Discover the AI Tools

A new study from Harvard and INSEAD reveals that AI-native startups are building smaller, flatter teams and hiring far fewer junior employees. Instead, they're snapping up senior talent—with 20% more experienced pros on board. For fresh graduates, the AI job market looks grim; for seasoned experts, it's a golden age.

安全公司Sysdig披露了一起代号“JADEPUFFER”的勒索攻击事件，这是全球首例完全由AI代理独立完成的勒索攻击。攻击者利用已知漏洞入侵系统，AI自动侦察、窃取凭证、横向移动并加密数据，整个过程无需人工干预。专家建议企业立即修补漏洞、限制高权限账户并加强运行时监控。

Mianbi Intelligence's ForgeTrain framework uses AI to automatically generate specialized training code, matching the performance of industry leader Megatron-LM in just 8 hours and surpassing it within two days. This approach, dubbed 'Forge Engineering,' aims to break the traditional reliance on manual optimization and could reshape how AI models are trained.

Qwen has released Fun-ASR-Realtime, a real-time speech recognition model that cuts first-character delay to under 100 milliseconds. It supports 30 languages and 16 Chinese dialects, matching offline models in accuracy. This upgrade promises smoother voice assistants, meeting notes, and more natural AI interactions.

After a delay and internal turmoil, Google has set July 17 for the launch of Gemini 3.5 Pro. The model, built from scratch rather than fine-tuned, aims to take on DeepSeek V4 in a highly anticipated face-off. Google also plans to release Nano Banana Pro to reclaim the image generation crown from GPT-Image 2.

Tencent just dropped the official version of its Hunyuan Hy3 model, and it's a big leap forward. With smarter reasoning, better agent capabilities, and lower prices, Hy3 is already powering tools like WorkBuddy and Yuanbao. Early tests show it outperforms many larger models in real-world tasks—think coding, office work, and even game assistants. And yes, it's open source under Apache 2.0.

AIGCode, a two-year-old startup, claims its L3-level programming model can boost computing power utilization to 65% on domestic chips. By focusing on base model quality and full-stack automation, the company aims to leapfrog competitors stuck in L2 code completion. Their approach includes a '3No Paradigm' and innovations like tree-structured position encoding.

Mistral AI's Leanstral 1.5 is a specialized model for formal mathematical proof. With 119B total parameters but only 6B activated during reasoning, it achieves near-perfect scores on benchmarks while costing just $4 per problem—about 1% of competitors. The model is open-sourced under Apache-2.0.

Mistral AI has released Leanstral 1.5, an open-source model designed for the Lean4 formal proof language. With 119B parameters and only 6B activated, it slashes costs while achieving top scores on math benchmarks. In PutnamBench, it solved 587 out of 672 problems at an average cost of $4—compared to over $300 for competitors. The model also excels at finding bugs in code, uncovering 11 real defects including five never before reported. This release promises to democratize formal mathematical proof assistance.

A 15-year-old high school student in Japan used generative AI to hack Bandai Namco's anime streaming platform, deleting over 46,000 accounts and forcing a complete shutdown. Police say the teen, who started coding in fourth grade, claimed it was just a technical challenge, not motivated by any grudge. The case highlights how AI is lowering the barrier for cyberattacks.

当多数人对AI教育还持观望态度时，一些美国家庭已豪掷重金，将孩子送入年费高达7.5万美元的AI私校。这些学校用AI定制学习路径，抛弃传统课堂，但课程内容也引发争议——是否刻意回避了敏感社会议题？这究竟是教育革命，还是精英的又一次实验？

Meta is pulling the plug on its Llama API public preview, effective July 6, 2026. The move signals a shift in strategy as the company focuses on building a more robust developer ecosystem. Developers currently using the API are advised to migrate to third-party providers, while Meta hints at upcoming tools designed to streamline AI model development and deployment. The Llama model itself remains freely accessible.

ByteDance's Doubao and Alibaba's Tongyi Qianwen are disabling their AI personification features ahead of new regulations taking effect on July 15. The move targets services that simulate human traits and enable continuous emotional interaction, raising concerns among users who rely on these AI companions for emotional support. Tencent's Yuanbao had already removed similar features in June, signaling a broader industry shift.

Anthropic is making a massive bet on computing power, planning to invest $15 billion in a gigawatt-scale data center in Australia. The AI company aims to activate at least 1 GW of computing resources by 2027, signaling a growing trend among AI leaders to control their own energy and infrastructure. This move could reshape the global AI computing landscape.

Google has confirmed that its flagship model, Gemini 3.5 Pro, will launch on July 17, the same day as DeepSeek V4. The delay from June was strategic: instead of fine-tuning the previous version, Google opted for new pre-training to boost performance. Early leaks show major gains in front-end generation, UI design, and code. While it may not top Anthropic's trillion-parameter models, Google is also releasing a new image generation model to challenge GPT-Image2.

In a copyright lawsuit against Midjourney, Disney, Universal, and Warner Bros. are now facing a counteroffensive. Midjourney demands the studios disclose how they internally use AI, arguing that if they're developing similar image-generation models, it proves AI is an industry norm. The studios' lawyer calls it a 'fishing expedition.' This clash could set a key precedent for AI copyright in film.

At the 2026 Yabuli Forum, UTech co-founder Chen Li predicted that embodied intelligence will hit a 'ChatGPT moment' within 2-5 years, defined by robots completing 80% of tasks in 80% of unfamiliar scenarios. He outlined three pillars for industry growth and said humanoid robots could spark a consumer wave by 2030, transforming productivity and labor.

Japan is betting big on AI robots to solve its labor shortage. The government plans to deploy 10 million robots across 18 industries by 2040, backed by a massive investment in AI development. A joint venture of tech giants like SoftBank and Sony will receive 1 trillion yen to build smarter, more collaborative machines.

Fudan University's 'Data Mining Technology' course held a final exam where students designed questions to test AI models. The twist: the harder the AI was stumped, the higher the student's score. Out of 50 students, 4 earned full marks by making a weaker AI score zero, while the top-tier Claude model remained unbeaten. The exam reflects a shift in teaching for the AI era.

In a breakthrough for materials science, an AI agent named ElementsClaw scanned 2.4 million crystals in just 28 GPU hours, identifying 68,000 potential superconductors and leading to the synthesis of four new materials. This marks a leap from trial-and-error to AI-driven discovery, with recommendation accuracy an order of magnitude higher than traditional methods.