SCR-SELF-1482026-06-01约 25 分钟阅读

大模型改变人工智能工程范式_1_11_translate

The AI revolution: Large models change the AI engineering paradigm

大模型改变人工智能工程范式111

The AI revolution: Large models change the AI engineering paradigm

Shang Can Technology

In July, 2023

abstract

The emergence of artificial intelligence large models has changed the training mode of artificial intelligence, improved the starting point of training, broken through the shortcomings and bottleneck of the original small models, improved the level of intelligence, and brought a new paradigm of artificial intelligence engineering. Enterprises should realize the significance of this revolutionary breakthrough, explore the application of large models to solve enterprise problems, build the enterprise Foundation model, and then build the enterprise intelligent machine matrix, to lay the foundation for the realization of dual-intelligent collaborative business.

The main discovery

Large models are not only language models, but also include multi-modal large models, which can process video, audio, images, and structured data

Large model is a revolutionary breakthrough in the field of artificial intelligence, and its intelligence emergence and generalization ability make artificial intelligence have a high intelligence

Large models can be used not only for content generation, but also for data analysis, modeling and algorithm extraction

The use of large models for data analysis and modeling will completely change the existing data analysis mode and bring a new paradigm of artificial intelligence engineering

The application of the large model can realize the idea of the foundation model, and a variety of small models and expert models can be trained on top of the foundation model

Recommendations

Enterprises should realize the revolutionary significance of the large model and take the foundation model created by the large model as the "base" for the business of dual intelligent collaboration

Explore the value of large models to the business, and find application scenarios in the difficult problems faced by enterprises

When selecting a large model, enterprises should comprehensively consider the applicability, safety and input-output ratio of a large model

To improve the ability to control large models, management ability is an indispensable ability, and technical ability is a plus

foundation model is an idea in the field of artificial intelligence. It is hoped that one model can be the foundation of other models, so that other models can be trained on top of the foundation model. Such a scenario is not fully realized before the large model is produced. It is hard to say that the large model is the only option for the foundation model, but there is no doubt that the large model can be the foundation model. As we have analyzed, a large model refers to the large number of model parameters (feature values of connections between neurons in the neural network). The generally agreed standard is more than 3 billion. At present, many Chinese enterprises say that the number of parameters is more than 1 billion, which can be regarded as a large model. In fact, the model is a relative concept. When the output accuracy of the model is higher than the random value, it indicates that the large model has appeared the emergent ability. The number of parameters at this point is the critical point. When the accuracy reaches the reliability requirements, the large model is available, and the number of parameters is large enough for the required task.

Large models were developed because of the need of AIGC, because machines have the ability to master a lot of knowledge through learning. When the amount of knowledge learned is not large enough, the model does not have enough intelligence to produce meaningful content, so a lot of training materials are needed. With the increase of training materials, the computational speed is faster and faster, and the model becomes more and more large, eventually leading to the generation of large models. Because large models are born to realize AIGC, and because most of the recent large models are mainly used for AIGC tasks, people mistakenly think that large models can only be used for content generation, while ignoring another more important value area, namely the value of the foundation model. The value of this field makes it possible to completely transform the paradigm of artificial intelligence engineering, from "training the model from a kid" to " from a college students ".

Augmented data analytics, a meaningful but less successful attempt at artificial intelligence

Since the concept of big data began to become popular, enterprises have established big data management system and big data analysis system, hoping to find the correlation and law between data through historical data and current data, and use these findings for the analysis of the current situation and the prediction of the future. However, before the Augmented data analysis technology, the design of the data analysis model mainly depends on the model designer's understanding of the business and the manual design. Manual design mode requires designers to have enough understanding of the business, so the designer's business ability and design ability will become the limitation of the data analysis model. On the one hand, the efficiency of manual design is very low, so it is impossible to complete the models required for the business in a relatively short time, and the total amount of models that can be completed manually is very limited. On the other hand, because the designer's understanding of the business is insufficient, the structure of the design cannot fully reflect the law of the business, forming a design bias. These things improved after machine learning was introduced into big data analytics. The automatic discovery of data (Pattern) through machine learning can greatly improve the efficiency of model generation, and break through the cognitive limitations of designers, so that the quantity and quality of analytical models can be improved to a greater extent. If the manual design analysis model is used, a designer can design a model in a year is also very limited. Using Augmented data analysis technology produces thousands of analytical models within minutes. Of course, it will take some time to verify the validity of these models, but the efficiency has been greatly improved.

