如果你是一家企业，想要在打败竞争对手的同时，更强大地度过经济衰退，那么开源并不是答案云，也不是。的确，两者都有帮助。两者都是企业应该如何重新思考其传统的IT方法的组成部分。但这两者都不会让你脱颖而出。
为什么？因为其他人也已经在使用开源和云计算了。曾经有一段时间，第一个拥抱开源项目(如Linux或MySQL)的经济可能会让一家公司脱颖而出，但现在不一样了。企业对云的采用仍处于起步阶段(据Gartner估计，到2022年，云将占到所有IT支出的10%左右)，但采用的速度如此之快，以至于你可能无法仅通过云来区分你的客户体验。什么能让你与众不同？
机器学习和人工智能。但也许不是你想的那样。
一、逐步思考人工智能
这不是那些吹捧AI/ML是某种定义不清的灵丹妙药的文章之一。是的，人工智能和ML在开发对抗COVID-19的强效药物方面发挥了重要作用，它们甚至有一天可以帮助找到治疗癌症的方法。但是，没有什么神奇的AI/ML肥料可以浇灌奄奄一息的IT项目，它们就会神奇地开花结果。谷歌或Uber等公司一直是AI/ML的先锋，但让我们面对现实吧：你没有他们的工程人才。
就连这些公司也在利用经济低迷减少登月计划的时间，把更多的时间花在渐进式进步上，正如《华尔街日报》最近的一篇文章(《大型科技公司停止做傻事》)所呼吁的那样：科技行业“长期以来一直致力于颠覆，现在则专注于增强已经存在的东西。”文章指出，“2023年最好的科技投资可能是那些愿意把钱花在(轮子上的)润滑上的公司，而不是重新发明轮子。”
企业做到这一点的一个重要方法是AI/ML，但不是神奇的飞行汽车。AI/ML正以更普通(也更有用)的方式被使用。
Zillow花了数年时间试图使用AI/ML模型在炒房上做大。然而，在2021年底，该公司退出了这项业务，理由是尽管有复杂的模型，但无法预测价格。相反，Zillow变得务实起来，正在使用AI/ML帮助潜在租客在城市中行走时查看房源，并使房东能够根据这些公寓的照片构建楼层平面图。远没有十亿美元的炒房生意那么吸引人，但对客户来说却有用得多。
就谷歌而言，它已经开始通过分析视频数据为零售商提供跟踪库存的功能。谷歌用超过10亿张产品图片的数据集训练模型。它可以识别来自手机或店内摄像头的图像数据。如果真如宣传的那样有效，对于传统上难以处理库存的零售商来说，这将是一个巨大的福音。这不是AI/ML的性感应用，但对零售客户很有用。
作为AI/ML领域的领导者，微软刚刚对OpenAI进行了巨大的投资，据报道，微软打算将gpt式的功能引入其生产力应用程序，如Word或Outlook。微软有足够的资源对Office进行大改造，或许可以让它完全由语音驱动。相反，它很可能会给Office一个严肃的Clippy升级，使用GitHub Copilot之类的方法。也就是说，GPT可能会取代编写文档或构建电子表格的一些无差别的繁重工作。少些性感，多些实用。
二、选择在AI上不失败
事实证明，增量方法是使用AI/ML进行构建的最聪明的方法。正如AWS无服务器英雄Ben Kehoe所说，“当人们想象将人工智能集成到软件开发(或任何其他过程)时，他们往往过于乐观。”他强调，一个关键的失败在于，相信AI/ML有思考的潜力，但没有足够的能力完全信任其结果:“我看到的许多AI都断言，AI将能够为一个人承担给定任务的全部责任，并隐含地假设这个人对任务的责任将会……消失？”
在现实世界中，开发人员(或其他人)必须对结果负责。例如，如果你正在使用GitHub Copilot，你仍然要对代码负责，无论它是如何编写的。如果代码最终出现bug，责怪AI是没有用的。拿着工资单的人将承担责任，如果他们无法验证他们是如何得出结果的，那么，他们很可能会在放弃工作之前放弃人工智能模型。
这并不是说AI和ML在软件开发或企业的其他领域没有一席之地。看看Zillow、谷歌和微软的例子就知道了。诀窍是使用AI/ML来补充人类智能，并允许同样的人类智能来检查结果。正如Kehoe所建议的，“当看到人工智能将自动化某些流程时，要寻找该流程真正困难的固有复杂性是什么，以及如果(通过黑箱人工智能)在复杂性中注入大量(新的)不确定性，该流程是否会成功。”
增加不确定性和加大问责难度是行不通的。相反，企业将寻找允许机器承担更多责任的领域，同时仍让相关人员对结果负责。这将是企业IT的下一个大事件，因为它将是许多小的、增量的东西。
原文
If you’re an enterprise looking for ways to come through a recession stronger while beating out competitors in the process, open source isn’t the answer. Neither is cloud. It’s true that both can be helpful. Both are ingredients in how enterprises should rethink their traditional approaches to IT. But neither will do much to distinguish you.
