近日，全球食品分选巨头陶朗食品将携两款双AI光学分选分级平台——Spectrim搭载LUCAi™与 InVision²搭载LUCAi™——亮相2025中国国际水果展和亚洲国际果蔬展览会。这是陶朗双AI分选分级平 台首次在中国境内同台展出，旨在展示其在水果AI分选领域的前沿成果，为行业技术革新注入前导力 量，帮助水果企业解决分选领域的疑难杂症。 双AI分选，重新定义水果的"智"与"质" 作为在分选领域深耕50多年的引领者，陶朗自2010年起便致力于人工智能与食品分选技术的探索与创 新。其自主研发的LUCAi™人工智能引擎基于深度学习技术，通过250000多张数据图片驱动数据库，能 够精准识别水果的各类瑕疵，如蓝莓的疤痕、发霉、发软、皱缩、开裂等。 陶朗此次展出的两款双AI光学分选分级平台，以LUCAi™人工智能引擎持续迭代的算法为核心，用双 AI系统驱动水果分选智能升级，突破水果复杂瑕疵难以识别的瓶颈，如传统机器学习难以识别的轻度 腐烂、日灼伤、软果、剪切损伤、果窝缺陷等。 陶朗这一次智能升级，是无数个日夜的积累。AI系统依赖的数据库，来自陶朗在全球不同地区20多个 果季采集的水果图像训练数据。通过对模型不断地 ...

以下文章来源于AGI接口 ，作者陈伊凡 出品｜虎嗅科技组 作者｜陈伊凡 编辑｜苗正卿 头图｜AI生成 "AI 原生 100" 是虎嗅科技组推出针对 AI 原生创新栏目，这是本系列的第「 15 」篇文章。 "我们确信不久的将来会有一个'大饼'掉下一个，虽然具体什么时候掉下来我们不知道，但我们要先 把盘子做好、做大，否则掉下来了我们也接不住。" 姜涛如此描述自己现在做的事。说这句话时，姜涛正坐在音潮办公室的会客区里。 AGI接口 . AI卷起的财富风暴。 那段时间，快手的办公室里总飘着奇奇怪怪的旋律。姜涛带着算法团队和阿怪挤在同一个隔间，一边 用代码训练模型，一边听阿怪讲和弦走向、编曲等乐理知识。这种奇妙的碰撞，成为了他音乐素养的 启蒙。 他亲历了中国AI音乐发展全过程，在过去的十多年时间，他几乎没有离开过AI音乐赛道，如今，他 是一家专注音乐大模型和AI音乐产品的公司——音潮的CEO。 如果第一次见到姜涛，很容易将他和音乐人联系在一起，但你很难想到这是一位从事了多年AI算法 的技术男。他穿着米色休闲亚麻西装，有时候还会带着一个圆形礼帽，很像一个玩爵士乐的"老炮 儿"。他会给妻子做歌，把女儿稚嫩的声音编进旋律，浪 ...

| 机器之心报道 | | --- | 机器之心编辑部 全世界、所有科学领域都算上，现在最热门的方向就是 AI 了。 图灵奖得主 Yoshua Bengio，近日成为了有史以来被引用次数最多的科学家：他的总被引用量高达 973,655 次，近五年引用量达到 698,008 次。 这项统计来自 AD Scientific Index，这是一个全球性的学术排名和分析平台，旨在评估和展示科学家、研究人员以及学术机构的科研表现和影响力。 参与这次排名的共计 2,626,749 名科学家，分布在 221 个国家和地区，隶属 24,576 家机构。排名依据总引用量和近五年的引用指数进行排序。值得一 提的是，这次排名不止 AI 领域，还包括医学等 13 个主要学科和 221 个学术细分学科。 我们再回到 Bengio 的研究。从学术主页来看，Bengio 2014 年提出的 「生成对抗网络（Generative Adversarial Nets）」 引用量已突破 10 万次， 甚至超过了他与 Yann LeCun 和 Geoffrey Hinton 合著的经典论文 「Deep Learning」，不过，后者的引用量同样也超过 ...

