Sequoia China recently released an article featuring an interview with their Partner Yue Ji on his framework for thinking about AI investments and how he's applied it to Sequoia's current portfolio. It covers AI's application in a number of industries, why Sequoia focuses its AI investments on clear use cases, the current AI talent gap, and the winner-take-all nature of the technology.
***
"Why have we been optimistic about our investment in Guazi (a C2C used car trading platform), and kept increasing our share in the company? Because Guazi can get data from its users during transactions, and conduct deep learning based on those data, to improve the consumer experiences in the future", Yue Ji, the Partner of Sequoia Capital China, said. Guazi has become one of the typical cases in Sequoia China's AI portfolio.
Why typical? Because the investment in Guazi matches two primary considerations of Sequoia's investments in AI. First, the company should have real use cases and address real world problems. And second, the system should have the ability to improve itself through a continuous stream of data.
(Source: www.guazi.com)
Recently, in an interview with PE Daily, Yue Ji gave a detailed explanation of Sequoia China's AI map. He has been in charge of or involved with the investments in a series of outstanding companies, such as Dianping, ele.me, Tuniu, Ganji.com, Noah Wealth Management, Maple Leaf Education Systems, and more. In recent years, Yue Ji and Sequoia China have reaped large returns from their AI investments, including with Guazi, 4Paradigm, Zuoye Bang, Patsnap, Cloudwise, Sensors Data, Infervision, and Ping++.
Behind the investments made by Sequoia lies one rule: the business plan for this AI application must be feasible. Lots of AI startups are holding a hammer and looking for nails. They do not know about traditional industries and are not aware of the pain point in those industries. Whether or not those nails exist is the problem the people holding the hammers have to think about carefully.
***
What does the AI Investment Map of Sequoia Capital China look like?
Security --- Yitu Tech, DeepGlint, Mininglamp
Finance --- 4Paradigm, JD Finance, 100Credit.cn, Ping++
Media / Information --- Toutiao, Kuaishou, Miaopai
Domestic Service --- Meituan, ele.me, JD Dada, WINNER Technology（汇纳科技）
Transportation --- NextEV, Didi, Mobike, PonyAI, Guazi
Healthcare --- Infervision, Voxel Cloud, Synyi
Hardware --- DJI, Ninebot, Horizon Robotics, Chumen Wenwen
Technology --- Cloudwise, Patsnap, Sensors Data
Those companies are the bellwethers in their corresponding fields. For instance, Toutiao's automated news recommendation app uses AI to generate content for around 700 million users. The AI product provided by Infervision could reduce the average amount of time needed for CT image analysis from around 15 to 30 minutes to just a few seconds. The electric car from NextEV, the ES8, is ready for mass production. Combined with Didi and Mobike, Sequoia Capital China's investments in the transportation industry have also changed the way people get around in China.
(NextEV SUV | Source: www.motorauthority.com)
Artificial Intelligence industry could be divided into three dimensions: fundamental research, technology and applications. Sequoia Capital China is currently focusing more on the application level.
“The expansion of artificial intelligence applications is not an abrupt process but a gradual one. If you look at the business plans of some previous products, it's highly possible that the whole document didn't mention AI. However, as projects develop, people come to realize that what they're doing has AI applications as well,” Ji Yue said.
Currently, most investments in AI industry are in the business sector. Some predict that it will take another 5 to 10 years before AI could be applied in the consumer sector. But Ji Yue doesn't agree with that. "Toutiao is a perfect example of AI application in the consumer sector, and same for Guazi. Have you used NetEase Music? It also has AI elements, so you are enjoying the convenience brought by AI unknowingly. It's not the prediction of experts, but the efforts of entrepreneurs, that could bring artificial intelligence to the consumer sector. It will happen once you address the consumers' real needs.”
***
Two core elements: scenarios and data
When we take a closer look at the portfolio of Sequoia Capital China, we find that there are two kinds of companies in the AI industry. One type is pure AI companies like Yitu Tech, Mininglamp, 4Paradigm, Infervision; the others are companies with an enormous amount of data, like Toutiao, Didi and Meituan. What they have in common is that they not only improve their operating efficiency with artificial intelligence but also bring users brand-new experiences through AI.
