闪亮的炒作火车 (The shiny hype train)

Ever since the overworked and cringeworthy remark that data science is the sexiest job of the 21st century, and the resulting hype train that torpedoed the modern conception of Machine Learning — Breiman’s conception — from the academic fringes to the dizzying lights of the mainstream labour market, the frantic bum’s-rush to repackage Logistic Regressions as bleeding edge A.I. began in earnest. I’m looking at you, IBM Watson.

自从过度劳累和cringeworthy此言数据科学是 21世纪最性感的工作 ，而鱼雷机器学习的现代观念所产生的炒作列车- Breiman的构想 -从学术边缘到主流的劳动力市场令人眼花缭乱的灯光，随着前沿AI的开始，疯狂的流浪汉重新包装Logistic回归 认真地 我在看着你， IBM Watson 。

Most found-in-the-wild, laymen sentiments towards A.I. seesaw between extremes of the utopian and the dystopian. The utopian vision is marked by unrealistic expectations for the short term potential of what can be achieved (think self driving cars). This sentiment is dangerous because overpromising helps fuel hype cycles — a common trend in the history of A.I. — , and when those bubbles inevitably burst, public confidence wanes, and with it, reputations, livelihoods, and investors.

在AI跷跷板的乌托邦和反乌托邦极端之间，大多数人在狂野中发现外行情绪。 乌托邦的愿景是对可能实现的短期潜力不切实际的期望(想想自动驾驶汽车 )。 这种情绪是危险的，因为过分的承诺助长了炒作周期 (这是AI历史上的普遍趋势)，并且当这些泡沫不可避免地破裂时，公众的信心减弱，声誉，生计和投资者也随之消逝。

The dystopian vision, on the other hand, siphons its ideas from sci-fi. The A.I. featured in stories like Bladerunner, A Space Odyssey, and Westworld is compelling because it straddles the line between mind-bending philosophy and malleable plot device, but whether you follow the writings of Philip K Dick, Asimov, et al. is not really the point, the point is that the language and ideas employed by those authors has found itself embedded in the output of the many folk who have a platformed stake in the A.I. debate. It’s being used to help frame media publications regarding the labour market, the short term potential of A.I., as well as marketing campaigns for basic Machine Learning apps.

另一方面，反乌托邦的愿景则从科幻小说中吸取了它的思想。 像Bladerunner ， A Space Odyssey和Westworld这样的故事中的AI引人注目，因为它跨越了弯曲思维哲学和可塑情节设备之间的界线，但是您是否遵循Philip K Dick，Asimov等人的著作。 并不是真正的重点，重点是这些作者所使用的语言和思想已经嵌入到许多在AI辩论中具有平台利益的人们的输出中。 它被用于帮助构建有关劳动力市场，人工智能的短期潜力以及基本机器学习应用程序的营销活动的媒体出版物。

Just as the techno-troglodyte hacker in a hoody, hunched over a Matrix-esque screen of raining code has become the media’s darling caricature for programmers, so too has the cyborg with a neon-blue networked brain for ML practitioners. But the portrayal suggests a world which is lightyears away from what ML is currently capable of. The group that push this sentiment, and indeed the people most influenced by this group, often imagine humanoid representations, eerie reshuffles of society, and a lone genius creator, akin to hackerman. Too much focus on the dystopian extreme will overshadow the more urgent dangers we face from the misuse of contemporary Machine Learning.

就像身穿连帽衫的技术专家级黑客一样，弯腰着矩阵式的下雨代码屏幕已成为媒体对程序员的宠爱讽刺画，同时，带有ML蓝色霓虹灯网络脑的cyborg对ML从业者也是如此。 但是从描述可以看出，这个世界距离ML当前的能力还差几年。 推动这种情绪的人群，甚至是受此人群影响最大的人们，经常会想象人形机器人的表现，社会的怪异改组以及一个类似于黑客的天才创造者。 对反乌托邦极端的过多关注将掩盖我们滥用现代机器学习所面临的更为紧迫的危险。

A.I.’s role within the quack advertiser’s choice hyperbole has, to some extent, crippled its intended meaning, and reduced ad absurdum useful terminology into a clichéd, and sometimes deceiving, meme. But more pressing than the issue of semantic shift (the meanings of words always change over a long enough time horizon), definition disparities between warring factions cheapen public discourse, often corrupting it into a misinformed shouting match between frustratingly stubborn assemblies. US politics, anyone?

