The development of AI has gone through three waves to this day. The first two waves were blocked due to the technological environment at that time and other reasons. In the end, the tide receded, leaving sighs and regrets.
Now, with the maturity of the technological environment and the continuous upgrading and improvement of algorithms, computing power, and data, the third wave of AI has come again, and it is even more violent than the previous two.
The dog named "Alpha" defeated humans in Go, and people discovered that AI was really better than humans in some aspects.
The streets and alleys are talking about artificial intelligence, and everyone can look forward to the changes that AI will bring to life in the future.
Algorithm engineers have become the darling of corporate recruitment, and large Internet companies have begun to pay high salaries to introduce algorithm personnel, hoping to take the lead in this wave.
Whether it is a technology startup company or a traditional startup company's external promotion PPT, the concept of AI must be immersed in the concept of AI, hoping to get more attention from investors.
It seems that the scenes that people see in science fiction films will soon be realized. However, what ordinary people don't know is that the third wave of AI still encountered some difficulties.
01 AI landing encounters difficulties
The difference between this wave and the previous two waves is that the first two waves were mainly participated by academia. This time, in addition to academia, more industries have participated. The goal of this wave is to pass For commercialization to be profitable, it is necessary to be able to truly sell AI.
For AI to be paid by someone, it must be able to solve practical problems.
So now referring to the key elements of AI, in addition to the three elements of "algorithm", "computing power" and "data", a "scene" element has been added. AI technology must be implemented in specific scenarios in order to realize its true value.
There is no doubt that this "scene" element is the key element that allows the third wave of AI to continue.
The AI community chanted the slogan of "landing" and began to recruit troops and acted boldly.
However, the actual situation is that the AI craze has encountered a cold wave in the past two years, a large number of AI companies have closed down at a loss, and investment enthusiasm has fallen sharply. The remaining companies are silent, some are struggling to survive, and some are accumulating strength and looking for opportunities.
02 Cause analysis
So what causes the difficulty of AI landing? I summarized the following reasons:
1. It is difficult for pure technology-oriented companies to survive
AI companies on the market are roughly divided into basic type, technology type and application type according to their types. Basic companies mainly provide low-level software frameworks or hardware platform services. Application companies generally focus on packaged solutions in several industries, while technology companies mainly focus on deep learning algorithms and encapsulate them into APIs for external system calls. Kind of upper-level application empowerment.
First of all, this kind of pure technology-oriented company focuses too much work on algorithm research, and ignores user, product and market research, and gives up control of user needs. This business model is incomplete.
If a company is obsessed with publishing papers and playing games, and is immersed in self-set conditions every day and is complacent, it will be farther and farther away from real needs.
The first step for any product is to clarify the problem and the user, not what technology you have. This "holding a hammer to find a nail" method is difficult to really solve the problem, but unfortunately, this is the norm. Because they label themselves AI, they must use AI methods to solve problems.
Large companies research algorithms, but not just algorithms. Companies such as Google, Microsoft, and Ali have succeeded not only because they have first-class technology, but also because they have a stable business model.
Secondly, AI algorithm barriers are too easy to be broken. Since the algorithm update speed is very fast now, new algorithms will be open sourced every once in a while. For example, the BERT model shines in a short period of time after its release. It has set records in more than a dozen NLP fields. Far more than other other algorithms.
Faced with this phenomenon, will customers still pay a high price to buy an algorithm that may be surpassed at any time? Isn't it cheaper and more efficient to directly call the latest open source algorithm?
Third, AI companies understand technology but not business, and customers understand business but not technology. When the two interact, this cognitive bias will cause a lot of additional learning costs.
This kind of problem is especially common for small and medium-sized AI companies. On the other hand, large-scale companies have their own AI teams and their own independent businesses. They engage in AI mainly to serve their own businesses. For example, Baidu's search system, Douyin's recommendation system, Taobao's search for images and other algorithms are all applied to their own business, which can be trial and error. At the same time, because of the continuous flow of data, the effect of AI evolution is remarkable.
