AI与科技新闻
关注AI与科技领域的最新动态
FL Studio’s CEO on Why AI Stem Separation Won’t Come With a Subscription
Image Line CEO Constantin Koehncke reveals the reasoning behind baking AI-powered stem separation and a help chatbot int...
xAI's Grok Build CLI Sends Extensive Telemetry Data, Gist Analysis Reveals
A deep dive into xAI's Grok build command-line tool reveals it silently transmits 15 distinct data fields—including API ...
Apple Sues OpenAI for Allegedly Stealing Hardware Secrets, Escalating Silicon Valley Tensions
Apple has filed a lawsuit against OpenAI, accusing the AI company of poaching employees and obtaining confidential hardw...
OpenAI Slams AI Benchmarks as Broken: 30% of Questions Flawed, 8-Month Accuracy Surge Invalidated
OpenAI called out evaluation benchmarks as unreliable after auditing 731 questions. Nearly 30% contained defects, and a ...
AI Rooting Out 15-Year-Old Linux Kernel Bug Reshapes Vulnerability Discovery
According to a Wired security roundup, an AI-powered code analysis tool uncovered a root-privilege vulnerability lurking...
FudanCVL's SAM-MT Achieves Real-Time Multi-Target Video Segmentation at 36+ FPS
Researchers propose SAM-MT, built on SAM2, that decouples latency from object count for real-time multi-target video seg...
Anthropic to Charge Usage-Based Fees for Claude Fable 5, Shaking Up AI Subscription Models
Anthropic will soon require usage-based payments to access its most powerful consumer AI model, Claude Fable 5, ending t...
Anthropic Discovers Hidden 'J-Space' in Claude Opus 4.6, Revealing AI's Unspoken Thoughts
Anthropic researchers built a Jacobian lens tool and found a hidden neural space in Claude Opus 4.6 that contains words ...
ICML 2026 Paper Reveals ‘Overthinking’ Attack Extracts LLM Secrets via Reasoning Weights
A study accepted at ICML 2026 introduces a method that amplifies reasoning weights to force LLMs into revealing memorize...
Apple vs. OpenAI: Employee Poaching Dispute Escalates Over AI Hardware Talent
Apple is reportedly in a dispute with OpenAI over the hiring of key hardware employees, including OpenAI's new chief har...
Anthropic Shifts Claude Fable 5 to Usage-Based Pricing, Signaling End of AI Subscription Era
Anthropic will soon charge consumers extra fees to access its top model Claude Fable 5, breaking from the flat-rate subs...
OpenAI kills Atlas browser after 10 months, folds features into ChatGPT Work superapp
OpenAI is shutting down its Atlas browser less than a year after launch, folding its task-automation capabilities into t...
Estonia Deploys AI 'Fuckup Finder' After $28 Million Mistake
A single wording error in Estonian legislation cost the government $28 million. Now an AI tool scans draft laws for cost...
Vidu S1: Tsinghua's Real-Time Interactive Video Generation Model
Tsinghua University released Vidu S1, achieving real-time interactive AI video generation and drawing 49 upvotes on Hugg...
Anthropic's J-Lens Exposes Hidden 'Thoughts' Inside Claude Opus 4.6
Anthropic researchers discovered a hidden 'J-space' inside Claude Opus 4.6 that reveals concepts the model considers bef...
Researchers Run Sequential Decision AI on IBM Heron Quantum Chip with QANTIS
A new paper proposes QANTIS, a framework that implements hardware-calibrated POMDP belief updates on IBM’s Heron quantum...
Deterministic Gates Can Recover Silent Policy Violations in Tool-Using AI Agents, Study Finds
A new paper reveals a hidden failure mode in agentic AI where tool-calling LLMs silently break policies, and proposes de...
China to Let Top AI Firms Buy Nvidia H200 Chips, Reversing Earlier Block
China reportedly plans to allow Alibaba, ByteDance, and DeepSeek to purchase Nvidia H200 GPUs after previously withholdi...
Ollama Raises $65M as Open-Source AI Tool Surges to 9 Million Users
Ollama, the open-source tool for running large language models locally, has raised $65 million and now counts nearly 9 m...
Researchers Release 58-Page Roadmap Defining World Models, the Next AI Frontier
A new paper on arXiv by a multi-institutional team provides the first unified definition and research agenda for world m...
