编辑 ｜ LRS 来源｜ 新智元（ID：AI_era） 封面来源 ｜ pexels AI人才竞争白热化， 大厂纷纷高薪招募实习生。 AI人才大战的火，终于烧到实习生身上了！ OpenAI、Anthropic、Meta、Google DeepMind等AI顶流公司，过去只为顶级研究员和工程师开出30–40万美元的年薪，甚至通过上百亿美元级的投资、并购 来「打包」挖团队。 现如今，类似的竞争开始下沉到实习和驻留项目：短期、入门角色的薪酬，直接对标很多行业的正式工作，传统意义上的「廉价实习生」正在消失。 巨头们给实习生开出的工资最高已经达到了1.83万美元（折合 12.8万人民币 ）， 月薪 ！ 不光要「挖」走AI人才，也要「培养」AI人才。 招募超级实习生 1.Anthropic安全研究员 Anthropic提供一个为期4个月的全职研究Fellowship，给出的定位是「加速AI安全研究、培养相关研究人才」，核心目标是让Fellows产出可公开发表的AI Safety研究成果，Anthropic透露过往届80%以上的学员最终写出了论文。 项目聚焦AI Safety，特别是可解释性（interpretability ...

Group 1 - MiniMax, a Chinese AI startup, plans to set its IPO price at the upper end of the range, potentially issuing shares at HKD 165 each [1] - The IPO is expected to raise at least HKD 4.2 billion (approximately USD 538 million) if priced at HKD 165, with a valuation between HKD 46.123 billion and HKD 50.399 billion [1] - The company has received strong demand from investors, leading to an early closure of the order book on January 5, one day ahead of schedule [1] Group 2 - MiniMax focuses on developing general AI models and has launched several AI-native products, covering over 200 countries and regions with more than 212 million personal users as of September 2025 [2] - The company's revenue grew by over 170% year-on-year in the first nine months of 2025, with over 70% of revenue coming from international markets [2] Group 3 - The company is set to officially list on the Hong Kong Stock Exchange on January 9, with the stock code "0100.HK" [3]

Core Insights - Yann LeCun, a Turing Award winner and a key figure in deep learning, has left Meta to become the Executive Chairman of AMI Labs, revealing internal turmoil at Meta regarding its AI strategy and leadership changes [1][12] Group 1: Departure from Meta - LeCun confirmed speculation about his departure from Meta, citing a crisis of integrity related to the Llama 4 model's testing results and a significant shift in the company's AI strategy [1][5] - The internal conflict escalated after Meta's CEO, Mark Zuckerberg, made a controversial decision to invest approximately $14.3 billion in acquiring a 49% stake in Scale AI, appointing 28-year-old Alexandr Wang as Chief AI Officer [6][8] Group 2: AI Strategy and Leadership Changes - The introduction of Wang led to a restructuring of Meta's AI research, consolidating various departments under his leadership, which marginalized LeCun's role [8][11] - Wang's focus on large language models (LLMs) as the sole path to achieving superintelligence conflicted with LeCun's belief in the importance of foundational research and alternative approaches [9][10] Group 3: Cultural and Operational Shifts - The shift in strategy resulted in a loss of academic freedom within Meta's AI research labs, leading to a culture that prioritized commercial viability over scientific exploration [11][12] - A new policy mandated that research papers must be approved for commercial relevance before publication, causing discontent among researchers and contributing to significant talent attrition [11][12] Group 4: Formation of AMI Labs - Following his departure, LeCun founded AMI Labs, aiming to explore scientific paths that were sidelined in the competitive landscape of tech giants, with an initial funding target of €500 million and a valuation of €3 billion [12][14] - LeCun has chosen not to take on the CEO role at AMI Labs, preferring to focus on scientific endeavors while leaving management to experienced professionals [14]

