资讯列表
迈向人工智能的认识论：窥探黑匣子的新方法
鉴于上述困难，研究人员正在多个方面进行创新，以更好地理解和控制大型语言模型（LLM）的推理方式。 总体而言，两种互补的策略正在形成：机械分析和归因：分解模型的内部计算（电路、神经元、注意力头），将特定的决策或步骤归因于特定的组件。 行为评估和约束：设计评估指标和训练框架，区分真正的推理和表面模式，并鼓励模型在中间步骤中讲真话。
6/18/2025 2:00:00 AM
晓晓
Akamai 利用 AI 技术和 Kubernetes 平台实现云成本节省 70%
在当前的数字时代，云计算的开销持续高企，企业在计算资源的使用效率上普遍存在浪费。 根据最新统计，预计今年企业在不必要的云开销上将浪费高达445亿美元。 Akamai Technologies 作为一家大型的网络安全和内容交付服务提供商，面临着复杂的多云基础设施以及严格的安全合规要求，因此需要寻找有效的解决方案来优化云成本。
6/17/2025 11:02:12 PM
AI在线
夸克App推出“夸克老师” 具备个性化AI辅导能力
今日，夸克App正式推出全新升级的学习产品——“夸克老师”，这一AI家教的亮相，为学习领域带来了新的活力。 “夸克老师”功能强大，集讲题、批作业、出题、找试卷等多种能力于一身，堪称学生身边的得力学习助手。 其背后依托夸克灵知学习大模型，并以通义千问为基础，在解答数学、物理等高难度题目时，展现出强大的推理及讲解能力。
6/17/2025 11:02:12 PM
AI在线
LinkedIn 推出 AI 驱动的求职搜索，助力用户更精准找到理想职位
近日，LinkedIn 正式向所有用户推出了一项全新的 AI 驱动求职搜索功能。 这项新功能利用了经过提炼和微调的大型模型，旨在改善用户在平台上的求职体验。 LinkedIn 的产品开发副总裁 Erran Berger 表示:“这一全新搜索体验让用户可以用自己的语言描述求职目标，从而获得更符合需求的结果。
6/17/2025 11:02:12 PM
AI在线
OpenAI获2亿美元国防合同，预计将于明年7月完成
近日，OpenAI宣布与美国国防部达成了一项重磅合作，获得了一份价值2亿美元的为期一年的合同。 这项合同旨在为国防部开发先进的人工智能工具，帮助应对一系列国家安全挑战。 五角大楼在声明中指出，这项工作将主要在华盛顿及其周边地区进行，预计将于2026年7月完成。
6/17/2025 11:02:12 PM
AI在线
OpenAI 重磅升级 ChatGPT 搜索功能，让智能响应更精准！
在数字化快速发展的今天，信息的获取方式不断演变。 为了提升用户体验，OpenAI 近期宣布对其人工智能助手 ChatGPT 的搜索功能进行重大升级。 此次更新不仅优化了搜索质量，还让智能响应更加精准，用户体验更上一层楼。
6/17/2025 11:02:12 PM
AI在线
Gemini 2.5 Pro即将更新Deep Think的功能 提供额外计算能力
随着人工智能技术的迅猛发展，Google的Gemini系列模型持续引领行业创新。 AIbase最新获悉，Gemini2.5Pro即将正式发布全新“Deep Think”功能，这一增强推理模式被誉为AI复杂问题解决的里程碑。 Deep Think功能:多线程推理的突破根据近期社交媒体上的可靠信息，Gemini2.5Pro的Deep Think模式通过并行处理多个假设，显著提升了模型在复杂任务中的推理能力。
6/17/2025 11:02:12 PM
AI在线
智能眼镜支付再升级：Rokid“看一下支付”兼顾便捷、安全与隐私
在今日举办的 Rokid & 支付宝新品首发媒体体验会上，Rokid 隆重推出了其最新款 Rokid Glasses 智能眼镜，并宣布联合支付宝共同打造全球首个智能眼镜支付产品——“看一下支付”。 这项创新功能旨在为用户带来前所未有的便捷支付体验。 图源备注:图片由AI生成，图片授权服务商Midjourney支付宝方面强调，“看一下支付”具备便捷性、安全性、隐私性三大特点。
6/17/2025 11:02:12 PM
AI在线
Adobe Firefly 上线移动平台应用：支持文生图 / 视频，可调用第三方 AI 模型
Adobe宣布Firefly生成式AI应用现已登陆iOS/安卓平台，支持文生图、文生视频及AI照片编辑，还能使用Photoshop的生成填充和智能扩展功能。用户需Creative Cloud账号登录，部分功能需订阅。同时，Firefly Boards公测版新增视频创作能力。#AdobeFirefly# #AI生成#
6/17/2025 9:59:42 PM
漾仔
Former Google CEO's startup releases 24 billion-parameter chemical reasoning model with accuracy surpassing multiple leading models
In the field of artificial intelligence, research on large models continues to advance, particularly in improving reasoning capabilities. Recently, FutureHouse, a startup funded by former Google CEO Eric Schmidt, has open-sourced a chemical task reasoning model named ether0, with a parameter scale as high as 24 billion. This model demonstrates strong domain-specific capabilities in chemistry without requiring additional pre-training in specific fields, achieving remarkable results through post-training techniques while significantly reducing data requirements compared to traditional field-specific models.The application of reasoning models goes beyond simple multiple-choice tests.
