OpenAI 推出的 o3 推理模型，打破了传统文字思维链的边界 —— 多模态模型首次实现将图像直接融入推理过程。它不仅 "看图"，还能 "用图思考"，开启了视觉与 文本推理深度融合的问题求解方式。例如，面对一张物理试卷图像，o3 能自动聚焦公式区域，分析变量关系，并结合知识库推导出答案；在解析建筑图纸时，o3 可在推理过程中旋转或裁剪局部结构，判断承重设计是否合理。这种 "Thinking with Images" 的能力，使 o3 在视觉推理基准测试 V* Bench 上准确率飙升至 95.7%，刷新了多模态模型的推理上限。 然而，OpenAI 如何赋予 o3 这一能力，学界和工业界仍不得而知。为此， 小红书团队联合西安交通大学， 采用端到端强化学习，在完全不依赖监督微调（SFT） 的前提下，激发了大模型 "以图深思" 的潜能， 构建出多模态深度思考模型 DeepEyes，首次实现了与 o3 类似的用图像进行思考的能力，并已同步开源相关技术细 节，让 "用图像思考" 不再是 OpenAI 专属。 论文地址：https://arxiv.org/abs/2505.14362 项目地址：https://visu ...

Core Viewpoint - The article discusses the recent cautious stance of major tech companies, particularly Amazon and Microsoft, regarding the expansion of AI data centers, indicating a potential slowdown in the AI industry and a reevaluation of the demand for computing power [1][2][3]. Group 1: Company Actions - Alibaba's chairman expressed concerns about a bubble in the U.S. data center market, which has led to a negative impact on domestic AI computing power stocks [1]. - Amazon Web Services (AWS) has reportedly paused some leasing negotiations for data centers, particularly international ones, mirroring Microsoft's recent actions [1][2]. - Microsoft confirmed the suspension of a $1 billion investment plan for three data centers in Ohio, suggesting a broader trend of scaling back on new projects in the AI sector [1][2]. Group 2: Industry Trends - The collaboration between AWS and AI startup Anthropic is highlighted as a strong partnership, contrasting with Microsoft's relationship with OpenAI, which appears less stable [2]. - The emergence of open-source models, particularly DeepSeek, has led to a reassessment of the value of foundational large models, causing many AI startups to reconsider their strategies [2][3]. - The overall demand for data center computing power is expected to decline as fewer companies are developing AI models, leading to a lack of customers for data center leasing [3]. Group 3: AI Model Development - The pace of AI model advancements has reportedly slowed, with some entrepreneurs expressing disappointment over the lack of significant progress since August of the previous year [3][5]. - There is a discrepancy between AI model performance scores and user experience, with notable examples like Meta's Llama 4 and OpenAI's o3 model failing to meet user expectations despite high scores in competitive settings [3][4]. - The AI industry is experiencing a cycle similar to that of the smartphone industry, where the focus on performance metrics has overshadowed actual user experience [4][5]. Group 4: Market Sentiment - The article suggests that the current state of the AI industry reflects a broader disillusionment, as the anticipated killer applications have yet to materialize, leading to a lack of sustainable revenue-generating products [4][5]. - The rapid hiring and subsequent layoffs in Silicon Valley during the pandemic are cited as a cautionary tale, with companies now facing the consequences of overexpansion during a period of perceived growth [5][6]. - The optimism surrounding the internet industry's growth is contrasted with a more cautious outlook for AI, indicating that companies may struggle to maintain the same level of enthusiasm moving forward [6].

Core Insights - OpenAI announced that GPT-4 will be completely replaced by GPT-4o starting April 30, while GPT-4 will still be available through API [1] - GPT-4o has shown superior performance in writing, coding, and STEM tasks compared to GPT-4 during face-to-face evaluations [1] - A series of new AI models, including GPT-4.1, will be unveiled next week, which will be an improved multimodal version of GPT-4o [1] Model Developments - OpenAI will introduce smaller versions of the new model, specifically GPT-4.1 mini and nano [1] - Additionally, new reasoning models named o3 and o4-mini will be launched [1]