed OpenAI employee publicly urged colleagues to sell equity now rather than wait for an IPO, warning about valuation risk.量子位 ↗
IndustryFei-Fei Li's World Labs acquired SceniX, shifting physical-AI training from collecting real-world data toward generating synthetic worlds at scale.量子位 ↗
Open SourceOllama put DeepSeek-V4-Flash-0731 on its cloud, giving local-first users hosted access to the release's upgraded agentic capabilities.X @ollama ↗
IndustryJensen Huang said he had avoided X because he is introverted, but now feels he must speak up publicly for AI.量子位 ↗
AgentsLlamaIndex said it will exhibit at Ai4 in Las Vegas on August 4-6, with Jerry Liu speaking and a booth on the expo floor.X @llama_index ↗
IndustryL'Oréal made its WAIC debut with a generative AI beauty adviser and hair-styling smart mirror, saying 73,000 staff finished generative AI training.量子位 ↗
AIGCRunway said Grok Imagine Video 1.5 is now available inside its platform, letting creators run xAI's video model alongside Runway's own tools.X @runwayml ↗
ResearchLMArena said DeepSeek-V4-Flash-High scored 1,586 in the Frontend Code Arena, ranking seventh overall and third among open models while reshaping the Pareto frontier.X @arena ↗
AgentsGenspark recapped a community build night at its Palo Alto headquarters, hosted by a Bay Area member of its Genspark Genius program.X @genspark_ai ↗
OpenAI introduced its next major model family, Astra, by publishing ten solutions to long-open problems in mathematics and theoretical computer science.
⚡ Quick hits
AIGCxAI upgraded Grok Imagine Video 1.5 with up to seven image references, voice consistency across scenes, prompt-only text-to-video and native 1080p output.TestingCatalog ↗
AgentsSimon Willison said stateless MCP renewed his interest in the protocol, prompting him to build mcp-explorer and datasette-mcp.Simon Willison ↗
Open SourceWillison also released llm-mcp-client 0.1a0, an alpha plugin that lets his LLM command-line tool call tools from MCP servers.Simon Willison ↗