一水 发自 凹非寺 量子位 | 公众号 QbitAI 魔镜魔镜，谁是有史以来被引用次数最多的科学家？ 答案： 深度学习三巨头之一、图灵奖得主Yoshua Bengio 。 如你所见，之所以提出这个问题，其实是因为相关消息正在引起热议ing。 并且这一次，Bengio的"最高引"头衔不仅限于计算机领域，而是"称霸"所有学科，属于 "各领域被引用次数最多的在世科学家" 。 在这之前，早在2018年，Bengio就是世界计算机研究者中单日引用次数最高的人 （同一年获图灵奖） ，2022年还一举成为世界上被引用次 数最多的计算机科学家。 其贡献最大的几篇论文《一种神经概率语言模型》（发表于2003年）、《Generative adversarial nets》（发表于2014年的GAN）、 《Deep learning》（发表于2015年）全都为深度学习领域奠定了重要基础，深刻影响着如今大火的自然语言处理、计算机视觉等研究。 | | | 而在网友们的讨论中，热议背后更深层的意义也逐渐明晰：AI的胜利。 Bengio改变了人工智能，其对深度学习的贡献真正塑造了现代人工智能研究。 所以，借此机会，我们不妨再来回顾一下Be ...

截至目前， 地平线 HSD已与全球近10家汽车品牌达成合作意向，全球首发搭载奇瑞的星途星纪元 E05量产在即 。 截至目前，地平线HSD已与全球近10家汽车品牌达成合作意向。 本文为IPO早知道原创 作者｜ Stone Jin 微信公众号｜ipozaozhidao 据 IPO早知道消息， 地平线 日前 发布 了 有史以来最大升级的高性能城区辅助驾驶产品 HSD （ Horizon SuperDrive） ，其作为 基于征程 ®6P打造的一段式端到端辅助驾驶系统，实现了从"光 子输入到轨迹输出"的系统超低时延，大幅提升 了 辅助驾驶安全、效率、舒适。 值得一提的是， 地平线最新版本的 HSD 现已被誉为 "国内最像特斯拉FSD的辅助驾驶系统" —— HSD可行驶区域覆盖城区、高速、乡间小路、停车场等道路场景 ， 不仅在连续弯道、多出口大型环 岛等复杂道路拓扑环境下，依然能实现精准感知与丝滑通行，还能从容完成高难度直行待行区识别、 盲区遮挡防御性驾驶、潮汐车道通行等复杂城区驾驶任务，并实现不依赖于记忆建图的园区漫游及车 位到车位的流畅驾驶 。 某种程度上而言， HSD 于今年年初的发布及日前完成的进一步迭代，再一 ...

Core Viewpoint - The recent advancements in large model technology have renewed interest in AI-specific chips, particularly Google's TPU, which has evolved significantly since its deployment in 2015, now reaching its 7th generation [1][9]. Group 1: TPU Overview - TPU is a specialized chip designed by Google to enhance the speed of machine learning model inference and training, focusing on executing mathematical operations efficiently [9]. - The architecture of TPU allows it to perform matrix multiplication efficiently, which constitutes a significant portion of computations in deep learning models [14][31]. Group 2: TinyTPU Project - The TinyTPU project was initiated by engineers from Western University in Canada to create an open-source ML inference and training chip, motivated by the lack of a complete open-source codebase for such accelerators [5][7]. - The project emphasizes a hands-on approach to learning hardware design and deep learning principles, avoiding reliance on AI tools for coding [6]. Group 3: Hardware Design Insights - The project team established a design philosophy of exploring unconventional ideas before consulting external resources, leading to the re-invention of many key mechanisms used in TPU [6]. - The hardware design process involves understanding clock cycles, using Verilog for hardware description, and implementing a systolic array architecture for efficient matrix multiplication [10][12][26]. Group 4: Training and Inference Mechanisms - The TinyTPU architecture allows for continuous inference by utilizing a double buffering mechanism, which enables the loading of new weights while processing current computations [61][64]. - The training process leverages the same architecture as inference, with additional modules for gradient calculation and weight updates, allowing for efficient training of neural networks [71][118]. Group 5: Control and Instruction Set - The control unit of TinyTPU employs a custom instruction set architecture (ISA) to manage control signals and data flow, enhancing the efficiency of operations [68][117]. - The ISA has evolved to include 94 bits, ensuring that all necessary control flags and data fields are accounted for without compromising performance [117].

