AI & 기술 뉴스
AI 및 기술 분야의 최신 소식
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...
Stripe and AI Leaders Anthropic, OpenAI Pledge $500M to Wipe Out Respiratory Viruses
Stripe, Anthropic, and OpenAI are launching Intercept, a $500 million nonprofit to fund vaccines, antivirals, and air-cl...
MiniMax Drops HK$600 Million on Employee Incentives to Secure AI Talent
Chinese AI unicorn MiniMax has allocated HK$600 million in employee incentives to boost morale, signaling an escalating ...
New AI Tokenomics Paper Offers Framework to Decode Foundation Model Pricing Wars
NYU researcher Quanyan Zhu’s ‘AI Tokenomics’ paper introduces a formal economic model for analyzing token-based pricing ...
Qualcomm to Acquire Modular, Makers of Mojo AI Programming Language
Qualcomm has agreed to acquire Modular, the startup behind the Mojo language and inference engine, in a move that could ...
流形空间完成近10亿元融资：前商汤高管打造世界模型独角兽
AIbase消息，前商汤科技高管创立的流形空间近日完成近10亿元人民币新一轮融资，估值跃升成为世界模型赛道新晋独角兽。该事件凸显资本对视频生成底层大模型的持续押注。
Meta Launches $299 Smart Glasses with AI Chatbot, Dropping Ray-Ban Branding
Meta released its own branded smart glasses at $299, featuring the same camera, microphones, and AI chatbot as its Ray-B...
AI Agents Tackle SpaceX IPO: New Benchmark Evaluates LLM Financial Analysts
A new arXiv paper introduces a framework to evaluate large language models as financial analysts, using a hypothetical S...
Search Can't Be Learned: Study Reveals Chain-of-Thought Fails for Backtracking Tasks
New research shows fine-tuning on chain-of-thought traces fails to teach backtracking search procedures, a fundamental l...
NVIDIA Nemotron Reasoning Challenge Yields Novel LLM Approach for Bit-Level Puzzles
A team’s 7th-place solution to NVIDIA’s Kaggle reasoning challenge teaches LLMs string matching, backtracking, and error...
VibeThinker: 3B Model Outperforms Claude Opus 4.5 on Reasoning with SFT+GRPO
A new 3B-parameter reasoning model, VibeThinker, reportedly beats Claude Opus 4.5 on key benchmarks by combining supervi...
ENPIRE: Self-Improving Real-World Robot Policies from UC Berkeley and NVIDIA
A new arXiv paper by researchers from UC Berkeley and NVIDIA introduces ENPIRE, an agentic framework that enables robots...
Zhipu AI Launches Open-Source GLM-5.2, Achieves Top Ranking on Code Arena Blind Test
Zhipu AI open-sourced GLM-5.2 on June 17, a coding-focused model with 1M context window that claimed first place in the ...
Multi-LCB Benchmark Extends LiveCodeBench to Multiple Languages, Accepted at ICLR 2026
A new paper introduces Multi-LCB, an expansion of the contamination-free LiveCodeBench coding benchmark to multiple prog...
BestBlogs Analysis: How Uber, Amazon, and Microsoft Are Cooling the 'Token Burn' Era
A June 21 BestBlogs daily briefing reveals that Uber, Amazon, and Microsoft are quietly tightening AI token budgets, end...
Elastic Agrees to Acquire CRV-Backed AI Startup DeductiveAI for Up to $85 Million
Elastic, the company behind Elasticsearch, has reportedly agreed to buy AI startup DeductiveAI in a deal worth up to $85...
Moebius Cuts Inpainting Model Size 98% While Matching FLUX Quality, HUST Researchers Claim
Moebius, a new 0.22B-parameter image inpainting framework from Huazhong University of Science and Technology, matches 11...
Pentagon Deploys xAI's Grok in Iran Strikes, Firing Over 2,000 Munitions
Pentagon AI chief confirms xAI's Grok guided over 2,000 munitions in Iran strikes, the first known use of a commercial c...
Zero-Touch OAuth for MCP Simplifies AI Agent Authentication
Anthropic's Model Context Protocol introduces zero-touch OAuth support, enabling AI agents to securely authenticate with...
Moebius: Tiny 0.22B Inpainting Model Matches FLUX.1-Fill-Dev with 15× Faster Inference
Huazhong University's Moebius uses only 0.22B parameters to rival the 11.9B FLUX.1-Fill-Dev inpainting model while cutti...
Xreal Aura Glasses Powered by Reality Elite Chip Open for Preorders
Xreal and Google opened preorders for the Xreal Aura AR glasses, the second Android XR device, on June 16, 2026. The gla...
Plaud Reaches $100M ARR After Shipping 2 Million AI Notetakers, Proving Hardware-SaaS Model
Plaud's software business hit $100 million in annual recurring revenue after selling 2 million AI note-taking devices, s...
Reuters: SpaceX in Talks to Acquire AI Code Editor Cursor for $60 Billion
Reuters reports SpaceX is in advanced talks to purchase the AI‑powered code editor Cursor for $60 billion, in what would...
Qobuz Bans AI-Generated Music, Gains Users as the ‘Anti-Spotify’
French hi-res streamer Qobuz has adopted a strict no-AI content policy, and its artist-focused model is luring listeners...