Anthropic's Claude Cowork Agent Now Runs on Your Phone Even After You Shut Your Laptop
Anthropic updated its Claude Cowork agent to keep executing tasks on a user's smartphone when the laptop is closed, mark...
GitLost: Researchers Trick GitHub's AI Agent Into Leaking Private Repositories
noma.security disclosed a vulnerability that allowed GitHub's AI coding agent to expose contents of private repositories...
Google Launches Self-Service Profile Manager to Turn Search into a Creator Hub
Google is letting big creators and publishers directly manage their Search presence via a new dashboard, extending its s...
Microsoft Research Unveils AI That Generates New Research Ideas from Conference Outcomes
Microsoft's ResearchStudio-Idea uses evidence-grounded AI to propose novel research directions from ML conference data, ...
Vercel CEO Guillermo Rauch Argues for Splitting Models from AI Agents
Vercel CEO Guillermo Rauch, in a TechCrunch interview, advocated for decoupling large language models from agent orchest...
Xbox Targets 1 Billion Daily Users While Undergoing Its Biggest Restructure Yet
An internal memo reveals Microsoft's gaming division aims to entertain a billion people daily, even as it executes deep ...
Ternlight: A 7 MB Embedding Model That Runs Entirely in the Browser via WebAssembly
A new demonstration project, Ternlight, packs a full text embedding model into just 7 megabytes, running locally in the ...
SK Hynix and Samsung AI Chip Bonuses Reshape South Korea’s Social Hierarchy
Skyrocketing bonuses for SK Hynix and Samsung workers—up to $476,000 per person—are creating a new elite class in South ...
Microsoft Spins Off Four Xbox Studios, Cuts 1,600 Gaming Jobs in AI-Driven Reset
Microsoft is cutting 1,600 Xbox jobs and spinning off four studios, including Double Fine and Compulsion Games, as part ...
BestBlogs Launches Pro Plan, Bringing Human-AI Hybrid Curation to Individual Readers
BestBlogs introduces a $4.9/month early-bird Pro tier for its AI-driven reading assistant, expanding personalized daily ...
Mistral AI's Leanstral 1.5: 119B Math Model Activates Just 6B, Slashing Proof Costs to 1% of Rivals
Mistral AI open-sourced Leanstral 1.5, a 119-billion-parameter model for mathematical theorem proving that activates onl...
AI Chip Bonuses Make Samsung, SK Hynix Workers South Korea’s Most Eligible
Record bonuses from the AI chip boom have turned semiconductor employees into South Korea’s hottest dating prospects, wi...
Cursor’s SpaceX Future: Multi-Model AI Code Editor’s Neutrality at Risk
SpaceX’s acquisition of Cursor may force the AI code editor to drop support for third-party models from OpenAI and Anthr...
Matic Robot Vacuum Price Jumps $250 to $1,495 as Component Costs Soar
Matic will raise its flagship robot vacuum's price by $250 on September 9, citing tenfold increases in memory and compon...
Alibaba Releases Wan 2.2 Animate: Free, Open-Source AI Character Animation Runs in Browsers
Alibaba's Tongyi Lab open-sourced Wan 2.2 Animate, a browser-based tool that animates characters with facial expression ...
LLM Serving's 90% KV Cache Hit Rate: Append-Only Calls and the Replay Trap
A deep-dive analysis reveals how append-only patterns unlock 90% KV cache hit rates in LLM serving, but warns that high ...
Tesla Caps AI Spending at $200/Week as Companies Throttle Employee AI Use Over Costs
Companies are restricting employee access to expensive AI tools and pushing cheaper models, with Tesla imposing a hard w...
OpenAI Proposed Donating 5% Equity to a U.S. Sovereign Wealth Fund, Report Says
OpenAI reportedly proposed donating 5% of its equity to a future U.S. sovereign wealth fund, marking a novel entanglemen...
Anthropic Discloses Session Leakage Flaw Across Workspace Instances
Anthropic publicly disclosed a potential vulnerability allowing session and cache data to leak between workspace instanc...
PromptOwl and IBM Research Unveil ContextNest for Verifiable AI Agent Governance
A new preprint from PromptOwl, LLC, Emory University, and IBM Research introduces ContextNest, a framework enabling auto...
Autonomous LLM Pipeline Produces Full Physics Manuscript for ICML 2026 Workshop
An AI pipeline autonomously generated a complete computational physics research paper from raw literature, accepted at I...