Why? Because everyone else is already using open source and cloud, too. There was a time when being first to embrace the economics of open source projects like Linux or MySQL could set a company apart, but not anymore. Enterprise adoption of cloud is still nascent (roughly 10% of all IT spending in 2022, per Gartner estimates), but adoption is moving at such a pace that you’re probably not going to distinguish your customer experience through cloud alone. What will set you apart?
Machine learning (ML) and artificial intelligence (AI). But maybe not how you think.
Thinking incrementally about AI
This is not one of those articles touting AI/ML as some ill-defined panacea.
This is not one of those articles touting AI/ML as some ill-defined panacea. Yes, AI and ML have been instrumental in developing potent medicines to combat COVID-19, and they could even someday help find a cure for cancer. But there’s no magical AI/ML fertilizer that you pour onto moribund IT projects and they magically blossom. Companies like Google or Uber have been on the vanguard of AI/ML, but let’s face it: You don’t have their engineering talent.
Even these companies are using the downturn to spend less time on moon shots and more time on incremental advances, as a recent article in The Wall Street Journal (“Big Tech Stops Doing Stupid Stuff“) calls out: The tech sector “that has long worked to disrupt is now focusing on enhancing what already exists.” Instead of reinventing wheels, the article notes, “The best tech investments of 2023 might be companies content to spend their coin greasing [the wheel].”
One big way enterprises are doing this is with AI/ML, but not with gee-whiz flying cars. AI/ML is being used in far more pedestrian (and useful) ways.
Zillow spent years trying to use AI/ML models to go big on flipping houses. In late 2021, however, the company exited that business, citing an inability to forecast prices despite sophisticated models. Instead, Zillow has turned pragmatic and is using AI/ML to help would-be renters see listings as they walk a city and enabling landlords to construct floorplans from photos of those apartments. Much less sexy than a billion-dollar house-flipping business, and much more useful for customers.
Google, for its part, has started offering retailers the ability to track store inventory by analyzing video data. Google trained its models on a data set of more than one billion product images. It can recognize the image data whether it comes from a mobile phone or an in-store camera. If it works as advertised, it would be a significant boon for retailers that traditionally have struggled to get a handle on inventory. Not a sexy use of AI/ML, but useful for retail customers.
Microsoft, a leader in AI/ML, just made a huge investment in OpenAI, with the reported intention of bringing GPT-esque functionality to its productivity apps, such as Word or Outlook. Microsoft has the resources to bet big on a moon shot makeover of Office, perhaps making it entirely voice driven. Instead, it’s likely going to give Office a serious Clippy upgrade with a GitHub Copilot sort of approach. That is, GPT might take over some of the undifferentiated heavy lifting of writing docs or building spreadsheets. Less sexy, more useful.
Choosing not to fail with AI
The incremental approach turns out to be the smartest way to build with AI/ML.
The incremental approach turns out to be the smartest way to build with AI/ML. As AWS Serverless Hero Ben Kehoe argues, “When people imagine integrating AI … into software development (or any other process), they tend to be overly optimistic.” A key failing, he stresses, is belief in AI/ML’s potential to think without a commensurate ability to fully trust its results: “A lot of the AI takes I see assert that AI will be able to assume the entire responsibility for a given task for a person, and implicitly assume that the person’s accountability for the task will just sort of … evaporate?”
In the real world, developers (or others) have to take responsibility for outcomes.
In the real world, developers (or others) have to take responsibility for outcomes. If you’re using GitHub Copilot, for example, you’re still responsible for the code, no matter how it was written. If the code ends up buggy, it won’t work to blame the AI. The person with the paystub will bear the blame, and if they can’t verify how they arrived at a result, well, they’re likely to scrap the AI model before they’ll give up their job.
This is not to say that AI and ML don’t have a place in software development or other areas of the enterprise. Just look at the examples from Zillow, Google, and Microsoft. The trick is to use AI/ML to complement human intelligence and allow that same human intelligence to fact-check results. As Kehoe suggests, “When looking at claims AI is going to automate some process, look for what the really hard, inherent complexity of that process is, and whether the process would be successful if a large degree of (new) uncertainty [through black-box AI] was injected into that complexity.”
Adding uncertainty and making accountability harder is a non-starter. Instead, enterprises will look for areas that allow machines to take on more responsibility while still leaving the people involved accountable for the results. This will be the next big thing in enterprise IT, precisely because it will be lots of small, incremental things.
本文由数字化转型网（www.szhzxw.cn）翻译而成，作者：Matt Asay；翻译：数字化转型网郑亚茹；翻译审核：数字化转型网默然。
免责声明: 本网站(http://www.szhzxw.cn/)内容主要来自原创、合作媒体供稿和第三方投稿，凡在本网站出现的信息，均仅供参考。本网站将尽力确保所提供信息的准确性及可靠性，但不保证有关资料的准确性及可靠性，读者在使用前请进一步核实，并对任何自主决定的行为负责。本网站对有关资料所引致的错误、不确或遗漏，概不负任何法律责任。
本网站刊载的所有内容(包括但不仅限文字、图片、LOGO、音频、视频、软件、程序等) 版权归原作者所有。任何单位或个人认为本网站中的内容可能涉嫌侵犯其知识产权或存在不实内容时，请及时通知本站，予以删除。