Core Insights - The article discusses the importance of spatial understanding capabilities in multimodal large language models (MLLMs) for their transition into real-world applications as "general intelligent assistants" [2] - It highlights the limitations of existing spatial intelligence evaluation benchmarks, which either rely heavily on template generation or focus on specific spatial tasks, making it difficult to comprehensively assess models' spatial understanding and reasoning abilities in real-world scenarios [2] Group 1: Introduction of MMSI-Video-Bench - The Shanghai Artificial Intelligence Laboratory's InternRobotics team has launched a comprehensive and rigorous spatial intelligence video benchmark called MMSI-Video-Bench, designed to challenge current mainstream multimodal models [2][6] - The benchmark aims to evaluate models' spatial perception, reasoning, and decision-making capabilities in complex and dynamic real-world environments [2][7] Group 2: Benchmark Characteristics - MMSI-Video-Bench features a systematic design of question types that assess models' basic spatial perception abilities based on spatiotemporal information [6] - It includes high-level decision-making evaluations and extends task categories to cover complex real-world scenarios, testing models' cross-video reasoning capabilities, memory update abilities, and multi-view integration [6][8] - The benchmark consists of five major task types and 13 subcategories, ensuring a comprehensive evaluation of spatial intelligence [10] Group 3: Challenge and Performance - The benchmark's questions are designed to be highly challenging, with all models tested, including the best-performing Gemini 3 Pro, achieving only a 38% accuracy rate, indicating a significant performance gap of approximately 60% compared to human levels [10][14] - The evaluation reveals that models struggle with spatial construction, motion understanding, planning, prediction, and cross-video reasoning, highlighting critical bottlenecks in their capabilities [14][15] Group 4: Error Analysis - The research team identified five main types of errors affecting model performance: detailed grounding errors, ID mapping errors, latent logical inference errors, prompt alignment errors, and geometric reasoning errors [17][21] - Geometric reasoning errors were found to be the most prevalent, significantly impacting performance, particularly in spatial construction tasks [19][21] Group 5: Future Directions - The article suggests that introducing 3D spatial cues could assist models in understanding spatial relationships better, indicating a potential direction for future research [22][24] - It emphasizes the need for effective design of spatial cues that models can truly understand and utilize, as current failures are attributed to underlying reasoning capabilities rather than a lack of explicit reasoning steps [27]

机器之心编辑部 近日，X 知名博主、Hyperbolic 联创 & CEO Yuchen Jin 发帖称，如果在他读博士的时候就有 Claude Code、Gemini 和 ChatGPT 等各类 AI 工具出现，那么也许只要 一年就能毕业，而不是用了 5.5 年。 而他之所以发出这个感慨，缘由是最近一些硅谷 AI 大厂工程师表示，在用了 AI 工具后，项目完成时长被大幅压缩…… 先是谷歌首席工程师、Gemini API 负责人 Jaana Dogan 在 X 上发文称：「我不是在开玩笑，这也不好笑。从去年开始，我们就在谷歌内部尝试构建分布式 Agent 编 排器。有多种选择，大家并没有完全认同…… 我只是向 Claude Code 描述了问题，它就在一小时内生成了一个东西，而这几乎就是我们去年一年所做的东西。」 随后，她又发文补充，提示内容不算详细，也没有具体细节，只是一段三段式的描述。但由于不能分享任何东西，也不好具体展示出来，总结来说就是在现有一 些想法基础上构建一个玩具版本，用以评估 Claude Code。 随后此推文获得了上百次的浏览，而该网友也发文认真做起了自我介绍，原来 Rohan Anil ...

钛媒体App 1月5日消息，2025年，中国一级市场达成了一个昂贵的共识："AI 应用元年"。据IT桔子、钛媒体TMTBASE数据显示，2025年截至12月，标签包 含AI应用且拿到新融资的公司总数为930家，融资总金额高达1070.7亿元（人民币）。这意味着，在2025年的每一天，都有2.6家公司拿到融资，平均每小时 有1200万资金进场。 | 场景 | 平均融资金额 | | --- | --- | | | (万元人民币) | | Al+自动驾驶 | 45,286 | | Al+通用 | 17,839 | | 具身智能 | 17,411 | | Al+农业 | 10,000 | | Al+医疗 | 9,940 | | Al+内容生成 | 9,867 | | Al+新药研发/合成生 | 9,486 | | 物 | | | Al+编程 | 8,721 | | Al+科学 | 8,351 | | Al+交通 | 8,300 | 数据来源: IT桔子、钛媒体TMTBASE 从资金集中度来看，TOP 3场景（具身智能、自动驾驶、通用）的融资总金额占比，超过总融资金额的一半，达51.6%。其中，具身智能赛道以337.7 ...

这些注释字迹残损严重，夹杂着大量中世纪拉丁文缩写，几个世纪以来，学者们始终无法解释它的含义。 然而，Gemini 3.0 Pro仅在一个小时内，就清晰地给出了解读！ 它成功识别出：这段注释并非随意的标记，或者装饰性的涂画，而是与不同圣 经年代学体系之间的比较和计算有关。 2026开年王炸！Gemini 3.0 Pro仅用1小时，暴力破解533年未解的《纽伦堡编年史》天书。从0.02美元的算力成本到精准复原16世纪学霸的历法对账单， AI正以全知视角降维打击传统考古！ 就在刚刚，500年前的《纽伦堡编年史》天书，被AI破解了！ 其中的一段手写注释，难倒了人类历史学家整整500年。 也就是说，几百年前作者的逻辑，被AI精准地捕捉到，完成了整套推理！ 研究者们激动地在博客中写道—— 令人难以置信的是，LMM的视觉理解能力已经发展到Gemini 3 Pro能阅读 500 年前的手写缩写速记旁注，回过头去阅读整页印刷内容，并利用页面内容来 推演和澄清速记的含义，然后将所有这些信息整合起来，得出一个能契合所有拼图碎片的最终理解，而这一切都不需要任何形式的人类协助！ 老祖宗的古籍，被AI破译了！ 《纽伦堡纪事报》是一部出版 ...

