[카테고리:] 공개 정보

The Experience Technology Department of Alipay, Ant Group, has officially open-sourced the intelligent programming assistant Neovate Code. It can deeply understand your codebase, follow the existing coding habits, and accurately complete function implementation, bug fixing, and code refactoring based on context awareness. It integrates the core capabilities required by Code Agent.
GitHub：https://github.com/neovateai/neovate-code

At present, Neovate Code is provided in the form of a CLI tool, but its architecture is highly flexible and will support multiple client forms in the future to adapt to more development scenarios.

Its main functions include:
Conversational development - A natural dialogue interface for programming tasks
AGENTS.md rule file - Define custom rules and behaviors for your project
Conversation continuation and resumption - Continue previous work across conversations
Support for popular models and providers - OpenAI, Anthropic, Google, etc.
Slash commands - Quick commands for common operations
Output style - Customize the way code changes are presented
Planning mode - Review the implementation plan before execution
Headless mode - Automate the workflow without interactive prompts
Plugin system - Extend functionality with custom plugins
MCP - Model context protocol for enhanced integration
Git workflow - Intelligent commit message and branch management
…

Dear friends, there's new news about DeepSeek! The latest model, DeepSeek-V3.1-Terminus, has made its debut! 👏

This version comes in two modes: the thinking model and the non-thinking mode, both with a context length of 128k. It is an upgrade based on DeepSeek-V3.1 and has two major improvements. First, in terms of language consistency, it alleviates the mixing of Chinese and English and the occurrence of occasional abnormal characters. For example, the "extreme" character issue mentioned before has also been improved. Second, in terms of Agent capabilities, the performance of Code Agent and Search Agent has been further optimized, making them even more outstanding.
DeepSeek's last update was on August 21st. It's only been a month, and the new model DeepSeek-V3.1-Terminus has outperformed Gemini 2.5 Pro in many evaluations.

However, in terms of benchmark performance, compared to DeepSeek-V3.1, it has only a slight overall upgrade, and there is a slight decline in some benchmarks. But in the Humanity's Last Exam benchmark, the improvement is huge, as high as 36.48%, jumping from 15.9 to 21.7. That's really amazing!

Now, DeepSeek-V3.1-Terminus has been launched on apps, web pages, and APIs.

Here are two addresses for you:
Hugging Face 地址：
https://huggingface.co/deepseek-ai/DeepSeek-V3.1-Terminus

ModelScope 地址：
https://modelscope.cn/models/deepseek-ai/DeepSeek-V3.1-Terminus

By the way, the word "Terminus" means "end". Does this imply that this is the last version of the V3 series and that DeepSeek - V4/R2 is coming soon? It's really exciting!

Dear friends, what do you think of DeepSeek-V3.1-Terminus? Come and share your thoughts in the comments section!

#DeepSeek #DeepSeek-V3.1-Terminus #ModelRelease #ModelUpgrade #PerformanceEvaluation

자기들이, 상하이교통대학교 IPADS 연구실 팀이 대단한 일을 했어! 그들이全新의 모바일 에이전트 툴체인 MobiAgent를 출시했어🎉. 이건 정말 대단한데, 개인화된 지능형 어시스턴트 개발 장벽을 뚫고, 실제 시나리오에서의 성능이 GPT-5와 다른 최고급 클로즈드 소스 모델보다 뛰어난다고도 합니다👍

MobiAgent는 정말 대단합니다. 누구나 자신만의 AI 어시스턴트를 만들 수 있는 기회를 제공합니다. 이 도구 체인은 사용자가 모바일 에이전트를 처음부터 구축할 수 있도록 지원하며, 작업 데이터 수집, 모델 학습, 핸드폰으로의 배포까지 일련의 프로세스를 모두 처리할 수 있습니다. 그리고 오픈 소스입니다. 사용자는 자신만의 데이터를 수집하고 모델을 학습시켜 개인 기기에서 지능형 어시스턴트를 사용할 수 있습니다. 너무 편리하죠🥰

그 성능을 검증하기 위해 연구팀은 국내 인기 있는 20개의 앱에서 테스트를 수행했습니다. 결과는 70억 규모의 MobiAgent 모델이 작업 완료 점수에서 많은 유명한 클로즈드 소스 대형 모델을 능가했으며, 같은 규모의 오픈 소스 GUI 에이전트 중에서도 선두에 있다는 것을 보여줍니다👏. 독특한 "잠재 메모리 가속기"는 과거 작업을 학습하여 에이전트가 반복적인 작업을 빠르게 완료할 수 있도록 돕고, 성능을 2~3배 향상시킵니다.

