The Rise of AI 【转】

The Rise of AI

While it is never wise to try to foresee the future, the emergence of an era of feasible Artificial Intelligence applications in the IT industry is a trend that no entrepreneur wanna-be could ignore.

The term Artificial Intelligence seems to have been floating around for a long time. I remember once talking to a fourty-ish technology manager in an investment bank about this topic. He was very intrigued but all he could relate with from school days was chess playing. Yes. That was the AI in 1980s when rule-based systems, search and etc dominated the age. What made this term particularly warning to investors were the unrealistic promises made by AI startups in 1980s. All of a sudden, hard problems such as Machine Translation seemed to be within reach. But as how every bubble ends, AI entered a winter when peopled realized that machines could not surpass human in any near future.




However, AI made its comeback in the early 21st century. From my humble perspective, this is largely due to the wide adoption of statistical method. Instead of figuring out the exact rules, we only need to know which solution is the most likely in our solution space. Take Machine Translation as an example, computer scientists in the 80s approached this problem by designing an “inter-lingual” which is a universal language representation, and coming up with rules of how a language could translate into and from this inter-lingua. If we were to translate from Chinese to English, we first translate Chinese sentence into inter-lingua, then translate that into English. Very intuitive. Isn’t it? However, it does not work.. Because the sheer scale of rules are beyond the reach of human extraction, especially given the limited computing power at that time.

Today, we are no longer interested (or not that interested) in articulating exactly what these rules are. We want to see enough examples of Chinese-English sentence pairs. From these pairs, we can “learn” how likely a Chinese word/phrase/syntax could translate into an English word/phrase/syntax. With a little statistics trick, we can compute which English sentence is the most likely translation of the original Chinese sentence. Does it work perfectly? No. But is it becoming acceptable? Yes!

Artificial Intelligence is no more in the imagination of futurists or the rhetoric of wild entrepreneurs. It is actually an ongoing trend and we are using it everyday. Rosetta Stone would apply it to recognize and analyze speech. Amazon and Netflix would apply it to recommend items which may interest you. Google would apply it to rank web pages. Although it is still far from reaching the Holy Grail of universal intelligence, specific problems with specific domains are being solved. And we will see a lot more interesting applications coming out in the next decade, and we, as consumers, will be adapting to a world with machine intelligence.

Opportunities

Artificial Intelligence is a long wave. This is my favorite reason of why I think this is a good opportunity. A good analogy would be like PC industry in its infancy. The problem domain is exciting. The solution creates real value. Technology is reaching the tipping point. Consumers are not yet fully aware of. Unlike short term trends such like Mobile, or Social Gaming, the industry would take decades to develop and has the potential to grow massive. This is the kind of long hill that Warren Buffet was talking about, and I believe it is good for young people.




To really foresee all possible applications from the development of Artificial Intelligence is not possible, since the technology itself is being developed. My main criteria for an opportunity is that it has certain technology barrier, it is doable, it creates value or solves real problem for customers, the market is sizable and either the technology or market has just matured. The following ones are constantly grabbing my attention.

First, turn unstructured data into structured data.The internet has led to an explosion of data. In order for humans to interpret and utilize this data, we need structure, or put it in another word, make sense of them. The development of massive computing capabilities has just made processing huge quantities of data a reality. Twine has made a good attempt on solving this problem. But it seems it is not hitting the consumer’s sweet spot. A lot of semantic search companies have tried, and failed, and several are still struggling. But this job will be done.

Second, elevate personalization to the next level. Computers will not become a viable human assistant soon, but to some extent, they will start to “read” your mind and automate some of tasks. What does a personalized world look like? There will be automatically rearranged Wall Street Journal that caters to your interests and save your reading and navigating time by 50%. There will be financial planning assistant which understands your financial goals and investment styles, and handles all the stock/bond/fund mundanes for you. Whenever you go travelling, dinning or clubbing, advanced recommendation engines will make sure that the top 3 alternatives have what you would like and make sure you can focus on the fun part. The business of personalize tackles with one of the most compelling human need - sloth, and promotes another modern theme individualism, and it will certainly have its shot.