作者 | 陈勇 在海量视频检索场景中，传统方法往往"重视觉、轻听觉"，且网络结构设计更多依赖经验与人工尝试，难以同时兼顾高效存储与快速检索。那么，是否 存在一种能够自动找到最优结构、并充分发挥多模态价值的方案？ 近日，来自北邮与北大的研究团队提出 AV-NAS，在多模态视频哈希领域首次引入神经架构搜索（NAS），构建了一个同时覆盖 Transformer 与 Mamba 的统一搜索空间。该方法不仅使模型能够自动发现最优的跨模态融合机制（Cross-Mamba），还揭示了一个颇具启发性的结论——在音频时序 建模任务中，看似简单的 "CNN + FFN" 结构竟然优于复杂的 Transformer 方案。 论文题目： AV-NAS: Audio-Visual Multi-Level Semantic Neural Architecture Search for Video Hashing 论文链接： https://dl.acm.org/doi/10.1145/3726302.3729899 代码链接： https://github.com/iFamilyi/AV-NAS 目前，AV-NAS 已被 SIGIR 2 ...

Core Viewpoint - CNBC host Jim Cramer endorses a J.P. Morgan report indicating that physical and financial constraints, rather than a market crash, will limit tech giants' spending on artificial intelligence (AI) [1] Group 1: Physical Constraints - Cramer argues that fears of an AI bubble similar to the dot-com era lack nuance, with "electric power gating" being the main factor preventing overspending by hyperscalers [2] - The J.P. Morgan report highlights U.S. power generation constraints as a significant risk for the AI sector, with data centers expected to drive two-thirds of U.S. load growth while only adding 25 GW of reliable capacity in 2024 [3] - This scarcity of electricity acts as a hard cap on the speed at which companies can deploy new infrastructure, effectively limiting their capital expenditures [3] Group 2: Financial Constraints - Major players like OpenAI will face balance sheet constraints, with the J.P. Morgan report noting substantial financial commitments that may exceed current revenues [3] - OpenAI has committed to pay Oracle Corp. $60 billion per year for compute facilities that are not yet built, highlighting the financial strain [3] - OpenAI's commitments to corporate partners total $1.4 trillion, while its revenue primarily comes from subscription fees, making profitability a significant challenge [4] Group 3: Market Dynamics - Cramer suggests that tangible constraints on power and capital will slow AI spending, preventing the speculative behavior seen during the 2000 market bubble [4] - The J.P. Morgan report contrasts today's market with the dot-com bubble, noting that current high valuations are supported by high profit margins, with 42 AI-related companies contributing up to 75% of S&P 500 earnings growth since late 2022 [7] - A shift in financing is occurring, with companies like Meta Platforms and Oracle increasingly relying on debt markets for data center expansions, indicating a new discipline in capital management [8][9]

Core Insights - The emergence of Claude Code has led to significant discussions about the concept of the technological singularity, with Elon Musk declaring that "we have entered the singularity" and predicting 2026 as the pivotal year for this event [1][3][7]. Group 1: Technological Singularity - The concept of the technological singularity, which refers to a point where technology accelerates exponentially, has been brought to the forefront by recent advancements in AI, particularly with Claude Code [7]. - Ray Kurzweil, a prominent futurist, previously predicted the singularity would occur around 2045, but recent developments have prompted a reevaluation of this timeline [7][29]. - The rapid advancements in AI capabilities, particularly in coding and problem-solving, suggest that the singularity may be closer than previously thought, with some experts indicating it could happen as soon as 2026 [3][19]. Group 2: Claude Code and AI Advancements - Claude Code has demonstrated remarkable programming capabilities, with its efficiency reportedly increasing by 220% when used in conjunction with Claude Opus 4.5, which is considered the top coding model globally [17][19]. - The latest LiveBench rankings show Claude Opus 4.5 outperforming competitors like GPT-5.1 Codex MAX and Gemini 3 Pro, indicating its superior performance in real-world applications [19][20]. - The ability of Claude Opus 4.5 to handle long-duration coding tasks without failure marks a significant milestone in AI development, showcasing its robustness and reliability [21]. Group 3: Impact on Software Engineering - The integration of AI in software development is transforming the role of engineers, with many now focusing on code review rather than traditional coding tasks, as AI handles the majority of the coding workload [32]. - The shift towards using natural language as a new programming syntax suggests that the barriers to software development are lowering, allowing individuals without coding experience to create functional applications quickly [25][34]. - As AI continues to automate coding and related tasks, the implications extend beyond software engineering to operations and management, potentially reshaping the entire landscape of work in technology [35].