According to the survey, a considerable number of enterprises have adopted augmenteddata analysis technology. However, the overall application effect is not ideal. There are two main reasons. One is that the technology of augmented data analysis is not well mastered and can not effectively play its effectiveness. Second, the augmented data analysis itself also has defects, because the machine can only find correlation, not causality. At the same time, the focus of machine learning is to discover data analysis models, rather than improving the capabilities of machine learning models themselves. In some application scenarios, such as the application of various intelligent brain, although improving the ability of the intelligent brain is the key, after long-term training, it is found that when the ability of the intelligent brain is improved to a certain extent, it seems to reach the ceiling of the ability, no matter how the training can not be improved. For example, the driverless system cannot overcome the difficulties of complex traffic environment, and the error rate of the language translation system kept at about 20% cannot be further reduced. In the medical field, the recognition of medical images has also hit the ceiling, always not at the level of human experts. In the f

登录后查看全文

本报告免费开放给注册用户,登录即可阅读全文。

相关报告推荐

2282682026-11-12

工程师的能力再养成:从会写到会定义、会审验、会编排智能体的转型路径研究

工程师的核心价值正从代码实现向业务定义与智能体编排转移。在双智协同框架下,研发人员的个人能力模型需要全面升级。传统的代码编写只是基础,现代工程师必须掌握业务定义、智能体输出审验以及通过业务编排脚本调度数字员工的能力。企业在推进新双模架构时,必须同步重塑人才结构。技术团队需深入理解业务逻辑,提升AI应用力与协同创新力,实现从单一开发向人机协同的跨越。这种能力再养成并非单纯的技术培训,而是组织数字化成熟度跃迁的基础工程,直接决定企业智能转型的最终成效。

7856152026-11-12

软件过程标准与DevOps流水线的AI化改造:评审、门禁与发布流程为智能体重写研究

DevOps流水线的AI化改造正推动软件过程标准从静态规则向液态敏捷演进。传统研发门禁依赖人工评审与固定脚本,难以适应高频发布需求。引入智能体重写评审与发布流程,是践行新双模架构的有效路径。通过将过程标准转化为业务编排脚本,企业能够构建人机协同的研发环境。这种改造提升了代码质量管控效率,实现软件交付的自适应调节。管理层需关注研发团队的数字化能力转型,确保人类工程师与AI智能体在复杂业务场景中高效协同,达成软件生产过程的智能化升级。

5920772026-11-11

瓶颈转移下的排期重构:编码不再是关键路径后项目缓冲与里程碑重排研究

软件研发瓶颈已从代码编写向需求转化与业务验证转移。在双智协同环境下,项目关键路径发生位移,传统以编码时长为基准的排期逻辑失效。企业需重新识别研发流程中的真实约束点,将项目缓冲区和里程碑向需求确认、架构设计以及业务编排脚本测试环节前移。管理者应依托新双模架构调整资源分配,将团队能力评估重心从单一代码产出转向业务理解与系统编排能力。组织需动态更新项目管理标准,以匹配智能协作环境下的真实交付节奏,确保研发效能的整体提升。

5328592026-11-11

敏捷仪式的去留:站会、迭代与看板在AI分钟级产出下的保留与重设研究

传统以信息同步为核心的敏捷仪式在AI分钟级产出下正面临失效,亟需向双智协同模式重设。每日站会与固定看板在高频交付中易沦为形式,团队应将精力从机械汇报转向业务编排脚本设计与复杂异常决策。迭代周期需依据机器产出特性,从固定时间盒转向事件驱动,以匹配非线性的开发节奏。组织需构建新双模架构,让人类智能与机器智能在目标对齐与逻辑校验层面深度协作。敏捷仪式的去留不应拘泥于原有形态,而应重设为驱动人机协作的价值交付节点,确保业务逻辑的准确性与战略方向的一致性。

