资讯列表
AI三问③模型之问 | 直面模型之问，以大爱共塑 AI 未来 ——WAIC 2025 大模型论坛以问题破局引领技术革新
WAIC 2025世界人工智能大会论坛：2025年7月26日-28日展览：2025年7月26日-29日地点：世博中心、世博展览馆、徐汇西岸等为深入AI核心命题，今年WAIC重磅提出“AI三问”——直击数学、科学和模型方向亟待探讨与思辨的前沿话题。 数学之问以公理公式推演规律，建认知框架；科学之问扎根实证，探自然本质；模型之问融合二者，化抽象为实用。 三问共生，数学为科学量化，科学赋数学意义，模型促智慧落地，三者协同在多元领域中彰显创新的深层价值。
7/22/2025 10:35:00 AM
新闻助手
超越Claude 3.5和o1！8B模型靠「分层投票+测试时训练」逆袭
测试时训练（test-time training）是一种通用的训练方法。 该方法将单个未标记的测试实例转化为自监督学习问题，在对测试样本进行预测之前更新模型参数。 而对于大模型训练，通常会使用一种称为情境学习的技术来提高其模型在新任务上的性能。
7/22/2025 10:22:02 AM
Google Gemini Advanced Version Makes a Stunning Debut! Wins Gold Medal at IMO 2025, the New Challenger in the Mathematical Olympiad!
Google DeepMind recently announced that its latest Gemini Advanced AI model performed exceptionally well in the 2025 International Mathematical Olympiad (IMO) competition, successfully winning a gold medal. This achievement not only draws attention but also marks a qualitative leap in AI technology's ability to solve complex problems.According to the information, this honor was not easily obtained. Gemini Advanced solved five problems in the competition, scoring a total of 35 points, meeting the standard for an IMO gold medal.
7/22/2025 9:52:55 AM
AI在线
SoftBank and OpenAI Part Ways! 50 Billion Dollar 'Interstellar Gateway' Project Faces Crisis
On July 22 Beijing time, the Wall Street Journal reported that the "Star Gate" project announced by the White House in January this year is facing numerous difficulties. This ambitious project aimed at promoting the development of artificial intelligence (AI) in the United States is currently progressing far slower than expected.
7/22/2025 9:52:49 AM
AI在线
ChatGPT Daily Prompt Volume Exceeds 2.5 Billion: AI Conversation Popularity Continues to Surge
According to the latest data disclosed by OpenAI to Axios, its flagship product ChatGPT processes 2.5 billion prompts from users around the world every day, of which approximately 330 million come from users in the United States. This impressive figure fully demonstrates the popularity and user engagement of ChatGPT on a global scale.Although the number of prompts for ChatGPT is huge, it still lags behind traditional search engine giant Google.
7/22/2025 9:52:48 AM
AI在线
AI Gold Medal Showdown: OpenAI vs. Google DeepMind in a Tense Competition, Who is the True Winner?
Recently, two major giants in the artificial intelligence field, OpenAI and Google DeepMind, have announced that their AI models won gold medals in the 2025 International Mathematical Olympiad (IMO).
7/22/2025 9:52:46 AM
AI在线
DeepSeek Downloads Drop 72.2% as Competition in AI Applications Intensifies
According to the "2025 Second Quarter AI Application Value Ranking" released by QuestMobile, DeepSeek had an average monthly download volume of 81.113 million in the first quarter, but this number dropped sharply to 22.589 million in the second quarter, a decrease of 72.2%. This change marks the gradual fading of DeepSeek's shine, and its appeal in the market has significantly weakened.At the same time, Doubao stood out in terms of download volume, becoming the new leader in downloads.
7/22/2025 9:52:44 AM
AI在线
72% of American Teenagers Have Tried AI Companions, Study Reveals Usage Status
According to a new study by the U.S. nonprofit organization Common Sense Media, the majority of American teenagers (72%) have tried an AI companion at least once. An "AI companion" refers to an AI chatbot capable of more personal conversations, such as Character.AI or Replika, rather than just some assistants for homework.Image source note: The image is AI-generated, and the image licensing service is MidjourneyThe study, based on a survey of 1,060 teenagers, found that these teens showed great interest in chatting with AI.
7/22/2025 9:52:43 AM
AI在线
ChatGPT Receives 2.5 Billion Questions Daily, Sparking a Trend in Artificial Intelligence Usage!
According to OpenAI, which disclosed the information to Axios, ChatGPT now receives up to 2.5 billion questions per day from users around the world. users contributed about 330 million questions. These impressive figures demonstrate the extensive application of OpenAI's flagship product.Although Alphabet, Google's parent company, has not released daily search data, recent disclosures show that Google receives approximately 5 trillion queries per year, or about 1.4 billion searches per day on average.
7/22/2025 9:52:42 AM
AI在线
Aliyun Tongyi Qianwen Launches Qwen3 Model, AI Technology Upgraded Again!