The logic behind these investments, according to Ji Yue, is that people shouldn't focus too much on the technology part when making investment decisions in AI industry, but should concentrate on the use cases. It's still too early to create an AI application that could solve all the existing problems; this might not happen within the next 30 years. However, it would be much easier for AI to solve problems in a particular industry, like in healthcare, finance and personal security. Those AIs are "vertical AI" applications; by focusing on the vertical market, boundaries of problems get defined more precisely and therefore the technical difficulties encountered during data processing get reduced.
***
First and foremost: real use cases. AI that could solve specific problems.
Behind the investments made by Sequoia, there is one rule: the business plan for the AI application must be feasible. Lots of AI startups are holding a hammer and looking for nails. They do not know about traditional industries and are not aware of the pain point in those industries. Whether or not those nails exist is the problem people holding the hammers have to think about carefully.
So, is it necessary for the founding members of AI startups to have industry background? How important is technology in AI startups, compared with a deep insight of the pain points in the industry? According to Ji Yue, a background in the vertical market is not necessary; an understanding of the needs of users is much more important. Take Mobike as an example. Its founder, Hu Weiwei, was not in the bike industry herself. What she did was to find out the real needs, and then realize what value AI had in addressing them.
***
China's AI in a Macro Perspective
China is experiencing a late-comer advantage in the AI industry. The application of big data is emerging in multiple industries; people have already realized the value of big data, while at the same time artificial intelligence is mature enough to be commercialized. Therefore, big data, cloud computing, artificial intelligence and SaaS are getting combined together in China, creating tons of start-up opportunities. Meanwhile, Ji Yue also reminds the founders not to apply those concepts mechanically. Whether or not a business model could succeed is based on reasonable use cases; customers are willing to pay when your product or service could bring values to them. The users of Toutiao don't care if AI technologies have been running behind the application. They only care about seeing what they want to see, and that's enough.
Companies also must incorporate big data into the practice of AI, since AI produces the best possible results when the algorithms are trained with large datasets.
Take DiDi Chuxing as an example. Data shows that DiDi had already reached 1.43 billion of ride requests annually as early as in 2015, making it the second largest number of daily transactions worldwide (only TaoBao has more). Currently, DiDi is accepting over 20 million ride requests each day, with more than 30,000 requests per minute from users during rush hours. Didi is now beginning to predict users' destinations using AI.
(source: www.thebeijinger.com)
Another example would be Mobike. At this moment, Mobike has over 100 million registered users, with 20 million ride requests daily. It is the largest online traveling platform worldwide per number of requests. In just July this year, Mobike's mobile app has hosted 648 million client sessions and 46.8 million hours of client session time.
Currently, all Mobike's bicycles are equipped with IoT communication chips and satellite navigation chips, which are compatible with major satellite navigation systems worldwide such as China's BeiDou, United States' GPS and Russia's GLONASS. These chips are producing 20TB memory of traveling data on daily basis, upon which Mobike is able to make forecasts of supply and demand for rentals, provide guidance for operation and deploy geo-fencing to address illegal parking incidents with the assistance of Mobike's own AI data monitoring platform “Magic Cube”.
***
AI's Talent Gap
It is possible that, in the foreseeable future, to be a successful company in any industry you'll have to use AI. With this assumption, there would be a huge gap in the supply of AI talents. It is the why Andrew Ng founded Deeplearning.ai and 4paradigm group founded 4paradigm university.
“Acquisition of talent is a challenge for any startup, but especially for those in a new industry. Every AI company right now is faced with the scarcity of high end talent. Therefore, it requires the CEOs of these startups to have strong technical background, so that they can locate and hire suitable people.”
Indeed, other than the acquisition of talents, the founders of startups would also have a series of problems to consider: sales, management, operation, and the coherence between product and client's needs. But the major issue lies in the founders themselves - it is entirely different to manage a company with 10 people than one with 200, and founders must make quick adaptations, and maintain a grasp over an an ever-changing field.
According to Ji Yue, while all the merits from a successful startup should be credited to the entrepreneurs, what the investor can do is to provide the CEOs with advice on strategy, financing, acquisition of talents, and operation, based on the combination of his or her network and resources, as well as their own experience.