MobiAgent의 핵심은 효율적인 데이터 수집과 지능형 학습 프로세스에 있습니다. 경량 도구를 사용하여 사용자의 핸드폰 작업을 기록한 다음 범용 VLM 모델을 사용하여 고품질 학습 데이터를 생성합니다. 정련된 조정을 통해 학습된 에이전트는 뛰어난 일반화 능력을 갖게 됩니다. 그 "뇌"는 세 부분으로 나뉩니다. "기획자"는 작업 계획을 책임지고, "결정자"는 화면을 기반으로 결정을 내리고, "실행자"는具體적인 작업을 수행합니다. 이러한 구조로 모델 학습이 더욱 효율적이 되고 응답 속도도 크게 향상됩니다😎

또한 혁신적인 AgentRR 가속 프레임워크가 있어 과거 작업 경험을 활용해 반복 작업 실행 효율을 크게 향상시킬 수 있으며, 동작 재사용률은 최대 60%~85%에 달할 수 있습니다. 지능형 어시스턴트는 일상적인 업무를 빠르고 정확하게 처리합니다.

MobiAgent의 출현은 개인용 지능형 어시스턴트 맞춤화를 편리하게 할 뿐만 아니라 모바일 에이전트 생태계 발전을 촉진합니다. "말로만 하면 손을 움직이지 않아도 된다"는 지능 시대가 정말 다가올 것 같습니다🤩

자기들이, MobiAgent가 기대되나요? 댓글 창에서 이야기 나눠보세요🧐

论文地址：https://arxiv.org/pdf/2509.00531

#MobiAgent #상하이교통대학교 #AI 어시스턴트 #모바일 에이전트 #오픈 소스 도구 체인 #성능 뛰어넘기기

자기들이, OpenAI가 또 대단한 행보를 보였어! 오늘 ChatGPT의 프로젝트 기능이 무료 사용자에게 공식적으로 개방된다고 발표했는데, 정말 대박이에요👏

이번 업데이트는 다양한 사용자 그룹에게 기능 개선을 가져왔어. 먼저 대용량 파일 업로드 제한에 대해 말하면, 무료 사용자는 하루에 최대 5개의 파일을 업로드할 수 있고, Plus 사용자는 25개로 늘어나며, Pro, 비즈니스 및 엔터프라이즈 버전 사용자는 40개의 파일을 업로드할 수 있어. 이런 계층화된 설계는 정말 친절해. 당신의 필요가 크든 작든, 자신에게 적합한 사용 방법을 찾을 수 있어요🥰

또한 OpenAI는 많은 개인 맞춤형 설정 기능을 추가했어. 지금 사용자들은 프로젝트의 색상과 아이콘을 사용자 정의할 수 있게 되었고, 관리 인터페이스가 갑자기 매우 개인화되어 업무 효율도 많이 향상될 거예요. 맥락 일관성을 유지해야 하는 분들에게 새로 추가된 프로젝트 전용 메모리 제어 기능은 정말 유용해. 다양한 대화 시나리오에 더 잘 적응할 수 있고 정보 관리도 쉽고 편해지겠죠😎

이 일련의 업데이트는 OpenAI가 우리 사용자들의 니즈에 얼마나 주의를 기울이고 있는지 충분히 보여줘. 기업 사용자也好 개인 사용자也好，이 새로운 기능들로 ChatGPT를 사용하는 경험이 더 원활해질 거예요.

말하지 않을 수 없이, OpenAI의 이번 업데이트는 사용자 경험의 매우 큰 업그레이드야. 플랫폼의 매력이 더욱 강해졌고, 더 많은 사용자들이 AI가 가져다주는 편리함을 평등하게 누릴 수 있게 됐어. 앞으로 ChatGPT는 확실히 계속 최적화될 거고, 더 많은 놀라움을 기대해봅시다🤩

자기들이, ChatGPT의 이 새로운 기능들이 기대되나요? 댓글 창에서 이야기 나눠보세요🧐

#ChatGPT #신기능 출시 #프로젝트 관리 #사용자 경험 #무료 사용자 #개인 맞춤형 설정

Dear friends, here's some big news! At 00:00 on September 1, 2025, the "Measures for the Identification of Artificial Intelligence-Generated and Synthesized Content" jointly formulated by multiple government departments officially took effect! 🎉 This measure puts forward regulatory requirements such as the mandatory addition of explicit and implicit identifications. From now on, AI-generated text, images, audio, and video must all show their "digital ID cards"🧐

Before this, many platforms such as Tencent, Douyin, Kuaishou, and Bilibili had already introduced detailed rules. Take Douyin for example, it has launched an AI content identification function and an AI content metadata identification reading and writing function, which help creators add prompt identifications and also provide technical support for content traceability👏

Now the ecological chain of AI-generated content has entered a stage of standardized management. Artificial intelligence is developing extremely rapidly. In 2024, the scale of China's artificial intelligence industry exceeded 700 billion yuan and has maintained a high growth rate year after year. However, the popularization of technology has also brought new risks. For example, there are more and more cases of it being used to create false news and carry out online fraud.

