Discover the AI Tools

Google just dropped LiteRT.js, a new library that makes AI run way faster in web browsers. By tapping into WebAssembly and hardware acceleration like WebGPU and WebNN, it can speed up AI inference by up to three times compared to the old TensorFlow.js approach. That means smoother AI features on websites without needing to install anything—even on older devices.

Tencent is in talks to take over Manus, a Singapore-based AI startup, becoming its largest shareholder in a $2 billion deal that replaces a previously planned acquisition by Meta. The move aims to bolster Tencent's AI ecosystem, particularly its WeChat AI agent, while Manus eyes a Hong Kong listing.

Step Star is set to unveil its first AI agent phone on July 13, making it the first major large model company to enter the smartphone market. The device, manufactured in partnership with Huqin Technology, aims to deliver a native AI experience from the edge. OpenAI had previously announced plans for a similar terminal but not until 2027.

Samsung is quietly developing a 4nm AI PC chip called GAIA, designed to accelerate generative AI tasks. The chip, which uses a novel memory-centric architecture, has already been sampled by Lenovo and HP. Mass production is expected by 2027, positioning Samsung as a key player in the competitive AI PC market.

Kimi, American Express, and Agricultural Bank of China have jointly launched the world's first AI-native credit card, the 'Kimi Credit Card.' Cardholders can earn points that convert into Kimi membership and computing power, creating a 'spend to produce' value loop. This marks a significant step in integrating AI services into financial products.

Samsung is making a comeback in the PC chip market with its self-developed AI accelerator, GAIA. Built on a 4nm process and centered around an NPU, the chip is designed to speed up generative AI tasks. Prototypes have been sent to Lenovo and HP for testing, with mass production slated for 2027. Samsung also eyes robotics applications, aiming to compete with NVIDIA, Qualcomm, and others in the AI PC arena.

Meta has launched Muse Spark 1.1, a flagship AI model designed to handle complex multi-agent tasks with ease. It features a context compression mechanism that manages up to a million tokens, allowing it to juggle multiple sub-agents without losing track. In coding tests, it scored 72.2 on the Vibe Code Bench, a huge leap from its predecessor. The model is now available for preview via Meta's API and chatbot.

今天的AI圈热闹非凡：OpenAI推出GPT-5.6系列，主打安全与性价比，但Codex不见了；腾讯计划以20亿美元估值入股Manus，成为最大股东；MiniMax完成160亿港元融资，创始人闫俊杰宣布零薪直到AGI实现。此外，ChatGPT扩展进驻Chrome侧边栏，Anthropic推出AI使用健康报告，字节跳动Seedance 2.5即将开放API，蚂蚁发布具身世界模型，三星开发4nm AI芯片。

Kuaishou's KwaiKAT team has released KAT-Coder-Pro V2.5, an AI model that can autonomously handle complete software engineering tasks. It upgrades long-term engineering, general agentic capabilities, and uses large-scale reinforcement learning. The model achieves a 57.2% success rate in building runnable environments and scores 65.2 on SWE-Bench Pro. It's now available on StreamLake platform.

OpenAI 更新了 ChatGPT 的 Chrome 扩展，让 AI 助手直接嵌入浏览器侧边栏。现在，ChatGPT 不仅能回答你的问题，还能主动浏览网页、提取信息、控制标签页，甚至访问本地文件。这意味着你无需切换应用，就能在浏览器里完成更多任务。

OpenAI has unveiled the GPT-5.6 series, featuring Sol, Terra, and Luna models. Sol leads in coding benchmarks, while all models emphasize cybersecurity and cost-effectiveness. The release includes ChatGPT Work for enterprise teams, signaling a shift from parameter competition to practical value.

OpenAI has launched a new desktop application that merges ChatGPT and Codex into a single tool. The app features three modules: Chat for conversations, Work for office tasks, and Codex for programming. This move aims to simplify workflows for both developers and non-technical users, offering a unified AI assistant experience.

Lyzr, a three-year-old enterprise AI startup, just closed a $100 million Series B round—and the whole thing was handled by its own AI agent, SivaClaw. The system fielded questions from over 130 investors, drafted investment memos, and even tracked how long investors spent on slides to gauge interest. The company's valuation now sits around $500 million, and the founder never had to leave his desk for meetings on Sand Hill Road.

China Mobile has launched a new service called 'New Message Claw' that lets users remotely control desktop AI agents via SMS—no app download needed. It solves the problem of missing agent notifications when away from your computer. Compatible with several open-source agents, it uses operator SMS channels for low-latency, reliable control.

OpenAI's second-in-command, Fidji Simo, is stepping down from her full-time role due to a recurring neuroimmune disease. She'll stay on as a part-time consultant. The departure comes at a critical time as OpenAI faces slowing ChatGPT growth and prepares for a potential IPO, putting pressure on CEO Sam Altman to find a successor quickly.

Google has revamped its Android Bench code ranking, introducing a secure Harbor sandbox framework for fairer testing. The new benchmark sees Anthropic's Claude5 top the list with 84.5% accuracy, while Google's own Gemini models struggle—Gemini 3.5 Flash took 28 hours and $165 to run. The move highlights the industry's shift toward autonomous development workflows and positions Android Bench as an authoritative AI evaluation platform.

Apple is reportedly in talks with AI startup PrismML to bring large language models directly to iPhones. Their 1-bit compression tech shrinks model size to 1/14, boosts speed 8x, and cuts power use by 80%. A 27B-parameter Qwen model already runs on an iPhone 17 Pro, hinting at a major shift in on-device AI.

Meta is collaborating with Broadcom and TSMC to develop its own AI data center processor, codenamed Iris, which is set to enter mass production in September 2026. The chip aims to optimize ranking and recommendation systems on Facebook and Instagram, reducing reliance on third-party GPUs and cutting costs. Meta also plans to deploy 7 gigawatts of computing power by end of 2026, with infrastructure spending reaching $14.5 billion.

Anthropic has launched a test version of its Reflect feature, turning Claude into a personal AI usage analyst. The tool generates a 'health check' report covering your chat patterns, task types, and collaboration style across four dimensions. It's like getting a physical exam for how you work with AI—minus the awkward gown. Privacy is a priority: incognito chats and health data stay off-limits.

OpenAI released GPT-5.6 on July 9, positioning it as the preferred model for Microsoft 365 Copilot, integrating deeply with Word, Excel, and PowerPoint. This move counters rumors of a rift between the two tech giants, as Microsoft had been developing its own MAI model. The partnership underscores ecosystem binding as a key strategy amid fierce AI competition.