AI is beyond the hype and entering a phase of industrialization. The belief about the imperative of AI is widespread— universally across industries. However, only about 5% of global enterprises is ready for industrialized growth of AI (according to this study). To reap the real benefits, organizations need to be able to scale AI solutions. This all sounds obvious: everybody talks about AI and scaling is the name of the game.

人工智能已经超越了炒作，进入了工业化阶段。 关于AI必不可少的信念在各个行业中普遍存在。 但是，只有大约5％的全球企业准备好AI的工业化增长(根据本研究 )。 为了获得真正的利益，组织需要能够扩展AI解决方案。 这听起来很明显：每个人都在谈论AI，而缩放就是游戏的名称。

But what does scaling AI mean exactly? And what requirements does that pose to an organization’s data and technology stack? How is that different from most legacy infrastructures? And how do you build the technical capabilities to meet these new requirements?

但是扩展AI到底是什么意思？ 对组织的数据和技术堆栈有什么要求？ 与大多数传统基础架构有何不同？ 您如何建立技术能力来满足这些新要求？

Reflecting on my own experience supporting large organizations on their AI journey, I will try to shine some light on these questions.

回顾我自己在支持大型组织的AI历程中的经验，我将尝试阐明这些问题。

扩展AI (Scaling AI)

The adoption of AI exhibits the same patterns as any other big technology revolution of the past. Widespread electrification took 30–40 years. AI will be no different. According to Amara’s law, the impact of a new innovation is typically overestimated in the short term, but underestimated in the long term. The underlying reason is that the context in which technology is applied needs to change before its real potential can emerge. For AI, this phenomenon is referred to as the AI productivity paradox. It is deceptively easy to build a proof of concept for AI and show all its amazing feats. But implementing and scaling AI in operational processes throughout an enterprise is a whole different ball game. Let’s look at what scaling AI means exactly.

人工智能的采用与过去的其他任何重大技术革命都呈现出相同的模式。 广泛的电气化历时30-40年。 AI不会有什么不同。 根据阿玛拉定律，通常在短期内高估一项新创新的影响，但从长期来看却低估了这一影响。 根本原因是，在真正的潜力出现之前，需要改变应用技术的环境。 对于AI，这种现象称为AI生产率悖论。 构建AI的概念证明并展示其所有惊人的壮举似乎很容易。 但是，在整个企业的运营流程中实施和扩展AI完全是一回事。 让我们看看扩展AI的确切含义。

ScaleConsider a global chemicals producer that uses sensor data from machinery to predict failures and apply preventive maintenance measures. Pumps or compressors are a typical example. They have thousands of them on different plants across the world. There are different types of compressors, with different failure modes and corresponding predictive data patterns. Developing an AI solution for this type of opportunity would start with a machine learning model for one specific type of equipment and testing different machine learning techniques. Scaling then means rolling out these models across the globe for the same equipment and subsequently repeating these steps for each different type of pump or compressor. You could even automate these steps by building a system that automatically trains, implements and scales models for different types of equipment. Including the shadow models, this could amount to hundreds of thousands of algorithms all running in production.

规模考虑一家全球化学品生产商，该生产商使用来自机器的传感器数据来预测故障并采取预防性维护措施。 泵或压缩机是一个典型的例子。 他们在世界各地的不同工厂中拥有成千上万的产品。 有不同类型的压缩机，具有不同的故障模式和相应的预测数据模式。 为此类机会开发AI解决方案将从针对一种特定类型设备的机器学习模型开始，并测试不同的机器学习技术。 然后，扩展意味着在全球范围内针对同一设备推出这些模型，并随后针对每种不同类型的泵或压缩机重复这些步骤。 您甚至可以通过构建一个系统来自动执行这些步骤，该系统可以针对不同类型的设备自动训练，实施和缩放模型。 包括阴影模型在内，这可能等于数十万种算法都在生产中运行。

ScopeAnother form of scaling AI refers to the pervasive use of AI by different teams, departments and for different use cases. Deploying AI at scale then means that anyone in an organization uses machine learning to optimize work processes — a.k.a. the democratization of AI. Although impact from AI is not immune to the Pareto principle — 80% of the value will sit with 20% of the potential use cases— eventually the use of machine learning will show up everywhere. Allowing this to happen obviously requires some standardized and structural capabilities.