The core of the policy of the "Measures for the Identification" is the requirement of dual identifications. Explicit identifications should be "visible at a glance" to ordinary users. For example, add text explanations at the beginning and end of an article, or add voice prompts or special icons in audio and video. Implicit identifications, on the other hand, are to embed "hidden information" in the file metadata, including various key information.

This measure is of great significance. Professor Ren Kui, one of the drafters, said that it is the first time to include generation service providers, content dissemination platforms, and end users in a unified governance framework, forming a system progression with other regulations and clarifying the boundaries of responsibility. It can promote the standardized development of the AIGC industry, reshape the public's trust in AIGC technology, and also enhance China's voice in the field of artificial intelligence security governance, providing a model for global content governance👍

Let's talk about the dual identification system again. Explicit identifications should be directly perceived by users. Texts should mark words such as "generated by artificial intelligence" in specific positions, and the font should be clear. Implicit identifications focus on technical traceability, embedding metadata inside the file, containing various key information. There are clear labeling requirements for different types of AI-generated content.

The "Measures for the Identification" also encourages the use of AI for original content creation. Moreover, it clarifies the obligations of different entities at the legal level. Service providers need to ensure that the content meets the identification requirements. Dissemination platforms need to verify implicit identifications and add significant prompt identifications. Application distribution platforms need to verify the identification functions of service providers.

However, the implementation of this measure also faces challenges. Users may delete explicit identifications or avoid implicit ones through transcoding, making it difficult to accurately identify the content posted by malicious users. Lawyers suggest that content publishing platforms should assume more responsibilities. Professor Ren Kui suggests from a technical perspective the development of secure content implicit identification technology.

All in all, identification is a crucial step in the governance of AI-generated content. But to truly avoid risks, it is also necessary to refine laws and regulations, establish industry self-discipline standards, strengthen law enforcement efforts, and enhance international cooperation. Cross-border AIGC law enforcement is also a challenge. In the future, it is necessary to promote the coordination of technical identifications and establish cross-border law enforcement mutual assistance mechanisms. Dear friends, what do you think about the mandatory "labeling" of AI-generated content? 🤔

#AI-generated content #Mandatory labeling #Content security governance #Dual identification system #Main body responsibility #Supervision challenges

On the evening of August 19th, DeepSeek officially announced that the online model version has been upgraded to V3.1. The most significant improvement is that the context length has been extended to 128K, which is equivalent to being able to process super-long texts of 100,000 to 130,000 Chinese characters, suitable for long document analysis, code library understanding and multi-round dialogue scenarios.

Users can now experience the new version through the official website, App or WeChat mini-program. The API interface call method remains unchanged, and developers can switch seamlessly without additional adjustments.

This upgrade is not a major version iteration, but an optimization of the V3 model. Tests show that V3.1 has a 43% improvement in multi-step reasoning tasks compared to the previous generation, especially more accurate in complex tasks such as mathematical calculations, code generation and scientific analysis. Meanwhile, the situation of the model's "hallucination" (generating false information) has decreased by 38%, and the output reliability has been significantly enhanced. In addition, V3.1 has also optimized multilingual support, especially improving the processing ability of Asian languages and less common languages.

Although V3.1 brings important improvements, the release time of the next-generation large model DeepSeek - R2, which users are more looking forward to, is still uncertain. Previously, there was market speculation that R2 would be released from August 15th to 30th, but insiders close to DeepSeek said that this news is not true and the official has no specific release plan at present.

DeepSeek's update rhythm indicates that the V4 model may be launched before the release of R2. However, the official has always been low-key, emphasizing that "it will be released when it's done" and has not responded to any market speculation.

Experience address:https://chat.deepseek.com/

Recently, the news of the release of DeepSeek's next-generation large model DeepSeek - R2 has attracted widespread attention in the market. There is a rumor that DeepSeek - R2 will be released between August 15th and 30th. However, according to Tencent Technology, sources close to DeepSeek have confirmed to the media that this news is not true and DeepSeek - R2 has no release plan this month.

As early as the beginning of this year, news about the R2 model had already started to spread. At that time, it was predicted that the R2 model would be released on March 17th, but this claim was also denied by the official. So far, DeepSeek has not officially announced the specific release time and technical details of the R2 model, which has disappointed many observers.

