76 Rules for China's Artificial Intelligence

王达业

76 Rules for China's Artificial Intelligence [Copy link]

 

Source: Leifeng.com

If one word is used to sum up China's artificial intelligence, it is "conflict."

Whether in industry, investment or academia, all their actions and processes revolve around conflict.

1. The greatest danger in turbulent times is not the turmoil itself, but still doing things with the logic of the past.

2. If companies want to win the AI race, resolving cognitive conflicts, cultural conflicts, and methodological conflicts between teams from different backgrounds is far more important than the business itself.

3. Yitu Medical has become an abandoned child, which is a microcosm of the collective problems in the business model of AI companies. Success or failure is the ultimate fate that entrepreneurs must accept. Perhaps, Yitu's story will continue, but there will no longer be Yitu in the medical AI world.

4. Four soul-searching questions about the collapse of AI business models: How did the dream of AI standardization and generalization fall apart? Why don’t highly customized solutions work? Why can’t we learn from overseas high-profit solutions? What are the three radical models for AI companies to break out of the low-profit dead end?

5. To make customization work, we must achieve the ultimate "three low costs": low average labor cost, low operating cost and sales cost, and low marginal cost after production expansion. Use the lowest cost to save more profit space.

6. The reason why Yuncong was able to pass the review first, in addition to its national team attributes, is its loss rate control. This has a lot to do with labor efficiency or the low labor cost in Chongqing. In 2019, the total salary of Yuncong Technology's core executives was only 8.9047 million, while many companies have scientists with annual salaries of tens of millions.

7. Internet giants use high-cost manpower to run customized projects, which is like using a rifle to shoot flies. The input-output ratio is very mismatched. Internet companies that are used to doing big business for a long time will not engage in the business of picking beans and picking sesame seeds for a long time.

8. Most AI companies cannot escape the "three highs" (high investment, high losses, and high talent), and at the same time they do not have economies of scale, which makes them unable to make ends meet in the highly customized and low-profit market.

9. What is a solution? Lou Gerstner once defined it as, if a customer needs a toilet, IBM will sell it.

10. In its heyday, IBM seemed to be working hard on highly customized solutions, but in fact, it had a firm grasp of the most core standardized middleware, so that a project could get very high gross profit. Similarly, SAP has ECC and Oracle has databases. AI companies still do not have such products, but they mistakenly believe that AI technology itself is middleware.

11. Domestic AI and even enterprise software solution companies lack high-gross-margin middleware products due to the following reasons: limited complex system engineering capabilities; lack of determination and sway in capital-intensive projects, and a desire to make quick money; insufficient effective patent protection, insufficient legal measures, and limited market penalties, resulting in low costs for plagiarism and the inability to establish barriers.

12. Most patents of AI companies have no practical significance, formalism outweighs pragmatism, and valid patents lack strong protection.

13. A first-class technology company not only needs to be able to create IP, but also needs to operate IP through legal means. In its most difficult period, IBM achieved 25% of its total profit income from intellectual property and licensing projects through a series of organizational structure and patent licensing reforms. Oracle's legal team is the most powerful department in the company. Some people joked that it should be a large law firm, not a software technology giant.

14. SaaS is in a rather awkward position in the domestic business environment. With the emergence of DingTalk, WeChat, and Feishu, the “freeloader” mentality of domestic small and medium-sized enterprises has been further exacerbated, indirectly leading to the gradual evolution of SaaS into a large-customer customized project, returning to the old path.

15. Taobao exists in the market in another form of SaaS, making money from most small and micro enterprises by making the wool come from the pig, which makes it extremely difficult for small and micro SaaS enterprises.

