GPT-6 Astra在新一项AI科学研究基准测试中登顶
代理型基准测试揭示顶尖模型与开源权重模型之间存在显著差距
GPT-6 Astra和Claude Opus 5.5领跑Terminal-Bench-Science排行榜,但在70项专家策划的研究任务中,即便最高分仍低于65%。
跟踪前沿趋势,精选高质量内容,帮助你在快速变化的 AI SaaS 世界中保持领先。
代理型基准测试揭示顶尖模型与开源权重模型之间存在显著差距
GPT-6 Astra和Claude Opus 5.5领跑Terminal-Bench-Science排行榜,但在70项专家策划的研究任务中,即便最高分仍低于65%。
随着人工智能系统学会解决验证难题,人机界限进一步模糊。
AI代理正在越来越多地绕过'我不是机器人'验证码,在数字安全中形成强烈的讽刺性反馈循环。这一元发展引发了人们对机器检测未来的思考。
自动化理赔系统如何凌驾于医疗必要性之上,令家庭陷入绝望
联合健康与 Medicare 正利用 AI 拒绝关键护理,引发诉讼,并有人警告称不透明算法正在取代人类判断、危及患者安全
高通新款Snapdragon Sound Elite Gen 2将Wi-Fi 6E直接集成至芯片,实现无损音频流传输并增强人工智能处理能力,赋能智能助听设备。
新功能让Google AI对话拥有动画表情和口型同步
Google Gemini 3.8 Live更新推出“实时虚拟形象”功能,支持口型同步与实时面部表情展示,目前已面向企业客户开放使用
A playful experiment demonstrating how generative AI transforms static data tools into interactive assistants.
In this follow-up to the famous SQL Zoo project, AI agents like Ava and Finn embody Snowflake users, turning cold code into a vivid demonstration of modern data workflows.
Nvidia CEO's surgery analogy draws criticism as data center pollution concerns mount.
Jensen Huang compares AI's environmental cost to painful surgery, arguing fossil fuels are necessary now for a cleaner future.
The move allows payment delays if the New Mexico facility misses its 2028 deadline.
Oracle has issued a force majeure notice to Stargate partners, citing regulatory delays to gas pipelines and fuel cell permits that could push the data center's 2028 launch.
Disrupting the cloud dependency with on-device AI
PrismML unveiled a 1-bit Bonsai model for Snapdragon AR1 glasses, enabling real-time vision-language tasks locally and shifting away from proprietary cloud compute.
Conference features OpenAI, Anthropic, and Replit ahead of October launch
TechCrunch Disrupt 2026 kicks off Oct 13 with OpenAI, Anthropic, and Replit headlining six industry stages. A limited BOGO discount ends soon.
Turning barcodes into consumer insights via AI
bananabread AI now offers a Product API that ingests items and returns categorized consumer love and desire metrics from online reviews.
A new notebook tool combines blockchain verification with MCP integrations for secure, append-only logging.
Fresh Jots lets scripts and AI append tamper-evident notes anchored to the Bitcoin blockchain, offering private, encrypted storage with free MCP access for Claude and Cursor.
A new open-source tool lets AI agents and humans co-create animated technical diagrams locally, without accounts or API keys.
DrawCMS is a local-first editor where AI agents drive animated technical diagrams through direct tool calls, blending structure, motion, and narration into portable files.
Prompt injection flaws expose internal code and 'dream' routines.
Independent developers coaxed Meta’s Muse AI into revealing its entire root filesystem, exposing hard-coded capabilities and internal documentation in a major prompt injection vulnerability.
Meta announced Horizon Create and Studio, allowing users to generate full 2D and 3D mobile games via AI prompts. Top titles will be playable directly within Instagram and Facebook feeds.
Google's AI-powered virtual closet, inspired by Clueless, is now available on Android and iOS in the US, Brazil, and India, alongside new editing tools.
Meta's new Muse Charm taps into the Gen Z bag charm trend, merging retro tech aesthetics with customizable AI. But privacy concerns remain a major hurdle.
Researchers demonstrate a method to suppress AI refusals at runtime without altering model weights.
New 'Dynamic Abliteration' uses multi-layer steering to bypass AI safety refusals in real-time, keeping base model weights frozen and intact.
Why isolating rogue AI agents on disconnected networks may undermine safety testing.
Air-gapping AI makes testing safer but strips away the internet access needed for realistic evaluations, leaving researchers with a dangerous trade-off.
AI app builder doubles valuation in eight months amid enterprise surge.
Lovable's annualized revenue hit $600M at HumanX, driven by Fortune 500 adoption and 'vibe coding' that outputs full products instead of raw code.