A promising industry necessarily attracts the entry of large amounts of capital. As for the investors, the target eventually comes to the top 10 percent of the companies within the industry. “It is always the case that less than 10 percent of the companies generate more than 90 percent of values, whereas 5 percent of the companies produce 95 percent of return in TMT industry, a phenomenon which remains constant. Therefore, Sequoia doesn't care about whether or not there's a bubble in the industry, or whether the average valuation in the industry is high or low.” Ji Yue said. Sequoia is always looking for that 5 percent.
Sequoia has always been the fund that makes major investments in the "A" round; it is the DNA of Sequoia. Now, Sequoia also begins to pay more attention to earlier pre-"A" round investments. Ji Yue says, “Sequoia now actively starts earlier pre-'A' round investments. We welcome ambitious entrepreneur come to Sequoia at pre-'A' or 'A' round at anytime to gain best resources and race together for the long haul.”
Chinese Version
数十笔投资背后，红杉坚守着一个准则：AI的商业落地必须切实可行。现在很多AI公司被称为“拿着锤子找钉子”，不了解传统行业，不了解产业痛点。而需求是否真实存在，是需要创始人花大量时间去研究的。
***
红杉中国的 AI Map
红杉中国的AI Map什么样？在哪些点插上了杉叶旗？我们绘制了这样一张图，而这张图呈现了当前人工智能商业化的几乎所有关键场景。
安防——依图、格灵深瞳、明略数据
金融——第四范式、京东金融、百融金服、Ping++
传媒/信息——今日头条、快手、秒拍
生活服务——美团点评、饿了么、达达、汇纳科技
汽车/交通——蔚来汽车、滴滴出行、摩拜单车、PonyAI、瓜子二手车
医疗健康——推想科技、Voxel Cloud、森亿智能
硬件——大疆创新、Ninebot、地平线、出门问问
技术层——云智慧、智慧芽、神策数据
这些公司都是各自领域的领头羊：今日头条用户7亿，已成为国内最大的信息推荐引擎之一；推想科技的AI产品可以将医生平均每份15到30分钟的CT影像分析时间缩短至几秒钟；蔚来汽车的ES8智能电动汽车即将量产, 加上滴滴、摩拜，红杉中国在汽车/交通领域的被投企业正在深刻影响着出行生态的变化……
在计越看来，AI可以分三个层面：基础层、技术层、应用层，红杉中国正在将注意力更多的放在应用层。
“你会发现AI布局是慢慢形成的，把当时的项目Memo翻出来，可能整篇都没有提到AI这个词，但随着行业的发展，人们发现它们也是AI应用的一种。”计越表示。
当前，AI投资大多集中在B端，有判断说AI在C端落地还需要5-10年，对此计越并不同意：“今日头条就是很好的在C端落地的应用，瓜子二手车也是。你用网易云音乐吗？你不知不觉中也在享受着AI带来的便捷。To C的AI什么时候实现不是靠专家的判断，而是通过创业者的努力，只要把用户的需求实实在在解决了，它就会实现。”
***
两大核心要素：场景和数据
我们进一步梳理红杉中国的投资组合时发现，与AI相关的可以分为两类公司：1.垂直行业AI公司，比如依图、明略、第四范式、推想科技；2.有自身的核心业务，并且有海量数据的互联网公司，比如今日头条、滴滴、美团点评。这两类AI公司的核心共同点是给用户带来全新的体验，并且提升自身的运营效率。