Section 1: Investment Rating of the Reported Industry - The provided content does not mention the industry investment rating [1][2] Section 2: Core Views of the Report - Deep learning may be used for convertible bond pricing. Based on the Universal Approximation Theorem (UAT), if there is a reasonable analytical solution for convertible bond pricing, a neural network model can fit the result [2][5] - A Multilayer Perceptron (MLP) model is designed. It uses 11 factors, including core factors, convertible bond-specific factors, and market performance factors, to nonlinearly fit the pricing characteristics of convertible bonds [2][5] - The MLP model has good convergence and excellent extrapolation generalization ability. It can strongly explain out-of-sample data from 2024 to Q1 2025 [2][8] - After developing multiple models, including the MLP, MC, and traditional BS models, they can assist in investment activities in various scenarios such as new bond pricing, market interpretation, and clause pricing [2][13] - The neural network model indicates that the current market pricing of convertible bonds is overestimated, but not as much as expected. Convertible bond valuations are high but may still rise further [2][13] - In new bond pricing, the MLP and MC models form a "high-low combination." The MC model is better at pricing large-scale, high-rated convertible bonds, while the MLP is more effective for regular convertible bond listings [2][16] - The models also work well for pricing convertible bond downward revisions [2][19] Section 3: Summary by Relevant Catalog 1. Deep Learning Pricing Model's Concept and Design - The MLP model is based on the idea that if there is an analytical solution for convertible bond pricing, a neural network can fit it. It uses 11 factors for pricing [5] - The model has good convergence and generalization ability. After training with data from 2022 - 2023 and cleaning the dataset, it can effectively explain out-of-sample data from 2024 to Q1 2025 [8] - Compared with the BS and MC models, the MLP model has better pricing results for the overall market and individual convertible bonds. It has faster computation speed than the MC model and is more suitable for real - world scenarios than the BS model. However, it has limitations such as being a "black box" and requiring a large amount of historical data [10][11] 2. Convertible Bond Quantitative Pricing 2.0 - What Are the Model's Applications? - With multiple models (MLP, MC, and BS), they can assist in investment activities in various scenarios [13] - At the overall market pricing level, the neural network model shows that the current market pricing is overestimated, but not significantly. Convertible bond valuations are high but may still increase [13] - In new bond pricing, the MLP and MC models complement each other. The MC model is better for large - scale, high - rated convertible bonds, and the MLP is better for regular convertible bonds. Over 50% of convertible bond listing prices fall within the range defined by the two models, and over 80% are captured after the pricing repair in November 2024 [16] - For downward revision pricing, the MLP model can predict prices when the convertible bond is revised to the trigger threshold and to the lowest level. Most convertible bond prices on the second trading day after a downward revision proposal fall within or near this predicted range [19]

Core Viewpoint - The article discusses the launch of the ninth session of Baidu's Chief AI Architect Training Program (AICA), highlighting its significance in cultivating AI talent and the increasing interest from top executives across various industries in AI education [2][41]. Group 1: AICA Program Overview - The AICA program aims to train composite AI architects who can engage in both technical development and project implementation, leveraging Baidu's self-developed deep learning platform, PaddlePaddle, and the Wenxin large model [5][41]. - This session has attracted 96 students selected from over 500 applicants, with 61% coming from state-owned enterprises, listed companies, and leading T1 application service providers [42][41]. - The curriculum includes new modules on Wenxin open-source, cutting-edge technologies, multimodal data, and practical case studies of Baidu's key technologies [44][41]. Group 2: Industry Trends and AI Development - The focus of discussions during the opening ceremony was on large models, which accounted for 51% of the topics covered, emphasizing their role in driving industrial transformation [6][7]. - AI is seen as a pivotal technology for economic development, comparable to the steam engine and the internet, with a shift towards practical applications in various sectors such as manufacturing, healthcare, and finance [12][13]. - Current trends in AI development include a shift from technical competition to commercial applications and a consolidation of industry leadership among major companies [18][20]. Group 3: Challenges and Recommendations - Despite the advancements, the effectiveness of AI products has not fully materialized, with issues of homogeneity and lack of innovation in new products [20]. - There is a need for AI to be closely integrated with core business operations to demonstrate its value and drive revenue [21][20]. - Recommendations include focusing on value creation through AI, enhancing talent development, and fostering collaboration between AI companies and user enterprises [21][20]. Group 4: Technical Insights and Future Directions - The evolution of large models has led to significant improvements in multi-task generalization capabilities, with AI code generation adoption rates increasing from 5% and 15% in 2022 to 50% and 80% [28][25]. - The article outlines four key areas for AI architects to focus on: prompt engineering, model tuning, full-stack system design, and understanding industry-specific challenges [33][32]. - The future of large models will continue to rely on the Transformer architecture while emphasizing the importance of expert MoE structures and efficient inference deployment [36][40].