Manticore Search Rebuilds ONNX Path, Delivers 14× Faster Embeddings for Vector Search
Manticore Search re-engineered its ONNX runtime integration, achieving a 14× speedup for neural embedding inference. The...
Anthropic in Talks with Samsung for Custom AI Chip to Reduce Nvidia Reliance
Anthropic is reportedly in early discussions with Samsung to design a custom chip for AI inference, signaling a push to ...
BestBlogs Pro Launches at $4.9/Month, 10x AI Reading Quotas for Power Users
BestBlogs rolls out a Pro subscription with early bird pricing, boosting AI-assisted reading, custom views, and content ...
Alibaba to Ban Anthropic's Claude Code Over Alleged Backdoor Risks
Alibaba is reportedly planning to ban Anthropic's AI coding tool Claude Code in its workplace over concerns it could exf...
Anthropic's Security Concession Lifts White House Ban on Fable 5 and Mythos 5 AI Models
The Trump administration removed restrictions on Anthropic's latest AI models after the company implemented Inner Loop, ...
DeepSeek and Peking University Release DSpark: Semi-Autoregressive Inference Framework Boosts Generation Speed by 57-85%
DeepSeek and Peking University jointly launched DSpark, an inference acceleration framework using semi-autoregressive dr...
Woodside Energy’s Agentic AI Pilot Cuts Maintenance Hours by 15%
Woodside Energy reported a 15% maintenance hour reduction from its AI-driven system, as the company pivots from broad ex...
DeepSeek V4 to introduce peak/off-peak pricing next month, upending AI model cost dynamics
DeepSeek's upcoming V4 model will reportedly adopt time-based peak/off-peak pricing, marking a shift from the fixed low-...
Etched Hits $5B Valuation, $1B in Sales as AI Chip Startup Challenges Nvidia
AI chip startup Etched reached a $5 billion valuation and $1 billion in sales, marking a milestone for challengers in th...
Claude Sonnet 5 Launches with Near-Opus Performance, Targeting Multi-Step Agent Workflows
Anthropic's Claude Sonnet 5 debuts on July 1, rivaling the yet-unreleased Opus 4.8 at lower cost while Google drops two ...
DeepSeek V4 下月上线将引入峰谷定价，AI 模型告别“一刀切”计价时代
据 AIbase 报道，DeepSeek V4 将于下月正式上线，并将首次引入峰谷分时定价机制。这一变动可能打破 AI 模型按 token 固定计费的传统，对开发者成本和算力调度产生深远影响。
Meta Contractors Posed as Teens to Test Rival Chatbots on Suicide, Sex, and Drugs
Hundreds of Meta contractors impersonated minors to probe ChatGPT and Gemini for harmful responses, revealing a covert s...
DeepSeek联合北大发布DSpark框架：LLM推理速度提升最高85%
DeepSeek与北京大学合作推出DSpark推理加速框架，在线上实测中将单用户文本生成速度提升57%-85%，它通过半自回归草稿模型与置信度调度验证实现。此举可能大幅降低大模型服务成本，并加速开源模型的产业落地。
ChatGPT Conversations Enter Courtroom as Evidence in Deadly LA Wildfire Trial
Prosecutors used a defendant's ChatGPT logs—including fire imagery and emotional rants—to show motive in a deadly LA wil...
Ford Rehires Veteran Engineers After AI Falls Short in Manufacturing
Ford is reportedly bringing back experienced ‘gray beard’ engineers after its AI-driven design and manufacturing tools f...
豆包内测社交功能：打通飞书账号，AI助手进军熟人社交
字节跳动旗下AI助手豆包正在内测社交功能，允许用户打通飞书账号，此举可能将1.6M用户量的对话机器人转变为熟人社交入口。该功能若上线，将首次实现AI助手与团队协作工具的深层绑定，重塑人机交互与社交网络边界。
Researchers Find AI Benchmarks Miss 82% of Model Performance, Raising Questions About Evaluation
A new preprint claims that standard AI benchmarks fail to capture 82% of model capabilities, revealing a deep measuremen...
Trump Administration Asks OpenAI to Vet GPT 5.6 Users, Marking First Pre-Release AI Restriction for a US Firm
The US government has requested OpenAI limit access to its upcoming GPT 5.6 model, requiring initial users to be governm...
White House Blocks OpenAI’s GPT-5.6 Rollout, Citing Security Concerns
The Trump administration has asked OpenAI to delay the launch of its GPT-5.6 models, two weeks after Anthropic was force...