范围扩展AI的另一种形式是指不同团队，部门和不同用例对AI的普遍使用。 大规模部署AI意味着组织中的任何人都可以使用机器学习来优化工作流程，也就是AI的民主化。 尽管来自AI的影响并非不受Pareto原则的影响-价值的80％将占潜在用例的20％-最终，机器学习的使用将无处不在。 允许发生这种情况显然需要一些标准化和结构化的功能。

SpeedPerhaps the most trivial form of scalability revolves around infrastructural resources. Depending on the type of AI solution, easily scalable storage, memory or computational resources are required. Training a deep neural net for instance can require significant memory and compute power. Running an algorithm to produce a prediction, also called inference, typically requires less computational power than training but at the same time needs to provide low latency. That is, if you want to produce a next-product-to-buy recommendation to an online browsing customer, that has to happen in the order of milliseconds. When your website engages with millions of customers at a time (think Amazon), each of these interactions trigger inferences that have to be computed simultaneously and virtually instantly.

速度可能最简单的可扩展性形式围绕基础设施资源。 根据AI解决方案的类型，需要轻松扩展的存储，内存或计算资源。 例如，训练深层神经网络可能需要大量的内存和计算能力。 运行算法以产生预测(也称为推断)的算法通常比训练需要更少的计算能力，但同时需要提供低延迟。 也就是说，如果您想为在线浏览客户提供“下一个要购买的产品”建议，则必须以毫秒为单位。 当您的网站一次与数百万的客户互动时(例如Amazon)，这些交互中的每一个都会触发推理，这些推理必须同时且几乎立即进行计算。

Front running organizations that are able to industrialize AI at scale do the following three things exceptionally well:

能够大规模实现人工智能产业化的领先组织出色地完成了以下三件事：

• They set direction and follow course. AI adoption is a journey that is completed step-by-step. It requires vision, strategy and a game plan that combines quick wins with a route to scale. Sustained commitment from senior leadership is key to long term success. You have to go all in, but with smart focus and grit.

  他们确定方向并遵循路线 。 人工智能的采用是一个逐步完成的旅程。 它需要远见卓识，战略和博弈计划，这些计划要结合快速获胜和扩展路线。 高层领导的持续承诺是取得长期成功的关键。 您必须全力以赴，但要有聪明的专注力和毅力。

• They have a playbook for AI solution development and implementation. AI-powered business innovation follows a typical life-cycle from idea or proof of concept to a tested prototype, MVP and eventually production-grade solution that is implemented. A proven approach to doing this repeatedly enables to gradually develop and scale new AI opportunity areas and become a truly AI-powered organization.

  他们有一本关于AI解决方案开发和实施的手册。 人工智能驱动的业务创新遵循典型的生命周期，从概念或概念证明到经过测试的原型，MVP以及最终实现的生产级解决方案。 经过反复验证的成熟方法可以逐步开发和扩展新的AI机会领域，并成为真正的AI驱动的组织。

• They have a technology platform that supports AI at scale. AI solutions pose new requirements for technology architecture. Companies that manage to scale AI have developed standardized platforms that allow rapid development of AI solutions in a robust and sustainable manner.

  他们拥有一个可大规模支持AI的技术平台。 人工智能解决方案对技术架构提出了新要求。 设法扩展AI的公司已经开发了标准化平台，可以以健壮和可持续的方式快速开发AI解决方案。

The next sections will explore five principles for building a technology platform that allows to scale AI solutions.