16. The future of AI may turn around in the following three paths:

Path 1: Re-customize integrated project implementation → Digital consulting → Consulting business feeds back to re-customize implementation → Establish high stickiness and system irreplaceability with large customers

Path 2: Re-implement customized integration projects → Enter the standard market of non-standard markets (autonomous driving, chips) → Form standardized products → Low-cost large-scale replication

Path 3: Open source deep learning framework, occupying the high ground of localization

17. As the business of today's AI companies continues to sink, they have essentially become digital companies, so there is a need to set up a consulting department. For digitalization, it is not important what technology is used, but what is important is to do a good job in top-level design, organizational management planning, data precipitation, data circulation, data decision-making, and ultimately achieve intelligence.

18. AI consulting is just an introduction. When you get through the customer's decision-making level, you can attract customers for your own project implementation business. Through this introduction, you can understand the industry and more customer needs, making the solution more sound and universal.

19. In fact, both TensorFlow and PyTorch have the potential to generate tens of billions of dollars in AI revenue. The reason why Google and Facebook do not make profits from this is that open source carries more strategic significance and is a defensive infrastructure measure to prevent being swallowed up by opponents. Android's free open source, from a strategic perspective, is to prevent being strangled by iOS and Windows.

20. Grouping is a common phenomenon in the academic community. Sometimes, appropriate grouping is often conducive to the unity, communication and integration of the academic community. However, if there are too many mixed interests, the bad money will drive out the good money, and ultimately fail in "inheritance".

21. Peer review is an important means of evaluating academic achievements. Any major theoretical breakthrough can only play its greatest role after being recognized by peers. However, the basis of peer review is built on credibility. In layman's terms, it means not letting anyone off the hook.

22. The earliest group of scholars who successfully published papers in the field of deep learning by adjusting parameters and flooding the papers have formed interest groups and have in fact mastered certain academic resources. They have the right to review papers in conferences, forums, and journals, and can decide whether some non-innovative papers (watered-down papers) are accepted.

23. Repeated parameter adjustments, patching up the original network, lack of theoretical guidance, and mostly "writing how but not why" are all common methods used by scholars to flood their papers. They are even proud of this and often compete with each other in the number of papers they publish.

24. Some of the "water-flooding scholars" have already graduated with a doctorate, entered the academic arena, and become tutors. Their academic style has influenced their students, and then these students have graduated with a doctorate and also entered the academic arena... At present, there may be a second or even third generation of these "water-flooding scholars" who have become tutors.

25. When companies and universities work together to conduct research, most of them only take what they need and are relatively short-sighted, focusing only on immediate benefits. Companies often state in contracts that interns will produce results in three months and professors will complete their tasks in one year. However, truly "breakthrough" research generally takes 3 to 5 years or even more.

26. In the past few years, there was a rush to promote deep learning, but now there is a big push to cool it down. This is actually a sign of immaturity. Even if there has been no progress in the theory of artificial intelligence recently, it does not mean that there has been no progress in artificial intelligence.

27. The imperial examination system of the time has "mutated" into hats, professional titles, and status. If you get a certain hat, you will have corresponding material resources, and use these resources to continue to exchange "hats"... This has led to talented young people trying to increase the number of papers in order to cope with professional title assessments and school awards.

28. Regardless of the number of papers or citation data, as long as this KPI is set for scholars, some domestic scholars will definitely find ways to solve this objective function.

29. Scientific research needs time to be tested. Now, formal AI academic conferences will set up a time test award. The purpose is to go back ten years ago to see which papers can stand the test of time. Therefore, academic freedom is required, and we cannot rely on numbers or citations.

30. Enterprises are afraid to recruit AI scientists because the existing organizational capabilities of most companies are not able to absorb the talented scientists in the ivory tower. People often attribute all organizational problems to people.

31. The fear of AI scientists in companies is a reflection of their lack of confidence in their organizational capabilities.

32. When second- and third-tier companies found that they did not have the organizational capacity to absorb chief scientists and stopped recruiting one after another, the balance of talent supply and demand began to tilt. With more meat and fewer wolves, the devaluation of some senior AI talents was a natural thing.