Ali Tongyi Qwen recently announced the latest update to its flagship product, the Qwen3 model, introducing a new version called Qwen3-235B-A22B-Instruct-2507-FP8. The highlight of this new version is its non-thinking mode and a significant improvement in long-text processing capabilities, expanding to a text length of 256K.
7/22/2025 9:52:38 AM
AI在线
Instacart CEO Fidji Simo Takes Over OpenAI's Application Division
Instacart CEO Fidji Simo will officially join OpenAI on August 18 as the newly created position of **"Chief Officer for Applications."** She will report directly to OpenAI CEO Sam Altman and will be responsible for leading at least one-third of the company's team, focusing on expanding and developing use cases for OpenAI technology.This new position was first announced by Altman during the company's reorganization in May this year.
7/22/2025 9:52:35 AM
AI在线
AI时代设计师的未来在哪里？阿里D20峰会揭秘这6个机遇方向！
大家好，我是姝斐，今年 7 月，我又如期参加了阿里巴巴举办的设计界顶级盛会-D20 全球设计院长峰会。 大会汇聚了清华大学、中国美术学院等设计顶级学府的院长、行业内做出大成果的企业高管，他们都结合实际落地项目案例一一对行业做了非常深刻的剖析分享，让我们看到了在 AI 时代，原来 AI 已经落地到如此成熟惊人的地步，AI 已经不再是一个口号，而是实实在在给我们带来了便捷与红利，下面我将分享几点本次峰会给我带来的冲击思考。 更多D20的干货：一、设计师要积极拥抱变化.
7/22/2025 8:03:51 AM
姝斐suphie
阿里云通义千问 Qwen 3 旗舰版模型宣布更新：性能全面提升，超越 Kimi、DeepSeek 等行业顶尖水平
AI在线 7 月 22 日消息，阿里云今天更新了旗舰版 Qwen3 模型，推出 Qwen3-235B-A22B-FP8 非思考模式（Non-thinking）的更新版本，命名为 Qwen3-235B-A22B-Instruct-2507-FP8。 阿里云表示，在经过与社区沟通和深思熟虑后，决定停止使用混合思考模式，转为分别训练 Instruct 和 Thinking 模型，以获得最佳质量。 据介绍，新的 Qwen3 模型通用能力显著提升，包括指令遵循、逻辑推理、文本理解、数学、科学、编程及工具使用等方面，在 GQPA（知识）、AIME25（数学）、LiveCodeBench（编程）、Arena-Hard（人类偏好对齐）、BFCL（Agent 能力）等众多测评中表现出色，超过 Kimi-K2、DeepSeek-V3 等顶级开源模型以及 Claude-Opus4-Non-thinking 等领先闭源模型。
7/22/2025 6:34:44 AM
问舟
深度解析 MEM1：开启智能体长时序高效交互之门
大家好，我是肆〇柒。 相信大家在落地AI应用的时候，对长上下文的处理时所遇到的挑战，都深有感触。 大型语言模型（LLM）于单轮任务中表现卓越，然而现实应用场景日益复杂，多轮交互需求剧增。
7/22/2025 4:00:00 AM
四零柒
深度解密国内BAS如何实现AI赋能的智能化安全验证攻防新范式！
在网络安全形势日益复杂、各类威胁不断涌现的当下，人工智能（AI）已然成为推动各行业数字化转型以及提升安全防御能力的关键力量。 AI正从根本上重塑BAS(入侵与攻击模拟)，使其从传统的“模拟验证工具”进化为“智能对抗中枢”。 其核心价值在于：通过赋能动态攻击生成、自适应路径探索、智能化结果分析与自动化策略闭环，将BAS的全生命周期智能化。
7/22/2025 2:11:00 AM
安全牛
Dify 1.6.0原生MCP开发出行助手智能体
前言AI智能体通过感知环境、自主决策和执行任务，突破传统大模型仅限于语言交互的局限。 然而，智能体的开发长期受制于接口碎片化与工具兼容性难题。 开发者需为不同数据源编写适配代码，这一痛点催生了MCP协议的诞生。
7/22/2025 2:00:00 AM
AI大模型应用开发
AIGC、RAG、Agent、Function Call、MCP 到底啥关系？一次讲明白！
Hello，大家好，我是 Sunday。 最近很多同学特别关注 AI 相关的领域。 但是，AI 技术发展太快了，AIGC、RAG、Agent、Function Call、MCP 等等的各种热词层出不穷的。
7/22/2025 1:55:00 AM
程序员Sunday
AI Code Review Star Greptile Raises $30 Million in Funding, Valuation Reaches $180 Million
According to TechCrunch, the funding round amounted to $30 million, with a valuation of $180 million, led by Eric Vishria, a partner at the renowned venture capital firm Benchmark. However, sources indicated that the deal has not yet been finalized, and specific terms may still be adjusted.Greptile was founded by Dasksh Gupta, who started his entrepreneurial journey after graduating from Georgia Tech in 2023.
7/22/2025 1:42:29 AM
AI在线