下一部分将探讨构建可扩展AI解决方案的技术平台的五项原则。

建立可扩展AI平台的五项原则 (Five principles to build a scalable AI platform)

On the highest level, the scalability of AI requires platform capabilities that combine two previously separated domains in data architecture: the operational or transactional side of data and the analytical use of data for analysis or modeling. An AI platform forms the bridge: it consumes data to produce trained machine learning algorithms which are subsequently deployed as production-grade services in the operational domain — to be used by applications and users for decision making. As simple as that. Below are five principles to build a scalable AI stack.

在最高层次上，AI的可扩展性要求平台功能必须结合数据体系结构中两个先前分离的域：数据的操作或事务方面以及对数据进行分析或建模的分析用途。 一个AI平台构成了桥梁：它使用数据来产生训练有素的机器学习算法，然后将其部署为操作域中的生产级服务，供应用程序和用户用于决策。 就如此容易。 以下是构建可扩展AI堆栈的五项原则。

1.作为微服务的算法 (1. Algorithms as micro-services)

The classical way to put a machine learning model into production is to build a pipeline that ingests the latest input data, runs the model with that input, and stores the output in a database. The application(s) using the predictions can subsequently fetch the data from this database. Although there is nothing inherently wrong with this approach, it is not very scalable. Modern ML use cases require predictions on demand, potentially with thousands of requests every second. So the prediction pipeline cannot be scheduled in advance. Moreover, it needs to be able to scale rapidly depending on the workload.

将机器学习模型投入生产的经典方法是建立一个管道，以吸收最新的输入数据，使用该输入运行模型，并将输出存储在数据库中。 使用预测的应用程序可以随后从该数据库中获取数据。 尽管此方法没有内在的错误，但它的可伸缩性不是很高。 现代ML用例需要按需预测，每秒可能有数千个请求。 因此，无法预先安排预测管道。 此外，它需要能够根据工作负载快速扩展。

The emerging best practice for deploying ML models is to package them — along with an API built on top — as a micro-service with containers (such as Docker). Containerization is the latest paradigm for application development with many benefits that also apply to AI applications or cognitive services:

部署ML模型的新兴最佳实践是将它们以及基于顶部构建的API打包为带有容器的微服务(例如Docker)。 容器化是应用程序开发的最新范例，具有许多优点，这些优点也适用于AI应用程序或认知服务：

• Re-usability and loose coupling. Deploying the ML model as a container allows any other application to leverage this prediction service without any dependencies. The model becomes self-contained. It becomes quite easy to build multiple AI systems based on loosely coupled micro-services, sharing the same algorithms. Predictions are done on demand — hence the term cognitive service.

  可重复使用和松耦合 。 将ML模型部署为容器后，任何其他应用程序都可以利用此预测服务而无需任何依赖关系。 该模型成为独立的。 建立基于松耦合微服务，共享相同算法的多个AI系统变得非常容易。 预测是按需完成的，因此称为认知服务 。

• Auto-scaling & scheduling. Once the model is deployed as a container, the required resources (e.g. computation) can be scaled in line with workload with container orchestration services such as Kubernetes, AWS Elastic Beanstalk or Azure App Services.

  自动缩放和计划。 一旦将模型部署为容器，就可以通过容器编排服务(例如Kubernetes，AWS Elastic Beanstalk或Azure App Services)根据工作负载扩展所需的资源(例如，计算)。

• Portability. Since the algorithm is packaged as a generic container that can be executed anywhere, the use of algorithms in operational processes is not dependent on specific tooling anymore. Moreover, containers are platform-agnostic. They can move easily from one cloud to the other and operate in a highly distributed manner. This makes it easy to expose algorithms close to where the data is (e.g. in edge applications).

  可移植性。 由于算法被打包为可以在任何地方执行的通用容器，因此在操作流程中使用算法不再依赖于特定的工具。 而且，容器与平台无关。 它们可以轻松地从一个云迁移到另一个云，并以高度分布式的方式运行。 这样可以很容易地将算法暴露在靠近数据的位置(例如，在边缘应用程序中)。

Combined, containerization provides a highly standardized, flexible and scalable approach to putting algorithms into production.