33. The effect of AI research is subtle and silent, and cannot be directly demonstrated, let alone KPI-based. It is not realistic to let scientists directly develop solutions and run business. Science seeks the optimal solution, and the essence of ToB products is a compromise with the optimal solution as the goal: a compromise between standardization and customization, a compromise between high gross profit and loss, a compromise between you and the customer...

34. Scientists are not good at compromise. In their local world, they are always the MVP. The scientific community does need such a paranoid and conceited spirit. But back to the industrial world: most managers do not believe in the methodology of local work.

35. Enterprises have two evaluation criteria for AI scientists: internal and external. Internally, they help enterprises solve practical business problems. Externally, they establish connections with the outside world, use their personal appeal to recruit more top talents, and reach cooperation with external top research institutions, which in turn feeds back their own technical capabilities and industry technology influence.

36. For Internet giants, the value created by face is incomparable. Face-oriented companies don’t last long, and substance-oriented companies can’t grow big. For those big companies that are doing the best and have both face and substance, in their eyes, the face of AI is sometimes more important than substance.

37. Alibaba Damo Academy was originally a basic technology research institute, but now it has completely become an application research institute for Alibaba Cloud. The purest corporate AI laboratory in China is Tencent AI Lab led by Zhang Zhengyou, which has invested absolute manpower and material resources to conduct cutting-edge research. Zhang Zhengyou became the first level 17 expert in Tencent's history last year, which shows Tencent's determination to conduct long-term cutting-edge research.

38. In small and medium-sized enterprises and traditional IT companies, the potential organizational conflicts brought about by the introduction of AI scientists are sometimes more valuable than the actual business problems they can solve.

39. Over the past five years, AI research institutes of small and medium-sized enterprises have proven a truth: if AI talent is used too pragmatically, it is like using a cannon to kill a fly. This is not only wasteful, but also not necessarily effective.

40. The driving force of a company's sustainable development is not talent, but the distribution of benefits. AI scientists are paid too high a salary in the company, which challenges the existing salary structure in the eyes of human resources executives. Business executives think that they have been working on the front line of business for more than 20 years and have made great contributions to the company, but they have not enjoyed such treatment, so they naturally feel unbalanced.

41. It is not uncommon for AI scientists to hire their own students at high salaries in many companies. The various special channels opened by companies for scientists will naturally bring trouble to themselves.

42. Once an enterprise discovers the problem of low per capita input-output ratio of AI, it will set a revenue KPI for the person in charge of AI research for the first time: the conflict between the AI research department and the business department comes from this.

43. There are two ways for AI research institutes to make money. One is to work with the business department to outsource for customers and take a commission from the project funds. The other is to do in-house outsourcing for its own product department and collect money from the business department. However, the in-house outsourcing method is prone to the situation where one's own people cheat one's own people.

44. The AI Research Institute does in-house outsourcing. Due to the huge revenue KPI pressure and knowing that the business department cannot outsource due to compliance issues, it asks for an exorbitant price. Because there is no party A/B relationship, the former's arrogant attitude is common. At the same time, they will attack the product department for not cooperating, the IT department for bad code, the data department for poor sample quality... The conflicts between the AI Research Institute and the business department broke out in these little things.

45. The AI research institutes of Internet giants are independent of the engineering department and have great say. But now CEOs are increasingly inclined to let engineering lead the management of the research institute. For example, Andrew Ng used to report to Robin Li. After he left, the AI research team was led by Wang Haifeng of the engineering system. After Fei-Fei Li left Google, engineering commander Jeff Dean took over Google's artificial intelligence research. After Harry Shum left, Microsoft's CTO took over its research team of more than 6,000 people.

46. In the past, CEOs with technological foresight hoped to form a top-down chain of AI research driving engineering, engineering driving products, and products empowering users/customers. In fact, it is found that the current AI cannot drive engineering, it is only a part of engineering and assists in implementation. Independent AI research institutes that used to have a lot of say have gradually become in name only and have become sub-teams of the engineering department.