Third, human replacement comes into reality. A humanoid bot is the ultimate goal for Artificial Intelligence and it is far from being achieved. However, solving specific tasks with reasonable results are on the horizon. The industrial revolution has greatly increased human society’s physical capacity. In the information revolution, the “mind” capacity is made accessible for the machines by digitalization. Now in analogy, an era of Artificial Intelligence would greatly increase human society’s “mind” capacity. Once I talked with a friend about some recent development in speech machine translation. And he was somehow not very comfortable with that. Why? Because I later find out that he has a close family member who works as a translator. This is the kind of reality that the new generation will need to deal with. In the industrialization age, physical workers have been hit and cut; in the information age, office workers especially those have strong routine components have been hit and cut; now, certain occupations that even require human judgement may be hit and cut. It is both good and bad. But the trend will move on. Another example I like very much is Rosetta Stone, I think they are doing a great job in evangelizing electronic tutors and their product is a balanced set of matured AI technologies and deep understanding of language learning process. Are the mass used to talk to a computer like Rosetta Stone to learn? Not so much for today. But will that become popular in the future? Very possible. Just like what happens in the e-commerce space.

Problems 
There are several problems associated with today’s AI industry, some of them are business related and some technological.




The first is consumer readiness. Are consumers ready to let machines take over part of their work? How can they trust machines? On the other hand, do they really want these tasks to be automated? Take travel planning for example. Maybe part of the fun is just in the research process, browsing pictures and travel journals, drafting routes and schedules. This problem is very similar what happens to the development of e-commerce. In the beginning, the majority of the customers do not trust online transactions. It was perceived as unsecure and unreliable. And we can also argue that e-commerce rids the fun of shopping personally, being surrounded by the goods and greeted by the staff. However, as we have witnessed later, a group of early adopters evangelized the idea, the technology becomes matured and reliable, the mass realized the value of ecommerce, its convenience and competitive price due to the removal of middleman. The key lesson is that as long as an application creates genuine value, no matter if the market realizes it or not in the beginning, it could eventually fulfill its potential. As the value brought by AI is so obvious, I think it will have a chance sooner or later.

The second problem is data. The current statistical approach usually requires large quantities of data to bootstrap, to “train” your model and make it intelligent. But who has data? The giants. Google, Microsoft, Amazon, Facebook … This is particularly a problem for startups. As all of these giants are aggressively pushing the boundaries of AI, startups do not even get a chance to set their foot in the door. MyStrands is a visionary and well funded company that dedicates its resources into making usable recommendation and personalization a reality. It had this wonderful music recommendation engine Strands. However, as they possess no data of personal music usage, they eventually opted to design an iTunes plug-in for their application in order to get access to data. This strategy was not natural for a user at the first place, so it never took off. In addition, as the owner of all the data, Apple could easily counter attack once this market develops. Not surprisingly, this year, Apple launched their iTunes Genius program which eliminates all possibilities of third party recommendation engine on top of the iTunes platform. It is still unclear how this data problem would get resolved in the future. On one hand, data is undoubtedly a competitive advantage. On the other hand, by opening the data to the public to an extent would be a good strategy to draw the creativeness of the crowd and increase the competitiveness for the platform as a whole.

Finally, there is the gap of expectation and the reality. The current technology suffers from some inherent problems. It is usually domain specific, in other words, a speech recognizer that excels in handling Chinese speakers’ English may not be comparably competent in handling an Indian speaker’s English. It has its accuracy limit. With the current approach, it may never reach what a human can do. It may be slow … All of these pitfalls, if come with a controlled customer expectation, could still lead to reasonable success. But in reality, the need for PR stunt usually overwhelms technology. I think the reason for the general struggling of the “semantic search” market has partly resulted from this. Consumers soon realized that the improvement brought by these semantic companies (Powerset, Hakia, …) is not “revolutionary” but merely incremental. If the perceived benefit by the consumer is only 10% improvement, but we require 100% more investment in technology infrastructure, is it a sound decision by a startup? Maybe we are not solving the right problem, maybe we should apply it in somewhere the perceived value improvement is “revolutionary”.

http://www.andrewjunhu.com/?p=24