结合起来，容器化提供了高度标准化，灵活和可扩展的方法，可将算法投入生产。

Consider for instance a financial services company that developed a Probability of Default (PoD) model. This algorithm is leveraged for the underwriting process for new customers. But it can also be applied to existing customers as part of a periodic asset risk evaluation. The model is operationalized as a stand-alone service that can be called to serve predictions based on different input data. As such it has become a very scalable cognitive service to be used across different applications and processes.

例如，考虑一家开发了违约概率(PoD)模型的金融服务公司。 该算法可用于新客户的承保流程。 但是，它也可以作为定期资产风险评估的一部分应用于现有客户。 该模型可作为独立服务运行，可以调用该服务基于不同的输入数据提供预测。 这样，它已成为一种可扩展性强的认知服务 ，可以在不同的应用程序和过程中使用。

2.一种工厂方法来构建和管理算法 (2. A factory approach to building and managing algorithms)

Producing algorithms as micro-services requires an assembly line with proper quality management and control mechanisms. Sort of a six sigma methodology for the production of algorithms:

生产作为微服务的算法需要具有适当质量管理和控制机制的装配线。 用于产生算法的一种六西格玛方法：

• Standardized & automated workflow. Instead of building a specific workflow for each individual use case, a standardized workflow (assembly line) is needed to enable scalable production of algorithms. Consider that models are not developed one off, but require ongoing maintenance and retraining. Hence, doing this at scale repeatedly for thousands of models requires a fully standardized and automated assembly line that can be replicated for any new AI solution. The assembly line fetches source data, transforms it into a format for machine learning training, develops a model, sets up a validation set, tests model performance on the test set and finally deploys the model as a container.

  标准化和自动化的工作流程 。 代替为每个单独的用例构建特定的工作流，需要标准化的工作流(装配线)以实现可扩展的算法生产。 考虑到模型不是一次性开发的，而是需要持续的维护和再培训。 因此，要针对数千个模型重复进行大规模操作，就需要一条完全标准化和自动化的装配线，该装配线可以复制到任何新的AI解决方案中。 组装线获取源数据，将其转换为用于机器学习训练的格式，开发模型，建立验证集，在测试集上测试模型性能，最后将模型部署为容器。

• Performance monitoring & inventions. Six sigma aimed at reducing variance in the manufacturing process: the upper and lower limits of quality have to lie beyond 6 times the standard deviation, implying a defect rate of 3.4 in a million. Similarly, model predictions will fall within a bandwidth. Monitoring this out of sample performance, how the model performs in practice rather than on historic data, is critical to keep algorithms in check. This includes for instance improving prediction power, reducing biases and triggering retraining to mitigate concept drift.

  性能监控与发明。 六 sigma旨在减少制造过程中的差异：质量的上限和下限必须超过标准偏差的6倍，这意味着百万分之3.4的缺陷率。 同样，模型预测将落入带宽之内。 监视样本外性能，模型在实践中的性能而不是对历史数据的性能，对于控制算法至关重要。 例如，这包括提高预测能力，减少偏差并触发重新训练以减轻概念漂移。

• Traceability. In Europe, new legislation requires organizations to be able to explain algorithmic decisions if they impact the customer. Responsible AI is becoming a hot topic. Therefore, the ability to trace back model versions, including the accompanying training data, is becoming a sine qua non. Like equipment and manufacturing parts have serial numbers, algorithms require a transparent and traceable production flow. The model assembly line therefore needs to store all relevant artifacts (inputs, outputs and scraps) in an organized manner.