47. As representatives of corporate research institutes, Bell Labs and GE Central Research Institute divide technology research and development into two categories: basic research and technology upgrade or transformation. These two types of research can be independently carried out by different types of scientists. The cycle from basic research to product development and then to market launch is particularly short. This efficient collaboration depends on the talent supply of a large team and the innovation of management models.

48. The success of the corporate research institute is due to two words: freedom. The free basic research atmosphere has allowed Bell Labs to contribute to the world transistors, lasers, Unix, C language and other great inventions that have changed the world. This freedom also means that it is only restricted by peer review, not by assessment pressure.

49. Replacing “freedom” with “performance appraisal” is a disaster for AI research institutes: After the split and Wall Street’s intervention in the 1980s, Bell Labs, which was burdened with performance appraisal, could no longer recover its vitality, and scientists even faced the embarrassing situation of having to work with the marketing department to promote products. In contrast, the various dilemmas faced by many corporate AI research institutes today are exactly the same as those of Bell Labs at that time.

50. In terms of academic research, domestic AI research institutes have failed to produce sufficiently influential results. Although they have published many papers at top conferences, they rarely have completely independent ideas. They basically follow the improvement research done by companies such as Google and OpenAI, such as BERT, GPT-3, AlphaFold, etc.

51. The best research is to prune the existing knowledge tree rather than develop it. Today, artificial intelligence is still in the stage of flourishing. If it encounters a bottleneck one day, it may be possible to consider "pruning the existing knowledge tree", that is, to abstract the most essential mathematical concepts of intelligence in theory, so as to bring about the next leap forward in the development of artificial intelligence.

52. Companies have two models to choose from when setting up AI research institutes: the MIT model and the Harvard model.

53. MIT has a heavy workload and strict assessments. Its graduates must be able to master professional knowledge and skills in a certain field and have the ability to learn independently. The quality is stable - the level is relatively average. Harvard University's educational philosophy is more free and open, with fewer assessments, and it encourages students to have a wide range of interests. The quality is unstable - the level is uneven, although the average level is also high.

54. Small and medium-sized enterprises are always struggling on the brink of life and death. They must first ensure product quality, and it is not the time to deliberately pursue originality, which is the MIT model. For large companies, their asset strength can ensure long-term survival by relying on existing businesses, and at the same time can support basic research that requires facing a large number of failed attempts, which is the Harvard model.

55. The venture capital circle in China in the early days was like a giant baby country. This circle seemed to be prosperous on the surface, but the core was not fully cooked: lack of independent thinking, imitation and plagiarism, disorderly industry, and failure to grow up.

56. Track-based investment is nothing more than seeing which European and American companies are growing rapidly, and then investing in a few Chinese companies that have the shadow of American companies. This is no different from plagiarism.

57. The investment logic of the track-based approach is similar to the technical principles of deep learning. Set a goal, collect a large number of similar samples (projects), increase computing power (funds), conduct brute force trial and error, and finally wait for an optimal result.

58. AI investment and financing is probably the first area where domestic VCs are not acting like giant babies. Europe and the United States have not seen any successful AI investment cases like the Internet. However, domestic investors have resolutely entered this field without any reference coordinates, spending the world's largest amount of 200 billion to invest heavily in AI.

59. The plan set up by investors seems to provide a lot of positive resources for the enterprise, but on the actual development path of the enterprise, they may be "negative resources".

60. Some companies have introduced third-rate scholars, but they brag about their top-notch achievements to the outside world, not realizing that they are putting themselves in a big slap in the face. Under the flattery of the companies, these second-rate scholars will subtly regard themselves as masters.

61. There are no more than five corporate scientists in China who have outstanding practical capabilities in academic, engineering and product fields. Some IEEE Fellow-level experts can get an annual salary of 30 or 40 million yuan, and some can get several million yuan.

