机器学习是否被过度炒作？
Is Machine Learning Overhyped?
风投人 Tomer Tunguz 认为机器学习并非又一波炒作，2016 年第四季度的研究进展掩盖在热度之下：计算机已能像人类一样理解语音、生成近乎无法分辨的语音、在未读过目标语言的情况下完成翻译、无需人工输入生成新加密方案，并为图像自动生成描述。
In short : Explores whether machine learning is overhyped, highlighting significant real-world advancements and future implications.
For the nine years I’ve been a venture capitalist, there’s always been a buzzword of the year. Solomo (social local mobile). Mobile-first. Realtime. Big data. 2016 was the year of machine learning. Is ML just another wave to crash and dissipate on the trough of disillusionment?
I don’t think so. In this rare case, I think hype is masking quite a bit of true technical innovation. During last quarter of 2016, machine learning research has made huge strides.
Computers now understand human speech as well as other humans. Computers can talk in a way that’s close-to-indistiguishable from true human speech. Computers can translate from one language to another, never having read the second language. Computers can generate new encryption schemes without human input. Computers can write captions to describe a given image.
These innovations aren’t limited to the lab. Tesla’s self-driving car reduces crash rates by 40% and a brick-laying robot builds walls 3 times as quickly as a human.
While some may groan that every pitch deck is littered with the words machine learning or artificial intelligence, I think each deck ought to be. Because over the next five to ten years, nearly every company will use machine learning in some form.
Advertising optimization. Antifraud software. Intrusion prevention. Stock trading These first uses of ML in the 2000s reveal the characteristics of problems that benefitted from machine learning: frequently repeated processes whose decision-making could be measured, quantified and back-tested. And processes in which humans could determine which factors in the decision are important.
The algorithmic advances in 2016 enumerated above broaden the range of applications for ML. With the right volumes and types data and processing power, computers can develop enough of an understanding to predict, optimize, segment or detect anomalies in many new domains like speech, like language generation, like image recognition, like natural language understanding, like image and music creation. And they can do it with far less human guidance than before.
While the we may not yet be able to build many of the things we dream of in software, we’re getting much closer much faster. Startups are going to revolutionize existing software categories with ML, and they will create new categories of software with ML.
Get the next one in your inbox
The 1-minute read that turns tech data into strategic advantage.
Read by 150k+ founders & operators.
GP at Theory Ventures. Former Google PM. Sharing data-driven insights on AI, web3, & venture capital.
来源：Tomer Tunguz 博客（VC 分析） · tomtunguz.com