Discover the AI Tools

灵光App的闪应用平台创作者数量已突破400万，其中大部分是没有编程背景的普通用户。从学生到心理治疗师，从游戏爱好者到网文粉丝，他们用兴趣和专业知识打造出近万款模拟器应用，涵盖升学、职业、偶像养成等主题。AI应用的价值正从效率工具扩展到情感陪伴和生活决策，娱乐成为重要驱动力。

Google has temporarily pulled the plug on its AI image generation feature in Google Earth, just 48 hours after launch. The tool, powered by Nano Banana2, could create realistic fake scenes like the Eiffel Tower collapsing, raising fears of misinformation. Despite digital watermarks, BBC tests found verification gaps. This incident highlights the tightrope walk between creative AI and the need for trust in high-stakes information.

A nonprofit dedicated to AI safety, METR, is struggling to hire enough experts despite offering salaries up to $503,000. With only 35 team members, they can't keep up with the growing demand for AI risk assessment. As AI capabilities double every seven months, the shortage of safety talent is becoming a critical issue.

8月3日，MiniMax开源了其首款多模态生成模型H3，摩尔线程当天便完成了适配。借助MTT S5000智能计算卡和MUSA软件栈，H3在3小时内实现了快速部署。该模型支持文本、图像、音频和视频的全面输入，并能直接输出2K分辨率、最长15秒的影音内容，视频生成成本低至每秒0.8元。摩尔线程表示将继续深化与MiniMax的合作，共同探索多模态领域的更大可能。

Alibaba has officially launched Qwen3.8-Max, a massive 2.4-trillion-parameter model that's already turning heads. It ranks second only to Anthropic's Claude on the Arena leaderboard, with major upgrades in coding and office tasks. The API is live, and open-sourcing is expected next week—a move that could shake up the AI landscape.

On August 3, MiniMax announced the open-source release of its next-generation video model, H3. This multimodal system handles text, images, video, and audio, generating up to 2K resolution videos with synchronized audio. With flexible input options and multilingual support, it's set to democratize AI video creation for filmmakers and designers alike.

West Lake Xinchen, a prominent domestic AI company, has officially announced the completion of its Series B+ financing, securing hundreds of millions of RMB. The round attracted strong participation from various capitals, with Guangfa Xinde leading the investment. The funds will be directed towards advancing next-generation diffusion language models, enhancing multimodal real-time interactive products, and expanding global market presence.

据知情人士透露，蚂蚁集团旗下的具身智能公司——蚂蚁灵波科技，可能启动独立融资，这意味着这家由蚂蚁集团全资孵化的企业将正式走向外部资本市场，接受市场的检验。随着集团在AI领域的投入不断加大，内部算力资源面临压力，引入外部资本成为支持研发和业务扩张的必然选择。目前，具身智能赛道融资环境活跃，蚂蚁灵波的估值有望达到独角兽级别。此外，蚂蚁集团旗下蚂蚁数字科技也在筹备Pre-IPO融资，加上此前已启动独立融资的蚂蚁国际和OceanBase，已有四家子公司开启资本化进程，显示出蚂蚁集团正通过独立融资和市场化运作加速释放各业务板块的发展潜力。

IBM最新报告显示，AI驱动的攻击已占恶意数据泄露的四分之一，同比激增56%，每次事件平均损失600万美元，比整体平均水平高出20%。更令人担忧的是，62%的此类攻击瞄准关键基础设施，金融和能源机构成为重灾区。报告还揭示了攻防成本的不平衡：攻击更便宜，防御更昂贵。

腾讯元宝AI助手迎来图像理解功能全面升级，推出“稳住，先拍个照”新场景，覆盖医疗、合同、家电维修等日常需求。现在，拍张照就能获得专业解读，让AI帮你解决生活小麻烦。

Elon Musk recently demonstrated Grok's new video interaction feature by feeding it an AI-generated video of Kobe Bryant. The AI not only described the visuals and dialogue but also correctly identified the footage as a deepfake, highlighting its advanced reasoning capabilities. This marks a significant leap from audio-only analysis to full video understanding.

South Korea has officially started construction on its National AI Computing Center, a massive 2.5 trillion won project set to open by 2028. With Samsung SDS leading the charge and government backing, the center aims to provide cost-effective AI computing power globally and boost domestic chip development. It's a bold move in the global AI race.

DeepSeek-V4-Flash稳定版正式发布，并同步开放API公测。国家超算互联网平台已上线该模型的API调用服务和文件下载，开发者无需复杂配置，一键即可接入。该版本经过深度后训练，智能体能力和指令跟随显著提升，多项基准测试成绩媲美当前最强闭源模型。平台已汇聚超1700个开源大模型，为AI应用开发提供一站式服务。

AI大模型公司月之暗面（Moonshot）刚完成超35亿美元融资，估值达350亿美元，现又传出最快本月提交港股IPO申请，募资约30亿美元。虽然官方尚未回应，但市场已高度关注。这背后是AI行业加速商业化的缩影，也让人们更加好奇这家明星公司的下一步棋。

OpenAI's new model, Astra, is turning heads with its math skills, but not everyone is convinced. Cognitive scientist Gary Marcus warns that our excitement might be a classic case of overgeneralization. While math is a great testing ground for AI, it doesn't mean Astra is ready for the messy, open-ended problems of the real world. And with OpenAI keeping the details under wraps, we're left wondering: is this a genuine leap forward or just clever marketing?

A Munich court has ruled that AI music generator Suno infringed copyright during both training and output, rejecting its fair use defense. The case, brought by German music rights organization GEMA, centered on six well-known songs. The court found Suno's models could reproduce original elements, a phenomenon known as 'memorization.' This decision could have far-reaching implications for the AI industry's use of copyrighted material.

SenseTime has open-sourced SenseNova U1.5-Lite-Preview, a compact 8B model that supports native 4K image generation. This lightweight model promises better texture detail, improved text rendering, and more stable editing, all while being accessible to developers. It's a shift toward efficiency and practicality in the multimodal AI race.

Tencent's CodeBuddy now fully supports the official release of DeepSeek-V4-Flash, integrating it across WorkBuddy, CodeBuddy IDE, plugins, and CLI tools. This upgrade significantly enhances AI Agent capabilities, outperforming the previous DeepSeek-V4-Pro-Preview in benchmarks. Developers can expect smarter assistance in coding, debugging, and project management, marking a step forward in intelligent programming tools.

Moonshot AI has officially denied reports of filing for a Hong Kong IPO this month. Yet, the company's recent $3.5 billion funding round and its shift to a joint-stock structure suggest big plans ahead. What's really going on? Let's dive into the details.

北京一家名为AGI Bar的酒吧推出创新服务：顾客消费后连接Wi-Fi即可免费获得DeepSeek V4 Flash的API Token，在店内无限量使用AI算力。这一模式将AI服务从付费订阅转变为像水电一样的线下基础设施，引发行业关注。