  可追溯性。 在欧洲，新法规要求组织在影响客户时能够解释算法决策。 负责任的AI成为热门话题。 因此，追溯模型版本(包括随附的训练数据)的能力已成为必要条件。 就像设备和制造零件具有序列号一样，算法也需要透明且可追溯的生产流程。 因此，模型装配线需要以有组织的方式存储所有相关的工件(输入，输出和报废)。

The step towards a standardized assembly line for model production is a big challenge for many organizations that aspire to using AI at scale. Although there is huge industry innovation on this front and many components are available as standardized (cloud native) services — e.g. Sagemaker for AWS, or Azure ML) — still quite a bit of specialist engineering is required to fit all the pieces together.

对于许多渴望大规模使用AI的组织而言，迈向标准的装配线以进行模型生产是一大挑战。 尽管在这方面进行了巨大的行业创新，并且许多组件都可以作为标准化的(云原生)服务(例如，AWS的Sagemaker或Azure ML)提供，但仍需要大量的专业技术来将所有组件组合在一起。

3.多种数据整合途径 (3. Multiple data integration routes)

The previous two principles are fully new and inherent to the arrival of AI. The third principle relates to data management and is often a source of struggle with legacy environments.

前两个原则是全新的，是AI到来的固有原则。 第三个原则与数据管理有关，通常是与遗留环境斗争的根源。

Innovation with AI requires more flexibility in data integration than most data architectures traditionally allow. Enterprise data warehouses are set up to provide a single source of truth on which data products can be built. The high integrity and efficient re-use of data comes at the cost of rigidity and high upfront investment. This works well for stable use cases that can be defined well in advance and are based on structured and slowly moving data — such as enterprise reporting. AI solutions have broader requirements which lie between traditional BI products and business applications. As mentioned above, they marry the operational and analytical use of data. A more versatile approach to data management is therefore needed:

与AI相比，创新需要比传统上大多数数据架构更大的灵活性。 建立企业数据仓库是为了提供可用于构建数据产品的单一事实来源 。 高度的完整性和有效的数据重用是以牺牲刚性和大量前期投资为代价的。 这对于可以预先定义好的稳定用例非常有用，这些用例是基于结构化且缓慢移动的数据(例如企业报告)。 人工智能解决方案具有更广泛的要求，介于传统BI产品和业务应用程序之间。 如上所述，它们结合了数据的操作和分析用途。 因此，需要一种更加通用的数据管理方法：

• AI requires experimentation in production. After an MVP has gone live, the AI solution will enter a next life cycle stage of continuous refinement. This often leads to adding new features (data) to optimize performance. The long development cycles of DWHs do not match the agile and iterative nature of AI-driven innovation.

  人工智能需要在生产中进行实验 。 MVP上线后，AI解决方案将进入持续改进的下一个生命周期阶段。 这通常导致添加新功能(数据)以优化性能。 DWH的漫长的开发周期与AI驱动的创新的敏捷性和迭代性不符。

• AI requires big data. New innovations based on machine learning often come from using new data sources, potentially with higher volumes and unstructured data. Traditional data integration platforms do not cater to these requirements.

  人工智能需要大数据 。 基于机器学习的新创新通常来自使用新数据源，可能具有更高的数据量和非结构化数据。 传统的数据集成平台无法满足这些要求。

To facilitate rapid AI-powered business innovation, your AI platform needs to support data integration practices that strike the right balance between flexibility/innovation and robustness/stability. This can be achieved by allowing local data integration routes:

为了促进快速的AI驱动的业务创新，您的AI平台需要支持在灵活性/创新性与鲁棒性/稳定性之间取得适当平衡的数据集成实践。 这可以通过允许本地数据集成路由来实现：

• A data warehouse for core data only (ETL). The DWH concept is not dead, but needs to be re-framed from a one-size-fits-all solution at enterprise scale, to domain-based data integration for your most important data assets. It still makes sense to have a solid data warehouse for e.g. customer and product data, which is used as a single-source-of-truth by all information products.