6/17/2025 9:03:21 PM
AI在线
Akamai Achieves 70% Cloud Cost Savings Using AI and Kubernetes Platform
In the current digital age, cloud computing costs remain high, and enterprises generally waste resources in the efficiency of computing resource utilization. According to the latest statistics, it is estimated that this year companies will waste up to $44.5 billion on unnecessary cloud expenses. Akamai Technologies, a major provider of cybersecurity and content delivery services, faces complex multi-cloud infrastructure and stringent security compliance requirements, thus needing to find effective solutions to optimize cloud costs.Image Source Note: Image generated by AI, image authorized service provider MidjourneyTo address this, Akamai introduced the Kubernetes automation platform Cast AI.
6/17/2025 9:03:21 PM
AI在线
Disrupting Tradition! New Multi-Agent Framework OWL Gains 17K Stars, Surpassing OpenAI to Pioneer a New Era of Intelligent Collaboration
With the rapid development of large language models (LLMs), single agents have revealed many limitations when dealing with complex real-world tasks. To address this issue, a new multi-agent framework named Workforce and an accompanying training method called OWL (Optimized Workforce Learning) were jointly introduced by institutions such as Hong Kong University and camel-ai. Recently, this innovative achievement achieved an accuracy rate of 69.70% on the authoritative benchmark test GAIA, not only breaking the record for open-source systems but also surpassing commercial systems like OpenAI Deep Research..
6/17/2025 9:03:21 PM
AI在线
Silicon Base Flow Launches the World's First Open Source Large-scale Hybrid Attention Reasoning Model MiniMax-M1-80k
The Silicon Cloud (SiliconFlow) has officially launched the world's first open-source large-scale hybrid attention reasoning model — MiniMax-M1-80k (456B). This innovative model is designed to provide strong support for complex tasks such as software engineering, long-context understanding, and tool usage, and its performance can rival leading models like o3 and Claude4Opus.It is reported that MiniMax-M1-80k supports a maximum context length of up to 128K, greatly facilitating the handling of long texts. For users with special needs, the platform also provides backend support to meet the demand for 1M long contexts.
6/17/2025 9:03:21 PM
AI在线
Poe Platform Launches Dream 3.0 Image Model and Seedance 1.0 Lite Video Model
The Poe platform has officially launched the Image Generation Model Seedream 3.0 and the Video Generation Model Seedance 1.0 Lite, both developed by ByteDance (ByteDance), offering global users a more efficient and higher quality multi-modal content creation experience. This update marks another breakthrough for Poe in the field of image and video generation, providing creators with a seamless creation process from static images to dynamic videos.Seedream 3.0: A New Benchmark in Image Generation. As the latest image generation model under ByteDance's Volcano Engine, Seedream 3.0 has drawn industry attention due to its excellent image quality and semantic understanding capabilities.
6/17/2025 9:03:21 PM
AI在线
Kimi-Dev-72B: The AI Wonder Breaking the Boundaries of Code Repair
Recently, the much-anticipated open-source large language model Kimi-Dev-72B has officially launched, becoming a favorite among developers. This model was developed by the "Dark Side of the Moon" team and is specifically designed to solve coding problems, aiming to enhance programming efficiency.In the recent SWE-bench Verified test, Kimi-Dev-72B demonstrated extraordinary capabilities, particularly excelling in fixing code defects within Docker environments. This advantage makes Kimi-Dev-72B not only a valuable assistant for developers but also an important tool for optimizing development workflows.The core advantage of this model lies in its reinforcement learning-based optimization mechanism.
6/17/2025 9:03:21 PM
AI在线
Quark App Launches Quark Teacher with Personalized AI Tutoring Capabilities
Today, the Quark app officially launched its newly upgraded learning product — "Quark Teacher". The debut of this AI tutor has brought new vitality to the education sector."Quark Teacher" is highly functional, integrating capabilities such as explaining problems, grading homework, creating questions, and finding test papers, making it a valuable study companion for students. It relies on the Quark Lingzhi Learning large model and is based on Qwen, demonstrating strong reasoning and explanation abilities when solving complex math and physics problems.
6/17/2025 9:03:21 PM
AI在线
OpenAI Secures $200 Million Defense Contract Expected to be Completed by July Next Year
Recently, OpenAI announced a major partnership with the U.S. Department of Defense, securing a $200 million contract for one year. The purpose of this contract is to develop advanced AI tools for the Department of Defense to help address a range of national security challenges.
6/17/2025 9:03:21 PM
AI在线
开源新星 Kimi-Dev-72B：打破代码修复界限的 AI 神器
最近，备受关注的开源大型语言模型 Kimi-Dev-72B 正式上线，成为开发者们的新宠。 这款模型由 “月之暗面” 团队开发，专门为解决代码问题而设计，旨在提升编程效率。 Kimi-Dev-72B 在近期的 SWE-bench Verified 测试中展现了非凡的实力，尤其在修复 Docker 环境中的代码缺陷方面，表现得尤为出色。
6/17/2025 7:01:20 PM
AI在线