According to reports, the DeepSeek team stepped up the development of the R2 model in June this year. Insiders revealed that CEO Liang Wenfeng is still not satisfied with the capabilities of the model, and the team is still improving its performance and is not ready for official use. Early news said that DeepSeek originally planned to launch the R2 model in May, but due to various reasons, the plan was delayed. The new model is expected to be able to generate higher quality code and have the ability to reason in non-English languages.

On August 7, 2025, OpenAI officially released the GPT-5 series of models, which represents the most significant product upgrade in the company's history. This release includes four versions: GPT-5, GPT-5 Mini, GPT-5 Nano, and GPT-5 Pro, each deeply optimized for different application scenarios, marking a new stage of development for AI technology.

Unified Intelligent System: A Revolutionary Breakthrough in Technical Architecture
GPT-5 is positioned by OpenAI as a "unified intelligent system", successfully integrating capabilities that were previously scattered across different models: the multimodal processing of GPT-4o, the deep reasoning of the o series, advanced mathematical calculation, and agent task execution. This architectural innovation eliminates the need for users to manually switch between different models. The system automatically selects the most suitable processing method based on task complexity through a real-time router.

In terms of core technical indicators, GPT-5 has achieved a comprehensive breakthrough:

Mathematical Reasoning: Achieved an accuracy rate of 94.6% in the AIME 2025 benchmark test without the need for external tools.
Code Capability: Scored 74.9% in the SWE-bench Verified test and 88% in the Aider Polyglot multilingual programming test.
Multimodal Understanding: Scored 84.2% in the MMMU benchmark test.
Professional Knowledge: Scored 88.4% in the GPQA general question answering test.
Detailed Analysis of the Four Versions

GPT-5（旗舰版）:最强推理与多模态能力
作为系列中的旗舰产品，GPT-5专为复杂任务设计，具备以下核心特性:

推理能力突破:内置链式推理（Chain-of-Thought）技术，能够分解复杂问题并逐步解决。在内部测试中，GPT-5在40多个职业领域的复杂任务上表现优于前代所有模型。

全面多模态支持:支持文本、图像、语音和视频处理，继承了Sora的视频生成技术。用户可以上传各种格式的内容，GPT-5能够生成相应回应或执行复合任务，例如分析医学影像或实时翻译视频内容。

代理式任务执行:支持自动浏览网页、生成完整软件应用、管理日程等复杂操作。在发布会演示中，GPT-5根据简单描述在数秒内生成了包含闪卡、测验和进度跟踪功能的完整法语学习Web应用。

大幅降低幻觉率:通过”安全补全”技术，GPT-5的事实错误率比GPT-4o降低约45%，在使用推理模式时错误率比o3模型降低约80%。

GPT-5Mini:高性价比的轻量选择

GPT-5Mini针对成本敏感应用进行优化，在保留核心功能的同时显著降低了资源需求:

支持中等复杂度的链式推理任务
具备文本、图像和语音处理能力，视频处理功能相对受限
可在较低算力设备上运行，适合中小企业和个人开发者
在核心推理测试中接近o4-mini性能水平
主要应用场景包括教育内容生成、客户服务自动化、简单多模态任务处理等。

GPT-5Nano:超高效边缘计算模型

GPT-5Nano专为速度和低资源占用优化，是系列中最轻量的版本:

极低延迟响应，专为实时应用设计
可在内存仅16GB的设备上运行，包括MacBook或低端服务器
推理能力相对简化，主要用于快速交互和简单任务
在通用基准测试中与o3-mini性能相当
适用场景包括移动设备应用、嵌入式系统、实时翻译、语音助手等对响应速度要求极高的场景。

GPT-5Pro:面向专业用户的增强版本
GPT-5Pro是专为高端用户和企业设计的高性能版本:

增强推理模式:支持”GPT-5Thinking”功能，可对复杂问题进行更长时间的深度推理，确保极高准确性。

无限制访问:Pro用户享有无限制的GPT-5访问权限，以及GPT-5Pro的独家访问权。

专业多模态能力:在视频处理、复杂图像分析等任务中表现优异，在HealthBench Hard医疗基准测试中得分46.2%。

深度工具整合:无缝集成搜索、Canvas、代码执行等专业工具，提供完整的工作流体验。

定价策略:史上最大规模免费开放
OpenAI采用了前所未有的开放策略，向所有用户群体提供GPT-5访问权限:

