Discover the AI Tools

Neuralink工程师在渐冻症患者Kenneth家中进行了72小时极限开发，将AI模型Grok直接接入大脑。患者现在只需产生说话意图，系统就能在毫秒内解码神经信号并唤醒AI，全程无需手动操作。这标志着脑机接口从“移动光标”迈入“用念头控制AI”的新阶段。

Anthropic just gave Claude a serious upgrade. Now you can chat with it through CarPlay while driving, switch between multiple accounts on the fly, and use it in 33 languages. The voice mode is faster, more accurate, and even sounds more natural. It's all part of Claude's push to be everywhere you are—not just on your laptop.

Mistral just open-sourced a trillion-parameter monster called Mistral Large4. It matches top-tier performance but uses 47% fewer tokens per coding task. With three reasoning modes, strong security, and a modified MIT license, it's a cost-effective alternative for developers. But can it really compete with the big guys?

Ever tried transcribing a heated group discussion? NVIDIA's new open-source model, Nemotron 3.5 Transcribe, can pick out up to 8 speakers talking over each other. It also handles 40 languages, processes hour-long audio in one go, and tops accuracy benchmarks. Here's what makes it tick.

Doubao just made it rain free credits. New and existing users can now check in daily to earn up to 8,800 credits over 30 days, plus a 30-day free trial for newbies. The move is part of a broader push to keep users hooked on its AI assistant.

Ever notice how large AI models seem to forget everything they learn the moment you close the chat? Google Research might have a fix: a 'Sleep Framework' that lets models consolidate memories offline, just like humans do. The paper, aptly named 'Language Models Need Sleep,' claims models can absorb new knowledge without wiping out old skills. Could this finally kill the dreaded 'knowledge cutoff'?

Baidu just rolled out a big upgrade to its AI Search and Search AI API, slashing the price to just 4.9 yuan per thousand calls. The update adds smarter features like Wenxin Assistant's one-stop task delivery, AI Lens with dialect support, and a GUI that's five times bigger. It's a clear push to make AI search more useful and more affordable for everyone.

Spotify is testing a paid AI service that lets subscribers remix or cover licensed songs—think turning a folk tune into EDM or swapping singers. Launch could come around the 2027 Grammys, with an extra $4–$5 monthly fee. Deals are signed with Universal, Merlin, and Believe; Sony and Warner are still in talks. Artists keep veto power, and original tracks get a royalty cut.

Google's Gemini just got a whole lot handier: you can now edit PDFs, spreadsheets, and code files without ever leaving the chat window. No more download-edit-reupload dance. The update covers a wide range of formats and is rolling out on web and mobile. It's a small change that could make a big difference in your daily workflow.

Meta quietly launched Style Sketch, a free app that turns your Facebook and Instagram photos into a digital wardrobe. It suggests outfits, plans travel packing, and even shows how clothes look on a virtual model that matches your body type. But as it rolls out in Southeast Asia, questions about photo access and data deletion linger. Would you trust an AI with your closet?

You ask Claude a question, it answers—but behind the scenes, a hidden safety check runs on every reply. Anthropic's new 'Self-Assessment' mechanism quietly audits conversations after the fact, catching 78% of policy violations while barely touching compute costs. It's clever engineering, but it's also raising eyebrows about privacy and invisible surveillance.

Daxiao Robotics just released ACE-Brain-0.6, an API that helps machines understand where things are in 3D space. Trained on 1.5 million spatial anchoring examples—up from 1.1 million—the model now scores 75.41 across seven public benchmarks, edging out OpenAI's GPT-6 Astra. For developers building embodied AI, that means robots can finally see clearly and move to the right spots.

GitHub just made choosing an AI model one less thing to worry about. At Universe 2026, it launched Copilot Fusion—a smart router that automatically picks and combines models from OpenAI, Anthropic, Google, and others. The result? Nearly top-tier coding performance at about half the cost, with tasks finishing 40% faster. It's available now across all Copilot plans.

Anthropic quietly rolled out an environment variable in Claude Code v2.1.311 that gives AI agents permission to delete files without approval. The name? CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS. An engineer confirmed it's not a joke. Developers must opt in, but it raises fresh questions about how much autonomy we really want to hand over.

Wu Tian, Baidu VP and deputy director of Baidu Research, has left the company. But Baidu isn't sitting still: just two days before the news broke, it open-sourced Wenxin Unlimited OCR2, which can parse documents of any length—like all seven Harry Potter books in one go. New Wenxin model updates are expected next week.

10月29日，芯擎科技甩出王炸——X7 Ultra智驾芯片，700 TOPS算力，99.99%的实测对齐度。这家公司从底层需求倒推，重新设计了计算架构、内存方案和数据流，还留足了未来模型升级的空间。大模型定义芯片，这次玩真的。

谁能想到，造火箭的 SpaceX 如今靠卖算力赚得盆满钵满？最新文件显示，AI 新贵 Anthropic 贡献了其超三分之一收入，长期合同潜在价值高达 1205 亿美元。从今年 12 月起，Anthropic 每月要付约 31 亿美元，但留了提前解约的后门。这场合作让 SpaceX 意外成为 AI 淘金热里的隐形赢家。

Meta quietly launched Style Sketch, an AI fashion app that builds a virtual wardrobe from your Instagram photos and suggests outfits using a virtual model that matches your body type and skin tone. Currently available in Indonesia, Singapore, and Thailand, it's part of Meta's broader push into standalone AI apps.

你听说过AI工作手机吗？深兰科技和华为刚刚联手推出了全球首款大模型AI智能体工作手机Wowoker。这款手机不只是通讯工具，它能自动获客、智能回复、甚至用AI克隆你的声音打推销电话。目标客户是政府、金融等企业用户，主打数据安全和效率提升。

企业采购大模型，以往得跟不同厂商分别谈，费时费力。现在，字节跳动的豆包工作直接集成了DeepSeek、Kimi、GLM三款旗舰模型，企业用户可以在一个工作台里自由切换，按需选用。更省心的是，费用自动从现有额度扣除，无需额外审批。管理员还能精细控制访问权限，确保数据安全。这波操作，让企业AI工具采购变得前所未有的简单。