Hongjiang Zhang: AI's Long-Term Development Requires Someone to Sit on the Cold Bench

源码资本·November 29, 2018·0·0

We hope AI fundamental research can look further into the future.

Source: AI Tech Base Camp

ID: rgznai100

Author: Ruoming

Two years after stepping down as CEO of Kingsoft, Hongjiang Zhang has taken on a new role as the inaugural chairman of BAAI, the Beijing-based AI research institute founded not long ago, in addition to his position as venture partner at Source Code Capital. He is focused on pushing different industries to share big data and making it easier for basic research institutions and AI startups to access data.

As someone who has moved across industry, academia, research, and investing, he undoubtedly has a more comprehensive view of AI's industrial development.

Recently, as a keynote speaker for the upcoming 2018 Big Data Technology Conference, he sat down with AI Tech Base Camp for an interview covering three areas: AI industry investment, AI company development, and talent cultivation.

On AI Investment: Rationality Is Setting In

AI Tech Base Camp: From an investment perspective, what do you see as the weak spot in China's "Intelligence+" investment?

Hongjiang Zhang: Chinese AI investment has been concentrated on the application layer. There are indeed fewer companies in the foundational layer — algorithms, chips — compared to the United States. This relates to our stage of development, but it will gradually improve.

AI Tech Base Camp: At the application layer, which "Intelligence+" sectors do you see as most promising?

Hongjiang Zhang: I'm not an expert in application domains, but growth potential generally depends on two factors: whether the application scenario can generate big data, and whether it can generate substantial revenue or already commands significant resources — in other words, industries with abundant money and data, such as finance, healthcare, and mobile internet.

AI Tech Base Camp: Industries with deep pockets and rich data have strong prospects, but this year's investment and financing data shows intelligent manufacturing attracting the most funding. What does this indicate?

Hongjiang Zhang: The larger the application scenario, the larger the addressable market it can influence in the future, and obviously the greater the potential. All the concepts we discuss are relative. Manufacturing is China's largest industry. Further intelligentization clearly means the largest market, so attracting investment in this sector is normal and relatively healthy.

AI Tech Base Camp: Early last year, industry insiders like Kai-Fu Lee predicted that a winter would arrive by year-end and the AI bubble would burst. Yet investment and financing data shows that while the number of AI investment deals dropped by one-third compared to last year, total funding actually increased by nearly a quarter. It seems AI investment fervor hasn't cooled but rather heated up?

Hongjiang Zhang: When we talk about bubbles bursting or winters arriving, these are all relative concepts. In other words, people no longer look at AI companies the way they did three years ago, throwing money at a few PhDs or professors fresh out of university. From that perspective, the bubble has already burst.

The increase in investment volume comes from more money flowing into Series B and later-stage companies, so a decline in deal count is natural. This precisely demonstrates the market is maturing, investment is becoming more rational, and people increasingly understand what core competencies an AI company needs to possess.

As for investment winters, people are really looking more at economic cycles. When the economy slows, investment inevitably becomes more cautious.

AI Tech Base Camp: If you could say one thing to entrepreneurs at AI startups, what would it be?

Hongjiang Zhang: Think clearly about what you actually want to do. First question: do the users you plan to serve actually have this need? Second question: assuming the need exists, what is your core differentiation from companies already meeting that need?

On AI Companies: Building Closed Data Loops

AI Tech Base Camp: You emphasize that an AI company's moat is data and talent — algorithms and technology alone don't constitute a business model. But the crucial question is how they obtain high-quality data and use it correctly?

Hongjiang Zhang: The question of how to acquire high-quality data already has very good answers today, and this is one of the fundamental reasons why this wave of AI has been able to rise. Twenty years ago when people worked on neural networks, it wasn't that they had no idea how algorithms should evolve — it was that there simply wasn't this much data available. But especially over the past decade, mobile internet's rapid development has created extensive, deep interaction between people and data, which itself is a process of generating high-quality labeled data. For example, when you take a photo with your phone, you at least know who took it where, when, and with which phone — all the camera parameters are there — making scene recognition quite straightforward.

By the same logic, I believe future AI companies will necessarily be so-called closed-loop companies: you have a product, the product interacts with users, and this process generates substantial feedback to improve the product and user experience, attracting more users, which in turn produces more data to train better algorithms. Overall it's a cycle, especially the "loop" of mobile internet.

So I don't think generating high-quality big data is the problem. The bigger problem, actually, is how to share big data across different industries and how to let people doing fundamental algorithm research use it to train better algorithms. That's the core issue.

AI Tech Base Camp: Since data plays a decisive role, can we conclude that AI companies without data will all die off in the future?

Hongjiang Zhang: I wouldn't say they'll die off. There will always be some AI technology and solutions consulting companies remaining, but these firms will struggle to become truly leading platform-type companies.