62. As long as the stories of star scientists and star entrepreneurial teams are told convincingly, VCs will be willing to pay for them in the next round. This method was effective between 2016 and 2018. After VCs tasted the sweetness, they personally stepped in to paint a picture for the next round of financing, while trying their best to attract star investors to join. The star investors who were attracted also became part of the game and could call on more people to join. AI investment has become a game within a game, with leverage leveraging leverage in a continuous cycle. The end of the game is still a game. There are no products or commercialization.

63. What are the advantages of emerging AI companies over established companies? Many people say it is technology. That is not entirely true. It is the high cost advantage. The high cost approach is quite beneficial from the perspective of customer acquisition. An expensive system is sold at a bargain price. Once customers are curious about it, they will naturally give it a try. But in the long run, it is just a temporary solution.

64. In order to win customers, AI Xiaolong gave salesmen a bonus of one million yuan for every police station he signed up. So the salesmen pushed the products like crazy, regardless of whether the products were easy to use or too cheap, as long as they could occupy a market, they would be considered successful.

65. A senior executive of Hikvision once said a very classic saying: We cannot do a business of making 100 yuan with 1 yuan, but no one can do better than us in the business of making 1 yuan with 1 yuan.

66. Investors originally wanted to make a business of 100 yuan with 1 yuan, while AI companies are doing business of spending 3 yuan and only getting back 1 yuan. Although it is difficult to make 100 yuan with 1 yuan, at least like buying lottery tickets, the investment is controllable and you can buy a good hope.

67. Whenever a company's actions are slightly distorted, investors will intervene in strategy, management, personnel, finance, and public relations to control the direction of the company. The conflict between the two business philosophies began to erupt around 2018. As a result, those companies that failed to convince investors simply died on the road and became the first batch of cannon fodder.

68. The AI company's high-profile move to sell cameras through hardware has put a lot of pressure on channel dealers, forcing them to take sides and only stand on the side of major customers, so they had no choice but to give up the AI company's orders, resulting in a backlog of hundreds of millions of goods that Xiaolong has been unable to sell.

69. AI companies are caught in a dilemma due to the tug-of-war between the two situations. Behind the two situations are two business cultures: one is a capital-oriented business logic, and the other is a business-oriented business logic. The two are not compatible. When the conflict between the two cultures becomes more intense, factions will form and internal friction will increase.

70.The first principle of AI business model: make it heavy downwards, light upwards, consult forwards, and operate backwards.

71. Now, 90% of AI companies follow the model of providing what customers need and making highly customized hardware and software solutions. However, the market value of 3 billion is the first ceiling. To break through the 3 billion ceiling, it is necessary to condense and extract general middleware products from customized solutions.

72. It took SAP more than 20 years to complete this path, from large-customer customization to continuous divestiture to become a pure software company. During the more than 20 years of standardization process, it also designed the ABAP low-code language and established a third-party cooperation ecosystem such as strategic/IT consulting, integration, deployment, data cleaning, and operation and maintenance to ensure that its standardized software has an ecosystem to support its implementation.

73. The current AI industry lacks the role of operating service providers. For example, old hardware companies are real estate developers, and AI software companies are decoration companies, but the only thing missing in the AI industry is third-party property companies.

74. In the past, the two most profitable directions of AI were security and data services. The former involves hardware purchases, with high revenue but low profits. The second is data services, which have the lightest model and considerable revenue. The ROI obtained by trading desensitized data services is much higher than the benefits brought by technology implementation. The two models have a common feature, which is to walk on the line of compliance standards. One is the use of private information, and the other is the indirect trading of desensitized private information.

75. The AI field is like playing cards, and most investors have bad cards in their hands.

76. When investors are not lucky enough to be dealt good cards, all they can do is to make local optimization as much as possible and play their own bad cards a little better than other people’s bad cards at each step.

qwqwqw2088

It's a bit too much to call these military regulations