Discover the AI Tools

Moonshot's next flagship, Kimi K3.1, may arrive as soon as October, according to a leak on X. The model reportedly offers Low, High, and Max reasoning levels, a million-token context window, and Agent plus Swarm multi-agent collaboration. Here's what we know so far.

当大模型自己生成训练数据，生成与训练的规模却不同步，批量大小该怎么定？腾讯混元团队从经典的临界批量大小理论出发，重新推导了在线强化学习场景下的最优解。结果很实在：PPO吞吐量最高提升2.29倍，GRPO训练时间省下29%。这意味着一把可操作的效率旋钮，让现有硬件跑得更划算。

Imagine catching esophageal cancer early—without a tube down your throat. Alibaba's DAMO Academy just unveiled DAMO EAGLE, an AI that reads plain CT scans and flags tumors with 90% sensitivity. Validated on 80,000 cases across three countries, it could ride along with routine lung cancer screenings, no extra radiation needed. The goal isn't to replace endoscopy but to pinpoint who needs it most. In a country where half the world's cases occur, that shift matters.

China's live-streaming e-commerce is now a 7 trillion yuan giant. As AI anchors flood the market, slashing costs by up to 90%, the big question remains: will real hosts become obsolete? Industry insiders say no—but the game is changing fast. Here's what's really happening behind the screens.

Huawei just launched Xiaoyi Work, a 24/7 AI assistant that handles market research, PPT creation, data analysis, and more. Just tell it your goal, and it breaks down tasks, runs them on cloud or local computers, and keeps working even when you're away. New users get 1000 free AI points for 31 days. It's Huawei's big push into long-task automation, competing with other AI assistants.

Zhiyuan just delivered its 20,000th Yuzheng A3 Ultra robot — a massive jump from 10,000 units only six months ago. Meanwhile, over 300 of its robots have started regular shifts at Chimelong's Spacecraft Park in Hengqin, marking the first large-scale deployment of embodied intelligent robots in a theme park. From factory floors to tourist crowds, these machines are stepping into the unpredictable real world.

Tencent just dropped a new translation app called Hy Translate, and it's aimed squarely at travelers. The app handles 33 languages, works offline, and even supports five minority languages and dialects. It's rolling out in the US, Japan, South Korea, and Southeast Asia. Voice, photo, and conversation translation are all included. Could this be your next travel buddy?

At Meta Connect, Zuckerberg unveiled a trio of smart glasses and a camera-free Ray-Ban model for privacy buffs. The real star? Muse, Meta's AI assistant, now lives inside your frames and even has its own inbox. With a lightweight VR headset on the horizon, Meta is betting big on AI hardware—but can it avoid the fate of Humane and Rabbit?

OpenAI just gave ChatGPT a voice agent that actually does things. Pro and Plus users can now speak to create documents, summarize emails, or even build websites. Free and Go users get plugins and connected apps. It's not just Q&A anymore—it's getting stuff done.

A leak suggests Moonshot AI's Kimi K3.1 could arrive as early as October 2026, bringing three adjustable reasoning levels, a million-token context, and built-in Agent and Swarm modes. The clues come from internal code, not an official announcement—so treat the details as speculation for now.

At the 2026 Snapdragon Summit, Qualcomm and PrismML squeezed a 2-billion-parameter vision model onto smart glasses—no cloud required. The 1-bit Bonsai model slashes memory to a quarter of 4-bit versions and doubles token speed. So you can point, ask, and get answers without a connection, while sipping less battery and staying cooler.

Google DeepMind's Demis Hassabis just spilled the beans: Gemini 4 is in post-training and might drop well before the end of the year. The company is skipping Gemini 3.5 Pro to focus on this next flagship. Engineers are already using it internally for coding, and safety tests are underway. Hassabis says he's '100% confident' in the team, but what about AGI? He thinks we're asking the wrong question.

你听说了吗？DeepSeek的年化收入已经突破10亿美元，短短几个月就翻了一番！更劲爆的是，他们正计划融资500亿人民币，目标直指上海证券交易所。CEO梁文锋在投资者会议上透露了这个消息，看来AI大模型的商业化真的跑通了。

华为刚刚推出小艺工作，一个能帮你搞定调研、写文档、做PPT、分析数据甚至写代码的AI助手。它深度适配鸿蒙手机、平板和电脑，任务能在设备间无缝流转。新用户还能领31天1000 AI点数的免费试用，这波福利别错过。

At the UN, OpenAI, Anthropic, and Hugging Face urged countries to adopt shared AI safety tests and incident reporting. The Trump administration pushed back, rejecting any global governance with binding power. Both sides agree frontier risks are real — they just can't agree on who should manage them.

Tencent Hunyuan just rolled out its new translation app, Tencent Hy Translation. It handles 33 languages—plus five Chinese minority dialects—and works offline. You can chat, snap photos, or download language packs before you travel. The app is now available in 12 countries and regions, including the US, Japan, and Southeast Asia. It's built on the open-sourced Hy-MT2 model, which launched in May.

蚂蚁集团旗下的 InclusionAI 在 HuggingFace 上悄悄开了个叫 AI-Transparency 的仓库，把 Ling、Ring、Ming 这几个模型系列的训练内容摘要一股脑儿公开了。不涉及模型权重，也不给训练数据，但每份文档都严格照着欧盟《人工智能法案》第 53 条的模板来写。监管越来越严，这波操作既是合规，也让人窥见模型是怎么炼成的。

Imagine a TV drama with zero cameras and zero sets. That's 'The Later Journey to the West,' which just hit Hunan Satellite TV's prime time. All 60 episodes were generated by AI model Seedance. The cost? Tens of thousands per episode, down from millions. A 100-person team rethought everything — and it's changing how film and TV get made.

Ever notice Claude's writing feels off? An Anthropic engineer spills the beans: the model was optimized for math and code, so it learned to write for other AIs—dense, jargon-heavy, and hard for humans to digest. The good news? The latest version is finding its voice again.

Google just dropped two text-to-speech models that let you craft custom voices with simple descriptions—no technical know-how required. The Gemini 3.8 Flash TTS even replicates a voice from just 30 seconds of audio. With support for over 100 languages and 2,000 ready-made voices, this could change how creators bring characters to life.