计越表示这其中内在的投资逻辑是：AI投资不能过多地讨论技术，而是要聚焦其使用场景。“现在让AI解决所有的问题还是太早了，有可能未来30年才会出现。因此，在垂直方向上实现应用的可能性更高，比如医疗、金融、安防等等，从具体场景切入会更容易，按照这个思路，最终的结果看起来像‘垂直AI’。”计越进一步补充说，“实际上，垂直可以把边界定义得更清楚，处理数据的技术难度会相对降低。”
这也回答了我们一直关心的问题：红杉在AI领域投资时，衡量项目的尺子长什么样？重点看几个维度？几个指标？什么样的项目红杉是不予考虑的？
首先，要有实实在在的使用场景，可以解决具体问题。
数十笔投资背后，红杉坚守着一个准则：AI的商业落地必须切实可行。现在很多AI公司被称为“拿着锤子找钉子”，不了解传统行业，不了解产业痛点。而需求是否真实存在，是需要创始人花大量时间去研究的。
那么，创始团队有产业背景是标配吗？AI创业公司中技术的重要性有多大？比之对行业痛点的深刻洞察哪个更重要？计越表示，垂直行业背景不是必备条件，但是创业者要对用户的需求有深入洞察。比如摩拜创始人不是做自行车的，但是她发现了实实在在的需求，并且意识到在这个过程中AI能够发挥价值。
从宏观发展来看，中国的后发优势正在显现，大数据应用在很多行业出现，大家越来越意识到数据的价值，恰好AI发展到可以产品化的阶段。于是，在中国大数据、云计算、人工智能和SaaS四个创新周期压缩成一体，这带来了海量的创业机会。与此同时，计越也提醒，创业者不必去套用这些概念，商业模式能否成功运转的根本在于是否是一个健康合理的使用场景，你的产品和服务可以给用户带来什么价值，这样用户才会买单。比如今日头条，只要让用户看到想看的内容就好，用户并不关心是否有“AI技术”。
其次，必须要跟数据结合，没有大数据，AI将无的放矢，只有用大量数据来训练AI算法模型，AI才能够发挥实际价值。
比如滴滴出行。数据显示，其早在2015年订单数量就已经达到14.3亿，成为仅次于淘宝的全球第二大在线交易平台。目前滴滴平台上每天的订单数量超过2000万，高峰期每分钟接收超过3万乘客需求，每日路径规划请求超过200亿次，将近1389万次/分钟；单次路径规划计算时间小于1毫秒；每两秒就能做一次订单匹配：“猜你想去”能在2毫秒内预测用户目的地，准确率为90%。
再比如摩拜单车，当前已拥有超过1亿注册用户，日订单量约2500万，是全球最大互联网出行平台。7月摩拜单车APP启动次数达6.48亿次，摩拜单车APP的使用时长达4680万小时。
当前，摩拜单车每辆单车安装了兼容世界主流卫星导航定位系统的“北斗+GPS+格洛纳斯”多模卫星导航芯片和物联网通信芯片，每天产生超过20TB的出行大数据。在此基础上，利用人工智能大数据平台“魔方”，在供需预测、停放预测和违停识别等领域提升管理精度和运维效率。
***
红杉要找的正是这5% 更关注早期Pre-A与A轮企业
也许，在可见的未来，每一个产业成功的公司都将是AI公司。按照这个逻辑，AI人才将严重不足，于是吴恩达创立Deaplearning.Ai，第四范式成立范式大学。
“每个创业公司都会面临人才的挑战，尤其在一个新兴行业的早期，高端人才必然会非常抢手，这是当前每个AI公司都要面对的。这就需要公司本身CEO有比较强的技术能力，找到适合公司的人才，把他们招募进来。”
同样，除了人才招募，创业征途上，创始人会遇到一系列问题：产品是否与用户需求妥贴匹配的问题，销售的问题，管理的问题，运营的问题，而最主要的问题是来自于创业者本身能否适应创业公司的节奏变化——管理10人的公司跟管理一个200人的公司对于创业者来说是完全不一样的，尤其是行业处于不停的变化过程中，对行业的认知能否及时跟上，并且认知更深、更符合行业发展规律，尤为关键。
计越认为，创业成功的所有光环应该归企业家所有，投资人可以做的是通过自身网络与资源的对接，以及根据自己经验提供一些战略、融资、人才和运营的建议供CEO参考。
一个有前景的行业，必然会面临大量资本进入的情况。对于投资来讲，最终还是要找到行业中顶尖的10%的公司。“永远都是不到10%的公司创造90%以上的价值，在TMT行业，则是5%的公司创造了95%的回报，这个现象一直都存在。所以红杉不关心行业有没有泡沫，或者行业的平均估值高不高。”计越说，红杉要找的正是这5%的优秀公司。
红杉一直以来都是以投资A轮为主的基金，这是红杉的DNA，现在也开始更多关注早期的Pre-A轮的积极投资。计越表示，“红杉目前也开始了更早期的Pre-A 轮的积极投资。我们随时欢迎有雄心的创业者在Pre-A轮和A轮来找红杉，获得最好的资源，一起长跑。”
This article was originally shared by Aimee Charlotte via PEdaily on August 29 2017. Translated by Juzhi Zheng, Xiaotong Lin, Yang He, Zhixiang Liang, Celine Ding, Shaolong Lin, edited by Jordan Schneider.