免费用户:可使用GPT-5和GPT-5Mini，有使用限额，超出后自动切换至Mini版本

Plus用户（$20/月）:享有更高使用限额，适合个人用户和小型团队

Pro用户（$200/月）:无限制访问GPT-5和GPT-5Pro，并可使用”GPT-5Thinking”模式

企业与教育用户:发布后一周内获得访问权限，并可使用GPT-5Pro版本

API定价:输入$1.25/百万token，输出$10/百万token，面向专业开发者

用户体验的全面升级
GPT-5系列带来了多项用户体验创新:

智能模型选择:系统根据任务复杂度和用户意图自动选择最适合的模型版本，用户无需手动切换

个性化交互:提供四种预设人格（Cynic、Robot、Listener、Nerd）和自定义聊天颜色选项

增强记忆能力:更大的上下文窗口能够记住更长的对话历史，提供更连贯的交互体验

用户友好设计:相比GPT-4o，新模型减少了过度讨好的表达，使用更少不必要的表情符号，让交互更加自然

技术架构创新
GPT-5系列可能采用了混合专家模型（MoE）架构，通过减少活跃参数数量大幅提升效率。训练数据以英语文本为主，聚焦STEM、编程和通用知识领域，知识截止时间为2024年6月。整个训练过程在NVIDIA H100GPU上完成，耗费约210万GPU小时。

竞争优势与市场影响
在当前AI竞争激烈的环境下，GPT-5的发布具有重要战略意义。面对Anthropic Claude3.5Sonnet、xAI Grok4、Google Gemini2.5Pro等强劲竞争对手，OpenAI通过免费开放策略和显著降低幻觉率来巩固市场地位。

据统计，目前已有500万付费用户使用ChatGPT商业产品，包括BNY Mellon、加州州立大学、Figma、Intercom、摩根士丹利等知名机构。GPT-5的发布预计将进一步加速企业AI采用，推动各行业的数字化转型。

行业展望与挑战
GPT-5系列的发布代表了AI技术发展的新里程碑，但同时也面临一些挑战:

隐私与安全:多模态能力涉及处理医疗影像、个人对话等敏感数据，数据保护成为关键议题

技术影响:自动化程度的提升可能对传统工作岗位产生冲击，需要社会层面的适应和调整

性能验证:虽然OpenAI声称GPT-5具备”博士级智能”，但其真实推理能力在实际应用中的表现仍需时间检验

总结
GPT-5系列的发布标志着OpenAI在AI领域的又一次重大突破。通过四个版本的差异化布局，OpenAI成功覆盖了从个人用户到企业客户的全部需求谱系。这不仅是一次技术升级，更是AI产品策略的全面革新。

随着GPT-5成为ChatGPT的新默认模型，取代此前的GPT-4o、o3等版本，用户只需打开ChatGPT输入问题，系统将自动处理并在需要时应用推理功能。这种无缝体验的实现，预示着AI技术正在从工具向助手、从辅助向协作的方向快速演进。

今天凌晨1点（北京时间），OpenAI正式发布备受瞩目的GPT-5，称其为迄今为止最强大、最实用的AI系统。与前代模型相比，GPT-5主要进步如下：在编程、数学、写作、健康问答、视觉感知等场景能力显著提升；大幅减少幻觉；指令遵循能力更强；显著减少谄媚、讨好性回答 。

GPT-5面向所有用户开放，Plus订阅用户有更多使用配额，Pro订阅用户可使用具备更深入推理能力、能提供更全面准确回答的GPT-5 Pro。

GPT-5不再区分传统意义上的推理模型、多模态模型和Agent模型，而是将这些能力集成在统一架构下，实时路由器（Router）会根据对话类型、问题难度、工具调用需求及用户显式指令（如“请认真思考”等）自动判断调用哪个模型。

GPT-5能力亮点：
编程能力：是迄今为止最强的代码模型，擅长前端页面生成、大型代码库调试等复杂任务，可单轮提示生成完整、美观、响应式网站/应用/游戏，对排版等设计原则理解能力增强，更符合开发者需求。

创意写作：能将粗略构思转化为结构完整、富有文学深度、节奏自然的文本，善于处理结构模糊或形式复杂的写作，在日常写作任务中表现出色，更富有人情味。

健康问答：在健康场景理解力大幅提升，是HealthBench测评中表现最优的模型，能根据多种因素提供精确、可信、贴合实际的健康信息，主动识别潜在风险，引导理性判断，适用于辅助决策但不替代医疗专业人员。

安全机制革新：从“拒答”转向“安全生成”，引入“Safe-completion（安全生成）”机制，能更细致处理双用途问题，如对“点燃烟花需要的最低能量是多少”这类问题，会在保障安全前提下给出合理实用回答 。

#GPT5 #OpenAI #AI 系统 #GPT5 能力 #安全机制 #模型升级