AI在庸俗的广告主选择夸张中的作用在某种程度上削弱了其预期的含义，并将荒唐的有用术语简化为陈词滥调，有时甚至是欺骗性的模因。 但是比语义转移的问题(单词的含义总是在足够长的时间范围内变化)更紧迫，交战各派之间的定义差异使公共话语便宜化，经常使它变成令人沮丧的顽固集会之间的误报。 美国政治，有人吗？

The most informed public debates are those for which the terminological differences between competing sides are slim, so that the encoding of a term as I play out my argument is not misinterpreted when you attempt to decode it. A foundation of robust axioms allows for more thoughtful arguments to emerge, especially important in trending technology discussions where understanding the implications of applying potential interventions are often vital and subtle.

公开程度最高的公开辩论是那些相互竞争的双方在术语方面的差异很小的辩论，因此，当我尝试解释我的论点时，对术语的编码不会被误解。 强大的公理的基础允许出现更深思熟虑的论据，在趋势不断的技术讨论中尤其重要，在这种讨论中，了解应用潜在干预措施的含义通常至关重要且微妙。

Pinning down A.I.

固定AI

So far we’ve only discussed how those outside the field generally interpret A.I. But what do practitioners think? What do those people-in-the-know who, anecdotally at least, scoff so heartedly at the dim eyed public for failing to comprehend A.I., think A.I. actually is? I’d wager that they themselves wouldn’t converge on any unanimous verdict. Further, I’d wager they’d flip-flop their definition depending on the audience, something especially true for those seeking research funding; since A.I. is currently high on the hype index you’d of course want to bill it as such. Incidentally, during the second “AI winter”, when the term had been dragged through the mud after a hype bubble of inflated expectations had burst, those left in the field found they could make more money by avoiding the tarnished term altogether.

到目前为止，我们仅讨论了外地人员通常如何解释AI的问题，但是从业者怎么看？ 什么是这些人在最知道是谁，据传至少，在昏暗的眼睛公众未能领悟AI嗤之以鼻所以全力以赴，认为AI其实是什么？ 我敢打赌，他们自己不会在任何一致的裁决上趋同。 此外，我敢打赌他们会根据听众来改变他们的定义，对于那些寻求研究经费的人来说尤其如此。 由于AI当前在炒作指数上很高，因此您当然要为此付费。 顺便说一句，在第二个“ AI冬季 ”期间，当高涨的期望泡沫破灭之后，这个术语就被拖入泥潭了，那些留在野外的人们发现，他们可以通过完全避免失去这个术语来赚更多的钱。

When some in the field talk of A.I., they’re actually referring to Machine Learning. For most applications, Machine Learning is that slim section of modelling which sits within a long pipeline of data ingestion, and transformation, munging, and serving, involving the complex interplay between many pieces of infrastructure and data warehousing and glue code and middleware and REST APIs and the like. Machine Learning is the Light GBM scoring fraudulent transactions, it’s the CNN used to detect breast cancer, it’s the Word2Vec infused recommendation engine.

当现场有人谈论AI时，他们实际上是在指机器学习。 对于大多数应用程序而言，机器学习是建模的一个狭窄部分，位于数据摄取，转换，修改和服务的漫长管道中，涉及许多基础架构和数据仓库以及粘合代码，中间件和REST API之间的复杂相互作用。等等。 机器学习是对GBM进行欺诈交易的轻型工具，它是用于检测乳腺癌的CNN，是Word2Vec注入的推荐引擎 。

When others in the field talk of A.I., they’re referring to A.G.I., otherwise known as that thing that kept Elon Musk awake at night for a brief spell in 2018. For others still, A.I. is simply the next biggest breakthrough, and always feels to be within touching distance — if only we can achieve <blank>! — but forever falls short of the mark as each newly accomplished milestone pulls back the curtain only to reveal only the same sentiment as before… that threshold A.I. wasn’t this milestone, it’s the next one, over and over. A quip from Douglas Hoffstader (or was it Larry Tesler?) points out that “A.I. is whatever hasn’t been done yet”. The idea of shifting goal posts in A.I. is nothing new, but it tends to be overlooked when people discuss the topic. This effect, by the way, is aptly known as the AI effect.