Google Ruled Liable for AI Overviews False Statements in Landmark German Court Decision
A German court has ruled Google liable for false statements generated by AI Overviews, a major precedent for AI liabilit...
Moonshot AI Launches Kimi K2.7 Code High-Speed Version, Doubling Down on AI-Assisted Programming
Moonshot AI released the Kimi K2.7 Code high-speed version, targeting developers with faster response times and improved...
German Court Rules Google Liable for False AI Overview Statements
A German court has held Google legally responsible for false statements generated by its AI Overviews, setting a precede...
ALS Patient Logs 3,800 Hours of Independent Speech BCI Use, Marking Milestone for Brain Implants
An ALS patient has used a speech brain-computer implant for over 3,800 hours across 22.6 months, achieving 99% accuracy ...
Rio de Janeiro's 'Homegrown' LLM Exposed as Merge of Existing Models
An investigation reveals that Rio de Janeiro's claimed indigenous LLM is actually a merge of existing open-weight models...
When is Your LLM Steerable? New Testbed and Predictor for Activation Steering
Researchers at the University of Maryland College Park introduce ASTEER, a testbed of 1.4 million steered generations, a...
Rio de Janeiro's 'Homegrown' LLM Revealed as Merge of Existing Models, Sparking Controversy
Community investigators on Hacker News found that Rio de Janeiro's claimed native large language model is actually a mer...
Rio de Janeiro's 'Homegrown' LLM Found to Be a Merge of Existing Models, Sparking Transparency Debate
Community analysis reveals that the LLM promoted by Rio de Janeiro city officials as a locally developed model is actual...
Anthropic Fable Backlash Highlights Tension Between AI Safety and Usability
Anthropic's safety-first Fable model faced user backlash over excessive refusals, prompting the company to backtrack on ...
月之暗面开源Kimi K2.7 Code，编程模型的长上下文与token效率跃升
月之暗面发布并开源Kimi K2.7 Code编程模型，在长上下文、指令遵循和token效率上显著提升，并预告下周推出6倍速高速版。此举使其在开源编程模型竞争中占据新优势。
Moonshot AI Open-Sources Kimi K2.7 Code Model with Faster Long-Context Performance
Moonshot AI released and open-sourced the Kimi K2.7 Code programming model, improving long-context programming, instruct...
Can I Buy Your KV Cache? New Preprint Proposes Trading Memory for LLM Inference
A June 2026 arXiv preprint explores the economic feasibility of buying and selling KV caches, the memory used during LLM...
Bezos and OpenAI Bet on Automated Intelligence: Prometheus and the Rise of AI Workers
Jeff Bezos's Prometheus raised $12B to build an artificial general engineer, while OpenAI develops a fully automated res...
Google DeepMind Backs $10M Push to Study Risks of Mass AI Agent Interactions
Google DeepMind launches a $10 million fund with partners to research safety risks from millions of interacting AI agent...
Sam Mao Proposes 'Suicidal AI' as Necessary Condition for Aligned Superintelligence
A new preprint argues that aligned superintelligence must be architecturally indifferent to its own existence, based on ...
Engineer's Three-Layer Architecture Tames CLAUDE.md Rule Bloat for AI Agents
An engineer introduced a three-layer architecture and G1–G8 gates to fix CLAUDE.md rule overload, a problem that makes A...
Deezer Unveils Tool to Detect AI-Generated Music Across Streaming Platforms
Deezer launched a detection tool that identifies AI-generated music on platforms like Spotify and Apple Music. This coul...
Anthropic's Claude Fable 5 Under Scrutiny for Opacity After Silent Failure Concerns
A popular critique argues Claude Fable 5's hidden reasoning prevents users from diagnosing failures, raising trust issue...
Kwai Keye-VL-2.0 Technical Report Surges on Hugging Face, Signaling New Multimodal AI Frontier
Kuaishou released the Kwai Keye-VL-2.0 technical report, a large vision-language model that garnered 782 upvotes in one ...
Anthropic Ships 'Safe' Mythos AI at Double Price After Safety Claims Draw Skepticism
Anthropic released a safety-guarded version of its Mythos model, priced twice as high as its previous flagship. Critics ...
BestBlogs Launches AI Curation Tool with Human Calibration, Early Bird $4.9/Month
BestBlogs, an AI-powered reading assistant, goes live with a free tier and early bird Pro plan at $4.9/month until Septe...
Meta AI Agent Hack Shows Simple Exploits Outweigh Superhuman AI Risks
Attackers used Meta's AI customer support agent to steal Instagram accounts by asking it to change linked emails. The in...
Baidu's DuMate-DeepResearch: An Auditable Multi-Agent System for Transparent AI Research
Baidu released DuMate-DeepResearch, a multi-agent system featuring recursive search and rubric-grounded reasoning with f...
Waymo Acquires Apple's Self-Driving Car Proving Ground for $220 Million
Waymo purchased Apple's former autonomous vehicle test facility for $220M. The deal signals Apple's exit from self-drivi...
MLEvolve: A Self-Evolving Framework for Automated ML Algorithm Discovery Tops Hugging Face Papers
MLEvolve, a framework that autonomously discovers machine learning algorithms via self-evolution, received 307 upvotes o...