  仅用于核心数据(ETL)的数据仓库。 DWH的概念还没有完结，但是需要从企业规模的“一刀切”解决方案重新架构，再到最重要的数据资产的基于域的数据集成。 拥有一个可靠的数据仓库来存储客户和产品数据仍然是有意义的，例如，所有信息产品都将其用作真实的单一来源。

• Direct data pipelines (ELT). A second route is needed for use case specific data which is not designated for DWH integration. Many AI solutions require a combination of core data, which can be fed from the DWH, and very specific data from multiple sources (e.g. click stream data). Instead of building a large DWH with pre-processed integration steps (ETL), the data pipeline for the specific solution in question handles the data integration after loading the data from multiple source locations (e.g. the DWH for core data and a data lake for other solution specific data).

  直接数据管道(ELT) 。 未针对DWH集成指定的用例特定数据需要第二条路径。 许多AI解决方案需要结合使用可从DWH馈送的核心数据，以及来自多个来源的非常具体的数据(例如，点击流数据)。 与使用预处理的集成步骤(ETL)构建大型DWH相比，所讨论的特定解决方案的数据管道在从多个源位置加载数据后处理数据集成(例如，用于核心数据的DWH和用于其他数据源的数据湖)解决方案特定数据)。

• Tiered data lake. To support solution-specific data pipelines, a data lake can be used for storing source data which contains different zones or tiers with varying degrees of data pre-processing. These zones feed different components of the integration layers, e.g. raw data feeding the DWH or development sandbox. A zone with curated data can be set up to feed solution-specific ELT data pipelines.

  分层数据湖。 为了支持特定于解决方案的数据管道，可以使用数据湖来存储源数据，该源数据包含具有不同程度或不同级别的数据预处理的不同区域或层。 这些区域提供了集成层的不同组件，例如，原始数据提供给DWH或开发沙箱。 可以设置包含策展数据的区域，以馈送特定于解决方案的ELT数据管道。

4.多模式数据交换：从批处理到基于事件 (4. Multi-pattern data exchange: from batch to event-based)

Another key feature of the data architecture for scalable AI deployment is the support for multiple data exchange patterns, in particular for the use of real-time data. This is where traditional data architecture for IT integration — connecting the application landscape — merges with modern use of data for analytical purposes (providing real-time BI or AI).

用于可扩展AI部署的数据体系结构的另一个关键功能是对多种数据交换模式的支持，尤其是对实时数据的使用。 这是用于IT集成的传统数据体系结构(连接应用程序环境)与现代数据使用合并以实现分析目的(提供实时BI或AI)的地方。

The dominant paradigm for real time data exchange is to use an event-driven architecture (EDA) based on e.g. a Kafka messaging platform. An EDA is ideal for deploying models as micro-services, as described above under #1, but it is a much broader concept that can serve as the back-bone of all data exchange. Any updates in source data are published as separate events that can be processed by data integration platforms and pipelines who listen to these events in real time. The AI platform connects to the EDA for sourcing data and in return deploys fully trained algorithms that can react to events by providing predictions on demand.

实时数据交换的主要范例是使用基于例如Kafka消息传递平台的事件驱动架构(EDA)。 如上所述，EDA非常适合将模型作为微服务部署，但是它是一个更广泛的概念，可以用作所有数据交换的基础。 在源数据的任何更新发布为可通过数据集成平台和管道谁听实时这些事件进行处理独立的事件。 AI平台连接到EDA以获取数据，并相应地部署经过全面训练的算法，可以通过提供按需预测来对事件做出React。

5.利用云组件进行敏捷开发 (5. Leverage cloud components for agile development)

The fifth principle is to apply agile development methods to your platform infrastructure. Legacy data architecture and infrastructure is characterized by monolithic, vendor-based systems that were meant to function as one-size-fits-all platforms. These infrastructures have been notoriously behind the curve — rapidly becoming outdated or even obsolete even before they were fully implemented. Never ending platform migrations with nirvana end-state views became the norm.

第五项原则是将敏捷开发方法应用于平台基础架构。 传统数据体系结构和基础结构的特点是基于卖方的整体式系统，旨在充当一种适用于所有平台的平台。 众所周知，这些基础架构落后于潮流-甚至在完全实施之前就Swift变得过时甚至过时。 永不停止使用必杀技的最终状态视图进行平台迁移已成为常态。

The modern approach to platform development is to build more nimble architectures using best-of-breed components. This is both necessary and possible.