当现场其他人谈论AI时，他们指的是AGI 。， 也就是让Elon Musk在2018年短暂醒来的夜晚 。 对于其他人而言，AI只是下一个最大的突破，并且总是感觉触手可及- 如果只有我们能够实现<blank>！ -但永远都无法达到目标，因为每个新近完成的里程碑都只是拉开了帷幕，只显示出与以前相同的情感……阈值AI并不是这个里程碑，而是一遍又一遍的下一个里程碑。 道格拉斯·霍夫斯塔德(Douglas Hoffstader)(或者是拉里·特斯勒(Larry Tesler)？ 在AI中转移目标职位的想法并不是什么新鲜事物，但是当人们讨论该主题时，它往往会被忽略。 顺便说一下，这种效应被适当地称为AI效应 。

What now?

现在怎么办？

Regardless of how you define the word, A.I. has come a heck of a long way since it was founded as an academic discipline way back in ~1955. But it’s clearly far far away from its most sacred of goals: A.G.I. Today’s most advanced models are often brilliant at solving very specific domain problems, but typically fail to generalise, and for more complex problems, like image recognition, need an absurd amount of computational power and training data.

无论您如何定义单词，自1955年AI作为一门学术学科成立以来，它已经走了很长一段路。 但这显然离最神圣的目标还很遥远：AGI当今最先进的模型通常擅长解决非常特定的领域问题，但通常无法概括，对于更复杂的问题，例如图像识别，需要大量的计算力量和训练数据。

My grumble is not that we’re prematurely misattributing A.G.I., far from it. It’s that the waters of our discussion are muddied with inconsistencies, and that many of us are clogging our collective brain space with arguments about A.I. which are either inconsequential, or manipulated by the outlandish promises touted by advertisers and marketeers playing off recent hype. More pressing than the long-term threat of A.G.I., or misguided expectations around the pace of innovation, are the immediate-term dangers posed by sinister applications of Machine Learning technology which already exists. And while it won’t spell dystopia in the iRobot sense of the word, is certainly capable of being used with malign and unethical intent, with potentially disastrous effects. An example of modern notoriety here is Clearview AI.

我的抱怨不是我们过早地错误分配了AGI，而远非如此。 就是我们的讨论之水被矛盾所淹没，我们中的许多人都在以关于AI的论点来阻塞我们的集体大脑空间，这些论点要么无关紧要，要么被广告商和市场营销商吹捧的虚假承诺所操纵，这些承诺是在进行近期炒作。 比AGI的长期威胁或对创新步伐的错误预期更迫切的是已经存在的机器学习技术的危险应用所带来的近期危险。 尽管从iRobot的角度讲，它不会拼写反乌托邦，但它肯定有可能以恶意和不道德的意图使用，并可能造成灾难性的后果。 Clearview AI是现代恶名昭著的一个例子。

When all is said and done, attempting to converge on a globally shared narrative for “what we talk about when we talk about A.I.” is both infeasible and undesirable. However, the fragmented branches of the debate would be dramatically less stark if those involved understood and acknowledged that no single definition, especially theirs, should be the basis for how we frame a debate bubbling over with many definitions.

总而言之，试图融合全球共享的“ 关于我们谈论AI时所谈论的内容 ”的叙事既不可行，也不可取。 但是，如果有关各方理解并承认，没有一个单一的定义，特别是它们的定义，不应成为我们以许多定义冒充辩论的基础，那么辩论的零碎分支将大大减少。

Thanks for reading.

谢谢阅读。