Image source @ Visual China
Wen 丨 brain polar body
In 2018, everyone wants to think of the AI industry. By 2022, this matter may be marked with a big question mark.
Not long ago, the data report released by the US recruitment platform Dice showed that there was a significant salary decline in artificial intelligence-related positions in the United States in 2021. Salaries in machine learning, natural language processing, and artificial intelligence were reduced by 2.1 percent, 7.8 percent, and 8.9 percent, respectively, by more than $11,739. This is the first comprehensive salary cut in the US AI industry after the AI boom triggered by deep learning. If you pay attention to the AI industry and AI jobs, you will find that similar events have already spread in the field of AI in China.
Since the second half of 2019, it seems that every three or five minutes can hear "bad news" from an AI company or the AI department of some Internet manufacturers. Salary cuts, serious overtime, difficulties in listing, big cattle running away, and so on. In the second half of 2021, major AI companies began to frequently see rumors of layoffs and layoffs of AI departments. Compared with the layoffs of Internet companies a few years ago, ai crazy to absorb talents, people can't help but sigh.
Is it the bursting of the AI bubble, the return of value, or the resurgence of the "AI winter" in history?
China's technology market is being staged, and the high probability of continuous fermentation of AI retreat, what is its real core?
All kinds of unsmooth AI, is the cold winter really coming?
Since the Dartmouth Conference in 1951, the development of AI has experienced two famous "cold winters" in its history. The impact of these two events on the AI industry can even be said to be devastating. So when AI technology has a problem again, people's first reaction is, is winter coming again?
Judging from the development of the domestic AI industry in the past two years, the development momentum of the original fire cooking oil has indeed encountered many "water reversals". First of all, we can see that the AI gods who were originally involved in the industry from the academic world began to rethink the development plan.
In 2019, Zhang Tong, director of Tencent AI Lab, left his job to join the Innovation Factory, serving as the director of the joint laboratory between HKUST and the Innovation Workshop, and concurrently served as a research partner. In July 2020, Wei Xiushan, founding president of Megvii Nanjing Research Institute, left his post. Immediately after, Ma Weiying, vice president of ByteDance and director of AI Lab, left his job and joined the Intelligent Industry Research Institute of Tsinghua University.
Entering 2021, AI experts and technology leaders in the industry have accelerated their return to academia or the pace of entrepreneurship. In August 2021, Li Lei, director of ByteDance's AI Lab, left his post and joined the University of California, Santa Barbara. In November, Qi Yuan, former vice president and chief data scientist of Ant Financial, joined Fudan University as the dean of Fudan Artificial Intelligence Innovation and Industry Research Institute.
With the return of academic talents, AI companies that were once popular with capital have also been in trouble. First of all, squirrel AI, Bitmain and other controversial AI companies have problems, bursting out a variety of layoffs, salary cuts "melon". Companies like AI Four Dragons are also struggling in the vortex of capital and efficiency. Problems such as overtime, salary cuts, difficulties in listing, and frequent inquiries after listing have continued to appear. There is much skepticism about the continued high R&D investment of such companies and the difficult business market.
Next, the AI projects of internet giants have also been affected, and the myth of high salaries that is common in the industry has begun to shatter. With the controversy that AI projects have little income and the prospects for doing AI rise are slim, many Internet manufacturers have begun to shrink AI projects and lay off AI departments.
However, at the same time as a piece of bad news began to appear, the framework competition and the big model competition in the domestic AI market surfaced one after another. The methods of combining AI with vertical industries have gradually enriched, and emerging AI opportunities such as full-vision autonomous driving and biocomputing have begun to attract attention.
Different from the "AI winter" in history, today's Chinese AI industry is facing more of a problem of low efficiency, and some talents and enterprises and capital have gradually lost patience and chosen to withdraw from this money-burning competition. The historical AI winter is more manifested in the complete negation of the AI stage core technology route, and new technological opportunities will exclude AI from the mainstream vision. The former manifested itself in the release of the 1973 Lighthill Report, and the latter represented the elimination of the expert computer route by the PC.
Since it is different from the "AI winter" in the true sense, what is the core of the retreat and unsmoothness shown by China's AI industry today?
The core crux of the problem lies in the failure of the algorithm economy
Since today's situation is that some enterprise AI development is not smooth, layoffs and salary cuts, some development is still smooth, still continue to upgrade product technology; many AI bulls choose to stop losses in time and return to the academic community, but there are also many academic talents who devote themselves to the industry to develop smoothly, and there are still people who continue to join. Then it is necessary to analyze what exactly caused the "in" and "retreat" of the AI industry to appear synchronously.
A closer look will reveal that AI companies or projects that have "water retrogrades" are roughly divided into three categories:
1. Algorithm companies with basic AI algorithms such as face recognition, machine vision, and speech recognition as the core of their business have lost their commercial vitality when the security market is saturated and government orders are close to bottlenecks.
2. The AI department of an Internet company is essentially to serve its own product system with AI capabilities, and by the way, it also wants to monetize AI. But lacking the pioneering technology and capabilities of the B-side, the basic disk is to grab business from algorithm companies.
3. Some companies that use AI as a gimmick, such as AI mining, AI education, and AI fitness. Such companies have also absorbed a lot of AI talents, and there is a high probability of hasty ending after AI cannot arouse capital interest.
What these three types of companies have in common is that their goods and selling points are concentrated in AI algorithms. Algorithms are of course the core of AI, but they can't be the whole picture of AI. There are not many types of basic speech, semantics and machine vision algorithms, and the general problems that can be solved are relatively basic recognition problems. Such AI algorithms can indeed meet some needs in C-end and B-side scenarios, such as face recognition in urban security and public transportation systems, but they cannot meet the differentiated needs of enterprises for "intelligence".