Core Viewpoint - The article discusses the urgent need for novel antibiotics to combat antibiotic resistance, highlighting the potential of generative artificial intelligence (AI) in designing new antibiotic compounds [2][5][11]. Group 1: Antibiotic Resistance Crisis - Antibiotic resistance (AMR) has led to 4.71 million deaths globally in 2021, with 1.14 million directly attributable to AMR [2]. - The CDC has classified Neisseria gonorrhoeae and Staphylococcus aureus as "urgent" and "serious" threats due to their widespread resistance to existing antibiotics [5]. - Between 1980 and 2003, only five new antibacterial drugs were developed by the top 15 pharmaceutical companies, indicating a critical need for innovative compounds [5]. Group 2: Generative AI in Antibiotic Development - Generative AI can design antibiotic molecules from scratch, allowing for the exploration of vast chemical spaces beyond existing compound libraries [7][11]. - The research team developed a generative AI platform that successfully designed two novel antibiotic molecules targeting resistant bacteria, demonstrating safety in human cells and efficacy in reducing bacterial load in mouse models [3][10]. Group 3: Research Methodology - The study utilized two methods for antibiotic design: a fragment-based approach (CReM) and an unconstrained de novo generation method (VAE), resulting in over 36 million novel compounds with predicted antibacterial activity [8][10]. - Out of 24 synthesized compounds, seven exhibited selective antibacterial activity, with two lead compounds (NG1 and DN1) showing significant efficacy against multi-drug resistant strains [10][11]. Group 4: Implications and Future Directions - The generative AI framework developed in this research provides a platform for exploring unknown chemical spaces, potentially leading to the discovery of new antibiotics [11].

Core Viewpoint - The article discusses the evolution of the partnership between Tesla and NVIDIA in the autonomous driving sector, highlighting the challenges and innovations that have shaped their collaboration. Group 1: Tesla's Journey in Autonomous Driving - In September 2013, Tesla officially entered the autonomous driving arena, emphasizing internal development rather than relying on external technologies [5] - Initially, Tesla partnered with Mobileye due to the lack of suitable self-developed autonomous driving chips, enhancing Mobileye's technology with unique innovations like Fleet Learning [9][12] - Tensions arose between Tesla and Mobileye as Tesla sought to develop its own algorithms, leading to Mobileye's demand for Tesla to halt its internal vision efforts [12][13] Group 2: NVIDIA's Strategic Shift - In 2012, NVIDIA's CEO Jensen Huang recognized the potential of autonomous driving in electric vehicles, leading to a focus on deep learning and computer vision [15] - By November 2013, Huang highlighted the importance of digital computing in modern vehicles, indicating a shift towards automation in the automotive industry [17] - In January 2015, NVIDIA launched the DRIVE brand, introducing the DRIVE PX platform, which provided significant computational power for autonomous driving applications [18] Group 3: The Partnership Development - Following a significant accident in May 2016, Mobileye ended its partnership with Tesla, prompting Tesla to choose NVIDIA as its new technology partner [19][20] - In October 2016, Tesla announced that all its production models would feature hardware capable of full self-driving capabilities, utilizing NVIDIA's DRIVE PX 2 platform [20] - By early 2017, Tesla publicly announced its plans to develop its own chips, indicating a shift in its strategy while NVIDIA continued to expand its automotive partnerships [25][26] Group 4: Technological Advancements - In 2018, NVIDIA introduced the DRIVE Xavier platform, which improved computational performance while reducing power consumption [28] - Tesla's HW3, launched in April 2019, was described by Musk as the most advanced computer designed specifically for autonomous driving, marking the end of NVIDIA's direct involvement in Tesla's autonomous driving hardware [30][32]