Large companies have sufficient soil for AI to take root and grow rapidly, while small and medium companies do not have such conditions, leading to slow development.
In comparison, it is more reliable for traditional companies to engage in AI than AI companies to engage in business. As long as a similar algorithm based on the business is used first and iterated slowly later, it will always get better and better.
So the fastest growing AI must be those companies that have their own independent business. Adding AI to the original business will save you more effort than doing business with AI. Because of this, the position of pure AI technology companies is very embarrassing, and now it has become less and less.
2. The experiment is far from the actual
I often see an AI company claiming on its official website or official account that its algorithm accuracy has achieved a breakthrough. After testing on a certain data set, the accuracy rate has reached more than 99%, surpassing the world's leading companies such as Microsoft and Facebook... description.
But the significance of such indicators requires a big question mark. Not to mention the real effect, how can people be convinced just by letting participants keep trying and getting scores under conditions similar to telling the answer?
The most important thing is, are these algorithms really effective in real scenarios?
You know, the difference between the experimental scene and the real scene is very big. In real scenarios, there are many more factors to consider than experimental scenarios. We have achieved 99.8+% accuracy on the LFW face recognition data set earlier, but the face recognition effect in real scenes is still not ideal, and it will be subject to various conditions such as illumination, angle, occlusion, and camera resolution. Constraints. Instead of trying to improve the accuracy of the algorithm by less than one point, it is better to add a fill light or change the camera.
The face recognition algorithm is still a relatively mature algorithm, and it can still play a good effect in actual scenes. However, there are still many niche-demand algorithms with accuracy of only 60% and 70%, which are far below commercial requirements.
Regarding the accuracy of the algorithm, what should be solved is how to do it from 80% to 90% in a real environment, not from 99.5% to 99.7% in a laboratory environment.
Some companies disguise "artificial" as "intelligence" in order to package themselves. A banking robot can smoothly joke and be cute, but even the most basic arithmetic problems have to be calculated in the background for a long time. Is this not obvious enough? Not to mention incidents such as simultaneous interpretation fraud that have been exposed.
The expectation of the outside world is one thing. If you really believe it, you will deceive yourself and others.
3. Many customization requirements, low input-output ratio
Some things are difficult to do, not necessarily a problem of technical implementation.
Many startups have a puzzle. Should the initial development be project-driven or product-driven?
If it is product-driven, whether it is through market research or the boss's decision, a direction is determined, and the company begins to invest resources in product development. In this case, there is no profit in a short time. At the same time, various questions need to be solved, such as where does the data come from? What is the business process? Does this design meet the needs of the enterprise? What is the user usage scenario?
If it is project-driven, bosses or executives find one or two customers through their own channels, usually traditional customers with AI needs, hoping to achieve the goals of reducing costs and increasing efficiency through AI. In this case, customers often put forward various customized requirements based on existing businesses. Some requirements can be modified slightly and can be used as general product requirements, while some requirements are just for this customer. Therefore, for some small scenes, if they are not very important customers, AI companies are reluctant to accept them.
I remember that when I was working on AI projects, I often received various customization requirements, such as identifying the grade of scrap iron when unloading at the scrap iron factory, identifying the eye status of personnel during high-altitude operations, and identifying service personnel in the government affairs hall. The behavior of "talking to ears" for identification and so on is too numerous to mention. If you only look at whether the algorithm can be implemented, the probability is not too big a problem, but when it comes to a complete system-level solution, there will be many engineering issues that need to be considered.
For customers, a small scenario is not worth investing too much money. For AI companies, this little money is not worth investing so much resources.
03 Conclusion
Although AI technology has encountered various problems in the process of landing, I personally believe that this time the AI wave will not suddenly go silent as the previous two.
Because this is the wave of globalization, and AI technology has begun to penetrate in all walks of life, and it has really brought obvious changes. So in the future, although there will be reciprocation of the small time axis, the large time axis must be a gradual upward trend.
Sometimes "holding a hammer to find nails" is just a helpless move. You can quickly recognize your own ability boundaries and hit them as soon as possible to truly gain a foothold.