3455112026-09-16

面向液冷通道密闭环境的物理安防设备空间侵入式监测适配模式研究

本研究提出一种适配液冷通道密闭环境的物理安防设备空间侵入式监测新范式,核心在于突破传统安防部署对开放空间与可见光条件的依赖,转向以环境约束为设计原点的主动适配逻辑。针对液冷通道高密度、全封闭、强电磁屏蔽及温湿度动态变化等典型特征,研究构建了“感知—响应—验证”三级协同机制:通过多模态微扰信号融合识别非授权空间扰动,利用通道结构特性增强信号可辨识度,并引入轻量级边缘推理实现低延迟本地决策。该模式不依赖外部视觉覆盖或高功耗持续扫描,在保障监测连续性的同时显著降低系统与基础设施的耦合风险。实践表明,其在有限安装空间、受限供电及复杂热流场中仍能维持稳定感知效能,为新型数据中心、高性能计算设施等高约束场

3261012026-09-14

异构算力混布机房中CPU/GPU/DPU热密度梯度分布建模与散热冗余度评估框架

本报告提出一套面向异构算力混布机房的热密度梯度建模与散热冗余度评估框架,核心观点是:传统均质化散热设计难以适配CPU、GPU、DPU等多类型芯片在空间分布、功耗动态性及局部热强度上的显著差异,必须建立与设备物理布局、负载时序特征和散热路径耦合的梯度化热模型。框架以热流守恒与传热边界条件为理论基础,将机房划分为多尺度热域,通过耦合设备瞬态功耗曲线与气流组织仿真,量化不同区域的热密度时空梯度;进而定义散热冗余度指标,综合反映制冷系统在局部热点、负载突变及单点故障场景下的动态承载裕量。该方法避免依赖静态峰值功耗假设,转而强调热响应的结构性瓶颈识别——例如高密GPU区与邻近DPU通信节点间的热串扰、冷

9859542026-09-14

超算异构混布机房中GPU加速卡在液冷约束下的峰值功耗释放能力评估框架研究:聚焦冷板流速-入口温度-芯片结温三变量协同限界机制

本研究提出一种面向超算异构混布机房的GPU加速卡功耗释放能力评估框架,核心观点是:在液冷约束下,GPU的实际峰值功耗并非由供电或芯片规格单方面决定,而是受冷板流速、冷却液入口温度与芯片结温三者动态耦合所共同限界。该框架摒弃传统“功耗—散热”线性映射思路,转而构建三变量协同作用下的热力学可行域模型,揭示不同工况组合对功耗释放的非线性抑制机制。研究发现,微小的流速波动或入口温度偏移,在高负载持续运行时可能触发结温快速逼近安全阈值,从而迫使系统主动降频限功——这种限界效应在多类型GPU混布、变负载场景中尤为显著。框架支持在机房规划、液冷系统调优及任务调度策略制定阶段,量化评估功耗释放潜力边界,避免因

2551742026-09-14

面向智算中心集群的算电协同动态响应机制研究:基于负荷可调性与电网调节信号的双向耦合建模

本研究提出一种面向智算中心集群的算电协同动态响应机制,核心在于打破计算负载与电力供应之间的单向适配惯性,构建负荷可调性与电网调节信号的双向耦合关系。通过将智算任务调度、资源弹性伸缩与电网频率、电压、负荷指令等实时运行信号统一建模,机制实现了计算侧对电力系统动态变化的主动感知与快速响应,同时支撑电网在波动场景下获得可观、可测、可控的柔性调节能力。研究强调“算力即调节资源”的系统观,将传统视为刚性负载的智算集群转化为具备时间维度灵活性和功率维度可塑性的协同单元。该机制不依赖特定硬件架构或封闭生态,兼容主流虚拟化与编排框架,可在现有基础设施上分阶段部署。实证表明,其在保障关键算力服务SLA前提下,显