平台开发的现代方法是使用同类最佳的组件构建更灵活的架构。 这既必要也是可能的。

It is necessary. For two reasons. First, fungibility. Building a platform that uses different technologies for different components prevents vendor lock-in. Second, agility. Technology has a tendency to develop faster than anyone can keep up to. Building a platform of different technology components prevents placing one big bet on a single provider or technology. It becomes easier to swap technologies and benefit from multiple innovation sources.

这是必要的 。 有两个原因。 首先，可替代性。 构建使用针对不同组件的不同技术的平台可防止供应商锁定。 第二，敏捷性。 技术的发展趋势是任何人都无法跟上的。 构建由不同技术组件组成的平台可防止对单个提供商或技术下大赌注。 交换技术变得容易，并可以从多种创新资源中受益。

It is possible. The combination of cloud native and open source technologies make it possible to stitch together different platform components. The cloud has created an ecosystem that provides seamless integration: it tends to just work. The use of standard lego blocks allow a DIY approach without creating (too much) technical debt and custom infrastructure that will have to be cleaned up later (and so the cycle continuous).

有可能的。 云原生技术和开源技术的结合使将不同的平台组件缝合在一起成为可能。 云创建了一个提供无缝集成的生态系统：它趋于正常运行。 使用标准的乐高积木可以进行DIY，而不会产生(过多)技术债务和自定义基础结构，而这些基础结构必须在以后进行清理(因此周期是连续的)。

如何建立AI平台 (How to go about building an AI platform)

Once you agree on the destination, it is time to think about the journey. The biggest pitfall in platform development is to approach it in splendid isolation by a dedicated platform team. Instead, apply the agile development principle #5 and leverage the requirements from front running use cases to innovate the platform. Referring to the picture below, aim to travel from 1a to 3e by progressing on both axes in a synchronized fashion: develop solutions and platform as a well staged act of synergy.

一旦您确定了目的地，就该考虑行程了。 平台开发中最大的陷阱是由专门的平台团队以出色的隔离度对其进行处理。 相反，应采用敏捷开发原则5，并利用前端运行用例的需求来创新平台。 参照下图，旨在以同步的方式在两个轴上前进，以从1a到3e：开发解决方案和平台，作为良好的协同作用。

Your vision and and strategy for AI needs to translate from high-level value pools to concrete AI solution designs. Subsequently, the most promising solutions go through a structured development cycle — from ideation all the way to optimized solutions at scale. In parallel, these first solutions pose requirements to the platform. That provides an opportunity to build new platform components in a very pragmatic way, immediately applying and testing new capabilities on actual solutions. It is a matter of tightly managing the trade-off between fast time-to-market (the y-axis) and building high platform standards (the x-axis).

您对AI的愿景和战略需要从高级价值池转换为具体的AI解决方案设计。 随后，最有前途的解决方案经历了结构化的开发周期-从构思一直到大规模的优化解决方案。 同时，这些第一个解决方案对平台提出了要求。 这提供了机会，可以以非常实用的方式构建新的平台组件，并立即在实际解决方案上应用和测试新功能。 这是紧密管理快速上市时间(y轴)和建立高平台标准(x轴)之间的权衡问题。

The alternatives are risky. It won’t be the first time that business teams end up building point solutions (1c) that are hard to scale, or that IT teams spend years building a platform that is hardly being used (3b).

替代品是有风险的。 这将不是业务团队第一次构建难以扩展的点解决方案(1c)，也不是IT团队花费数年时间构建难以使用的平台(3b)。

I’m curious to hear your views and experience with scaling AI!

我很想听听您对扩展AI的看法和经验！

Wouter HuygenManaging Partner at MIcompany

Wouter Huygen MIcompany总经理