AI Tech Base Camp: Conversely, if existing big data companies or data platforms quickly catch up on algorithms and technology, they should have competitive advantages going forward.

Hongjiang Zhang: Exactly. Today's AI-advantaged companies — Google, Facebook, Microsoft, Amazon — their core strength lies in accessing large volumes of data from practical application scenarios. Of course they also have strong technical teams, especially Google and Microsoft. Similarly in China, BAT, Toutiao, and Meituan can continuously acquire large amounts of high-quality data. Combined with their inherent technical capabilities, their emergence as AI leaders is beyond question.

AI Tech Base Camp: At least half of the speaking topics at the upcoming 2018 Big Data Technology Conference (BDTC) are strongly AI-related. What do you make of big data conferences becoming "AI-fied"?

Hongjiang Zhang: Big data conferences have been running for many years now, and I'm quite glad they haven't changed their name because of AI's popularity. We've talked about big data for 15 years, but its practical applications have been quite limited. Breakthroughs in deep learning algorithms have provided powerful tools for big data applications. Before becoming intelligent, enterprises must first become data-driven — they need to recognize using data to drive business and industry. So for big data's future development, we should very optimistically recognize it as the core production material of this industrial revolution, and by leveraging AI as a tool, we'll find better and more extensive big data application scenarios.

On AI Talent: Herd Mentality Is Not the Way

AI Tech Base Camp: From a technical personnel perspective, what characteristics define excellent talent needed in the AI era?

Hongjiang Zhang: Industry demands on technical people are fundamentally the same in any era. The core requirements are passion for technology, solid technical foundations, and strong hands-on ability. It's just that the AI era may demand higher skill levels. AI engineers don't just write programs — they must continuously develop algorithms, possess strong data analysis capabilities, and understand application scenarios.

AI Tech Base Camp: Universities are important bases for supplying AI talent to enterprises. Since last year, a notable trend has been at least 50 universities establishing undergraduate AI schools or colleges. You've also engaged with universities on talent cultivation. What pitfalls do you see in how universities are approaching AI talent training?

Hongjiang Zhang: Any herd mentality approach is inadvisable, especially when universities all do it without examining actual requirements. For example, 15 years ago we all created software schools — today, can anyone tell me the difference between a software school and a computer science department? Ten years ago everyone started e-commerce majors — what exactly did e-commerce majors study? Today's AI rise means we need more skills from AI talent, so undergraduate education should strengthen AI course offerings and quality, making it an emphasis within computer science. If you create an AI major, do you skip computer fundamentals? Do you not need to understand computer system architecture? Basic algorithms? And before all that you need mathematics, algebra — all of this is required.

So I fail to see how rushing to create AI undergraduate majors benefits this industry. Five years from now, if we need more chip talent, networking talent, what then? Why didn't we establish AI undergraduate majors 10 years ago? What we need is to build consistent, enduring undergraduate computer science education that emphasizes fundamental capabilities, with more rigorous and comprehensive curriculum — including textbooks and lecture content that continuously keeps pace with technological development.

AI Tech Base Camp: On the other hand, the proliferation of AI schools also indicates strong demand for AI talent.

Hongjiang Zhang: I don't see top-50 U.S. schools rushing in herds. AI spent 30 years on the bench — if you open programs because it's hot, what happens when it's back on the bench in five years? What we need is solid undergraduate education, not herd mentality chasing buzzwords.

AI Tech Base Camp: What can China's AI talent education learn from the United States?

Hongjiang Zhang: For decades we've been learning from the United States about how to structure education and curricula, and this process should continue. But what we often learn are surface-level or "trendy" things — for example, today universities mostly emphasize papers, SCI indexing.

I was chatting with an academician recently about how China already has a considerable number of AI talents in academia. He said the quantity has increased a lot compared to before, but look at what they're doing. International AI conferences have no shortage of Chinese papers, but supposedly 90% of these papers are on the hottest topic of deep neural networks — exactly mirroring the situation with undergraduate AI majors. This is concerning.

Has anyone thought about the limitations of today's neural networks or deep learning? What problems can't they solve? How will these algorithms evolve? Should we do more in other areas of AI? We should emphasize sustained fundamental research — the kind of work that requires sitting on the bench for long periods.

AI Tech Base Camp: What do you see as the most important issue people need to pay attention to regarding AI industry development?

Hongjiang Zhang: The greatest risk is people thinking AI consists only of deep learning or only neural network learning. We know every algorithm has its limitations. If you only consider one algorithm, you should also look more at how to improve it, whether it can be combined with other algorithms, and thereby applied to better scenarios to create greater value.

Today's AI breakthroughs are all results of long periods sitting on the bench. The next breakthrough may also come from those sitting on the bench today. I hope AI fundamental research can look further into the future.

Source Code Capital

Creating Lasting, Genuine Value

WeChat ID:

sourcecodecapital