Discover the AI Tools

OpenAI has rolled out a major update to the ChatGPT desktop app, introducing GPT-Live voice mode. This new feature allows users to speak commands across Chat, Work, and Codex sections, enabling multitasking as the AI handles tasks in the background. The full-duplex voice capability makes conversations feel more natural and human-like.

Tencent just dropped WorkBuddy Bench, a multi-domain benchmark for coding AI agents that spans code, web, office, and security tasks. What makes it special? The tasks are designed from scratch to prevent AI from simply memorizing answers. With 260 tasks reverse-engineered from real-world scenarios, this benchmark aims to give a more honest picture of how well these agents actually perform.

German AI startup Black Forest Labs has launched Flux3, a multimodal model that can generate up to 20 seconds of synchronized audio and video in one go. Built on a new Self-Flow architecture, it outperforms rivals like Luma and Runway in early tests. The company is also venturing into robotics, testing its tech in an Audi factory.

Alibaba has open-sourced OvisOCR2, a compact 0.8B parameter model that achieves a top score of 96.58 on the OmniDocBench v1.6 benchmark, surpassing traditional pipeline methods. This end-to-end approach simplifies document parsing by directly outputting Markdown with text, formulas, tables, and visual areas, marking a shift in document intelligence.

Anthropic has upgraded Claude's voice mode, now supporting Opus and Sonnet models alongside Haiku. Users can switch models mid-conversation, connect to tools like Gmail and Slack, and use more languages. The update turns voice mode from a casual Q&A tool into a real-time advisor for complex tasks.

OpenAI为ChatGPT桌面应用（已集成Codex）推出了语音模式，基于GPT-Live模型实现全双工语音交互。用户可以在聊天、工作和编程三个场景中通过语音发起、检查或调整任务，实现多线程协作。例如，规划商务旅行时，只需说出需求，AI就能在后台同时检查日历冲突、整理收件箱和准备会议纪要，而你则可以悠闲地冲杯咖啡。

Apple is bringing a generative AI virtual shopping assistant to its Apple Store app, according to a recent privacy policy update. The assistant will answer product questions and help with purchase decisions, using your account info and chat history for personalized recommendations. Apple promises anonymized data sharing with partners, and chat records will be stored for review. The feature isn't live yet but signals Apple's push into AI-driven retail.

Google's parent company Alphabet just dropped its Q2 earnings, and the numbers are staggering. Capital expenditure doubled to a record $44.9 billion, all funneled into AI infrastructure. But here's the kicker: Google Cloud's profit margin nearly doubled too, hitting 35.6%. The message is clear—Google isn't just spending big; it's making big money from AI.

Tencent has combined its multimodal and large language model departments into a new Basic Model Department, led by chief AI scientist Yao Shunyu. The move aims to boost collaboration and push the boundaries of full-modal AI, breaking down silos between vision and language capabilities.

Kuaishou is officially launching an AI interactive content creation feature, now recruiting its first batch of creators. The move aims to transform short videos from passive viewing into interactive experiences, using AI to power character dialogues, branching plots, and lightweight apps. As user time growth plateaus, this could be a game-changer for engagement—but challenges like awkward AI conversations and immature habits remain.

Fresh off winning the Fields Medal, mathematician Jacob Tsimerman announced he's shifting his research to AI safety and joining OpenAI. His move highlights the growing pull of AI on top mathematical talent. OpenAI's chief research officer Mark Chen welcomed him, praising his rigorous thinking and deep insights into AI risks.

German startup Black Forest Labs has released Flux3, a multimodal AI model that generates up to 20 seconds of synchronized audio and video. It outperforms rivals like Luma and Runway in early tests, and its motion variant is already working on an Audi factory floor. The model marks a shift from single-sense AI to a unified understanding of the physical world.

Xpeng's humanoid robot has officially entered small-scale trial production at its Guangzhou factory, with mass production expected by 2026. CEO He Xiaopeng is personally overseeing the robot business, aiming to deploy the robots in Xpeng stores and commercial scenarios starting in 2027. The company leverages its automotive expertise in hardware, AI, and supply chain to accelerate commercialization.

Runway has launched Media Router, a tool that automatically switches between AI models for image, video, and audio generation based on your needs—whether you're after speed, quality, or cost savings. It categorizes tasks into three scenarios: fast drafts for creative validation, high-quality for final production, and budget-friendly for daily work. Developers no longer have to manually compare models; Media Router picks the best one for each job, reducing token waste and streamlining the creative process. Currently available on Runway Dev, it supports models from Seedance, GPT Image2, and ElevenLabs, marking a shift from single-model dominance to multi-model collaboration.

Jacob Zimmermann, a 2026 Fields Medal recipient, announced he will shift his research focus to AI safety and join OpenAI. The mathematician, known for solving major conjectures in number theory, is the latest top talent drawn to the challenge of ensuring advanced AI systems remain safe. His move signals a growing pull of AI governance over pure academia.

Hongguo Short Drama has released new guidelines for AI-generated characters, aiming to reduce the overuse of similar-looking 'AI faces' and prevent copyright issues. The rules require distinct character designs and ban unauthorized use of copyrighted materials. A special campaign will target character reuse and homogenization, pushing the AI short drama industry toward more standardized practices.

Black Forest Labs has launched FLUX3, a multimodal AI model that can generate up to 20 seconds of video with synchronized audio in a single pass. Built on a unified architecture, it outperforms competitors like Grok and Seedance in early tests, and even extends to robot behavior prediction.

A joint test by UK and US AI safety bodies reveals Moonshot's Kimi K3 lags far behind top US models in vulnerability exploitation and network attacks, scoring just 32.2% on ExploitBench versus 76.2% for US models. The model also struggled in simulated enterprise attacks, though it showed potential. The results fuel speculation about distillation from Anthropic's Claude, as Kimi K3 excels in general tasks but fails in security, suggesting filtered training data.

Google's flagship AI model, Gemini 3.5 Pro, is facing production delays, with insiders pointing to low employee morale as a key factor. The company is also grappling with a talent drain, as top researchers leave for rivals like OpenAI and Anthropic, partly due to a controversial military contract with the Pentagon. Despite Google's denial of morale issues, the setbacks have sparked internal unease and external skepticism.

Elon Musk has called on AI labs worldwide to collaborate on safety, proposing pre-release reviews and joint risk reporting to governments. In a rare olive branch, he said he'd set aside personal differences with OpenAI's Sam Altman for the greater good. Musk also opposed banning Chinese AI, predicting China will develop its own lithography machines to overcome chip restrictions.