When more and more enterprises provide basic AI algorithms, the market that simple algorithms can meet continues to saturate, and this algorithm economy will also dry up rapidly. In a few years, the AI algorithm call dropped from a few cents to a few cents at a time, and further became an algorithm to send it in vain, and it was enough to pay the traffic fee. When algorithms become cheaper and more abundant, no longer scarce products, companies that still focus on algorithms are in trouble.
Of course, we can see that every AI company today does not only provide basic algorithms, but also has a variety of product architecture and technical ideas. However, for these technologies to go down from the PPT and become the real enterprise service market share, there are many thresholds that need to be crossed during the period.
When the algorithm economy gradually fails, and enterprises do not have the cost, determination and ability to differentiate, customize, and solve the enterprise market, of course, they can only shrink the AI business, at least they can no longer afford the high jobs and salaries of AI talents.
In other words, the retreat that AI is experiencing is not that AI is useless, but that a large part of the AI companies sell things that are too simple. Mention that smart cities are always security, mention that industry is permanent is quality inspection, and mention that the Internet is beauty and voice assistants, intelligent recommendations. These scenarios lack the depth of development, and there has been supply saturation, which is difficult to form a driving force for further development. Some people say that the AI industry is invincible in the PPT world, the demo is not satisfactory, and it has gone to the market to lose money. If AI cannot be turned into high-premium software products and services, then it is indeed possible.
This is why we rarely see updates to the basic algorithms of AI today, but we can still see upgrades at the AI product level. Big models, development frameworks, solvers, simple machine learning, all point in one direction: escaping the algorithmic economy and moving toward standardized, high-premium software.
The essence of the AI problem is a cost problem
Some people may say that we can see the combination of AI companies and various industries, there are so many very wonderful cases, how can we say that AI companies have always stayed in selling algorithms?
Indeed, if you only look at the conference and PPT, the integration of AI and the industry and the enterprise market can be described as colorful. The value brought by each of these cases, if placed in the national or global industry market stock, is the market share of Weiguan.
But the question is, how much did AI companies spend to complete these cases? Are they really replicable?
This is the biggest problem of AI at this stage, and technology service providers will face huge comprehensive costs when they break away from the simple and low-cost algorithmic economy and embrace the industry market with high premiums.
First of all, the biggest cost of AI is still the cost of talent. At present, many AI projects with relatively high premiums require technology providers to mobilize a large number of experts to carry out on-site, support and even long-term presence. A lot of simple parameter tuning requires Ph.D.-level talent to do it. These people are extremely well-paid at first and foremost, and at the same time they are essentially more like academics than front-line engineers. They come to the front line of the industry with huge communication costs. If the accumulation of high-level talents is relied upon for a long time to achieve case success, then AI cannot be replicated at scale.
On the other hand, customized AI models also require more complex training environments and longer training cycles. This involves high AI hardware costs. Currently, training a relatively complex machine vision model often requires a large number of GPUs for months or even a year of training. The cost of hardware rental can reach millions, or even tens of millions of dollars. Such hardware costs are difficult for both technical service providers and end users to afford.
Another cost pressure of AI comes from the ambiguity of benefit returns. Most AI enterprises and business departments are technology-oriented and lack experience in enterprise market services. Therefore, it is difficult to judge which industry and which product can bring accurate market feedback. Therefore, there are often a large number of trial and error costs, as well as internal market relations and internal friction, which in turn leads to a large amount of cost waste in the vague perception of the market.
AI companies can't tolerate an AI bull with more than a dozen PhDs for a long time, and it took months to solve the AI needs of a certain company. Only by reducing development costs, realizing service standardization, and reusing industry cases, AI as an enterprise service technology can really have a market at all. The logic of pre-training large models once trained and reused multiple times, and the simplified automation of the model development platform are all aimed at achieving the goal of cost reduction with product capabilities.
From the current situation, AI is still the most certain core technology in the new round of technological change. But to make it worthwhile, companies must first be able to outperform cost pressures. In this race against cost, many companies must be down and countless bubbles will be blown out. American companies completed the commercial harvest of global IT software in the 1980s and 1990s, when it was also heavily invested, and the crowds were fighting, and finally a large-scale reshuffle. In the strategic space for Chinese companies to seize the software base of the AI era, they must also break through the layers of cost barriers. And in the end it must be the oligarchs who survive as the underlying platform.
Winning or losing the race against cost depends on three important elements of AI companies:
1. If there is no money to continue to invest, small and medium-sized companies will be directly trapped in this step. Problems such as salary cuts and layoffs have also been caused by this.
2. Have no willingness to continue to invest. This willingness includes both cost willingness and market willingness. It is difficult to bring a new technology to a specific industry, and it will even completely change the original business nature and working habits of the enterprise, and there will be many difficulties during the period. Huawei's establishment of the industry corps needs vertical command from the highest level, just to reduce the willingness cost of this matter.
3. Can we find technical methods to continue to invest in AI, provided that enterprises need to be clear in what situation can see the market turnaround, and this requires strong technical capabilities and technical paths. accurate. Big models, the combination of AI and knowledge, and the instrumentation of deep learning frameworks are the three most representative productization paths at present.
When the time comes to 2022, we can see that many excellent AI products and AI technology ideas continue to burst out, and we also see that AI companies and business departments that eat the old are retreating.
At high tide, everyone just wants to put the boat into the water as soon as possible.
Between the high tide and the low tide, the ship has the possibility of moving forward.
We see the retreat of AI companies, which may be the progress of the AI industry.