模型发布/更新
-
First unified scientific large model LOGOS officially open-sourced
LOGOS, open-sourced by ATH-Token Foundry in collaboration with Renmin University of China's Gaoling School of…
-
Volcano Engine launches Doubao Realtime Voice Model 3.0 API service, opens invite-only testing
Volcano Engine has launched the Doubao Realtime Voice Model 3.0 (Seeduplex) API service and opened invite-only…
-
Grok TTS tops blind test with a humanness score of 96
xAI's Grok TTS model ranked first in @Vapi_AI's Humanness Index blind test with 96 points (humans score 100). The…
产品发布/更新
-
GPT-5.5 Instant improves ChatGPT health intelligence
More than 230 million users get health information through ChatGPT every week. GPT-5.5 Instant shows significant gains…
-
OpenAI 联合多国医生:GPT-5.5 Instant 健康问答能力追平前沿 Thinking 模型
OpenAI 与全球 60 个国家、49 种语言、26 个专科的数百名医生合作,通过医生主导的评估大幅提升了 GPT-5.5 Instant 在健康相关问题的智能水平,现已能与公司前沿 Thinking 模型(推理模型)相当。该模型每周为…
-
AI employee Viktor lands on Microsoft Teams with $20M annualized revenue
AI employee Viktor has reached $20 million in annualized revenue on Slack with no sales team and no large-scale…
-
Claude Code now supports artifacts
Starting today, Claude Code can turn work in progress into artifacts — live, shareable interactive web pages covering…
-
Adobe adds AI agents to Photoshop, Premiere and other Creative Cloud apps
Adobe is expanding its "creative agents" to Photoshop, Premiere and other apps, offering AI Assistant in public beta.…
-
DeepSeek image recognition mode officially launches on App and web
DeepSeek's image recognition mode officially launched on the web and App on June 18, alongside Fast Mode and Expert…
-
A deep dive into the technical internals of Midjourney Scanner
Our technical deep dive into the all-new "Midjourney Scanner."
行业动态
-
Eight departments: leverage consumer loan interest subsidy policy to support purchases of AI-related products
The Ministry of Commerce and seven other departments on June 18 issued implementation opinions on accelerating the…
-
China's first mandatory national standard for L3/L4 autonomous driving published for comment, effective July 2027
On June 16, the Ministry of Industry and Information Technology opened public comment on two mandatory national…
-
Bernie Sanders proposes $7 trillion AI plan: 50% stock tax on large AI companies
Bernie Sanders proposed legislation imposing a 50% stock tax on companies with annual AI sales exceeding $200 million,…
-
OpenAI IPO前连下两城:招揽Transformer共同作者及前白宫AI政策官员
OpenAI在IPO前夕连招两位重量级人物:Google DeepMind AI先驱、Transformer架构共同作者Noam Shazeer,以及前特朗普白宫AI政策官员Dean Ball。Shazeer此前通过27亿美元收购协议重返…
-
AI data centers get government-mandated fast track to the grid
The US Federal Energy Regulatory Commission (FERC) ordered six major grid operators to provide a fast-track…
-
Grok 现集成 Databricks Agent Bricks
Grok 模型现已原生集成到 Databricks Agent Bricks 平台。在 2026 年 Data + AI Summit 上,Databricks 宣布与 SpaceXAI 合作,使 Grok 与其他前沿及开源模型在同一受控…
论文研究
-
Anthropic Project Fetch phase two: Claude Opus 4.7 completes tasks autonomously, about 20x faster than human teams
Anthropic released results from phase two of the Project Fetch experiment. In the original August 2024 experiment,…
-
MosaicLeaks: can your research agent keep a secret?
Deep research agents face privacy leakage risks when combining private local documents with external web retrieval.…
-
OpenAI and Harvard study: o3 Deep Research model aids diagnosis of rare pediatric diseases, yielding 4.8% additional diagnoses
Boston Children's Hospital, Harvard University and OpenAI published a study in NEJM AI. The team used the OpenAI o3…
-
OpenAI reinforcement learning yields broadly and persistently beneficial models
OpenAI trains models via reinforcement learning in real conversation scenarios to exhibit beneficial traits such as…
技巧与观点
-
Are AI agents up to the task? Evaluating open-source models on our own tools
Hugging Face released a benchmark framework for AI agent use cases, evaluating the agent-friendliness of the…
-
Navigating Claude Code: CLAUDE.md, skills, hooks, rules, subagents and more
Claude Code offers seven ways to customize instructions: CLAUDE.md (always loaded at the root, loaded on demand in…
-
Beyond LoRA: how to choose the best parameter-efficient fine-tuning technique?
Among parameter-efficient fine-tuning (PEFT) techniques, LoRA holds absolute dominance: of 20,834 model cards on…
-
Cloudflare 发布多阶段漏洞发现工具,详解对抗性审查与上下文绕过技术
Cloudflare 分享了其多阶段漏洞发现工具的技术架构,包含自动化分类循环。该系统通过管理状态控制、引入对抗性审查来压制误报,并围绕 LLM 上下文窗口限制设计路由策略。
-
Google 庆祝A2A协议发布一周年:协作智能体生态
Google 庆祝Agent-to-Agent(A2A)协议发布一周年。A2A专为生成式AI设计,相比传统REST API提供安全边界、零上下文污染、动态自主性和工作负载分布四大架构优势。应用实例FoldRun是一个独立的智能体接口,可在…
-
埃森哲:昔日与今朝,以及它如何预示未来
埃森哲去年九月高调宣称AI将改变其业务,但本季度财报令人失望,股价下跌约18%,本周跌幅近23%,较52周高点已跌超50%。生成式AI并未带来预期的大幅收益,MIT、麦肯锡、贝恩等多份研究均显示类似结论。Claude Code(特殊神经符…