Another AI bull has left.
Recently, He Changhua, chief architect of computing storage at Ant Group, left his job at the rank of P10 in Ant Group.
He graduated from Stanford University and has 7 years of experience at Google, and later worked at Airbnb to work on application architecture for back-office systems. He has since joined Ant as Chief Architect of Ant Compute Storage.
The departure of He Changhua is not an isolated case, and the departure of Ai Daniu, a major Chinese manufacturer, has not been uncommon in the past two years. They either return to college or start a business. Such as Zhang Tong, director of Tencent AI Lab, Ye Jie of Didi AI Lab, Ma Weiying, vice president and director of AI Lab of ByteDance, Zhou Bowen, founder of JD AI, Min Wan, chief scientist of Alibaba Cloud AI Intelligence, Wang Gang, vice president of Alibaba and head of the autonomous driving laboratory of Damo Academy, Qi Yuan, former vice president of Ant Group and head of the AI team...
AI bulls have left, there must be a common reason and underlying logic, AI Business Week tries to find a "magnifying glass", pull out the "onion cortex" of AI bulls leaving, and dig the "hidden corners" under the iceberg of the event.
AI Lab of a Major Chinese Manufacturer: The "Duckweed" Between Industry and Scientific Research
Scientific research and industry are like two ends of the scale, one is today, the other is tomorrow, but most of the domestic manufacturers have not found this balance.
Most AI Labs were founded with the original intention of doing basic research and rarely touching the business, but with the continuous high investment of enterprises, it is difficult to directly bring benefits, and they have to adjust to support the industrialization of business.
Then, some trivial, low-threshold work, such as marking data, modifying models, adjusting parameters, etc., are naturally arranged on the heads of these researchers, and even some netizens joke that "there is no difference between them and Foxconn's craftsmen." This is contrary to the idea of researchers, since they cannot find their own value for a long time, it is better to return to academia and continue their studies.
Admittedly, there are also some large manufacturers whose AI Labs do not have business pressure, but carry the KPIs with heavy paper quantity and quality
Publishing papers has two direct benefits: on the one hand, it is a good PR material for enterprises to promote the leadership of technology; on the other hand, it can attract more research talents to join, which is of great benefit to recruitment.
But the problem is that with the sharp increase in the number of submissions to CVPR, ICCV, ECCV and other academic summits in recent years, the "inner volume" of papers has become more and more powerful, and it has become more and more difficult for papers to be selected. When inputs are not proportional to output, companies begin to re-examine the value of AI Labs.
The contradictions within the big factories are also a non-negligible obstacle, taking Tencent as an example, AI Lab research and development needs data from the business side, but the competition between different departments is fierce, and the business side is not necessarily actively cooperating. What the big factory needs is the aura of the big cow, and the big bull needs the economic return and data resources of the big factory, and once it does not match, it will be good to gather and disperse.
AI Lab is like an in-house advanced outsourcing. Lab's research is not necessarily needed if it wants to be marketed to business units. Not to mention that Lab's research results may not be able to accurately solve the pain points of business departments, some financial BG has its own research department, and there is no strong motivation to cooperate with AI Lab.
Take a look at how the labs of foreign tech giants operate.
Since its inception in 2010, Google X, the secret lab owned by Google's parent company Alphabet, has conceived hundreds of ideas a year for major challenges. Space lifts? Balloons online? Seawater refining fuel? Smart glasses? ...... These seemingly whimsical ideas are all ideas that Google X Labs once had.
Google X encourages, and even requires, regular exploration of absurd ideas. At present, Google X has successfully realized many ideas, such as Google Balloon and Waymo autonomous driving. Of course, there are also many ideas that ultimately fail, such as Google Glass smart glasses.
Google X's purpose is not to solve Google's problems, and it does not have a mission of charity. The ultimate goal of the organization's existence is to create world-changing businesses that could eventually become the next Google.
Operating innovation itself with innovative ideas and culture, optimistic about long-term value and returns, is the source of Innovation that Google X gushes endlessly, and it is also worth learning from domestic enterprises.
Where is the AI industry headed?
AI bulls have left the Internet factories, but also the industry reshuffle and innovation of the appearance, so what kind of trend will the AI industry have?
At present, the outlet of AI has passed or is nearing the end, and capital investment in AI will eventually need to land and return, on the whole, there will not be much breakthrough in technology in recent years. If CV is not Transformer, the defoaming of the entire AI field will be faster. There is no substantial progress in NLP at the moment, and it is unlikely that in the next few years.
In this case, the algorithm team in the core department of the enterprise is not worse than the AI Lab, or even better, because they have richer data for the scene, and the scene forces the innovation of the algorithm, including the improvement of algorithm efficiency and productization capabilities.
In the next few years, artificial intelligence will enter the era of high-speed productization, and data-driven deep learning algorithms are still an important driving force.
Of course, the future trend is "small data" rather than "big data". Recently, AI bull Wu Enda said in an interview with IEEE Spectrum: Big data, big models as a deep learning algorithm engine has been successfully running for about 15 years, so far, it is still dynamic. That being said, it only works for some issues, and there are a bunch of other issues that require small data to solve.
Small sample learning will enable AI to quickly adapt to new tasks and meet new needs for rapid iteration. Examples include the latest language models from OpenAI and Meta, WebGPT, and BlenderBot 2.0, which can retrieve the latest answers to the questions they ask online.
Over the past year, we have witnessed great progress in large language models, and the AI big model competition will continue in 2022. In 2019, OpenAI's GPT-2 became the first model with more than 1 billion parameters; in 2020, GPT-3 swept the AI community with 175 billion parameters, dwarfing everything before, but GPT-3's dominance as the largest AI model did not last long; in 2021, Google Switch Transformer model (1.6 trillion parameters) and Beijing Zhiyuan Research Institute "Wudao" (1.75 trillion parameters) model breaks down trillion parameter barriers.
These large models have greater versatility and reduce dependence on data. The market expects the size of large language models to continue to grow in 2022. The largest model of 2022 is likely to come from OpenAI:GPT-4.
Of course, as artificial intelligence penetrates into more and more long-tail applications, the demand for computing power doubles. Nvidia just released its 2021 Q4 earnings report, in which data center business revenue reached $3.26 billion, up 72% year-on-year, higher than expected of $3.15 billion.
It is foreseeable that the AI industry is a long-distance run, and small breakthroughs continue to burst out in technology, which is a comprehensive contest of productization capabilities, capital and talents.
Write at the end
The core of ai bull's departure is that the company overestimated the value of its innovative technology and realized the realizability of commercialization, and underestimated the difficulty of landing new technologies. AI bulls choosing to return to academia/return to traditional industries may be the future trend. From the traditional industry to the Internet express, the dividend of a wave of new technologies from scratch has ended.
In the next decade, Internet giants such as BAT will encounter great challenges. The traffic-led Internet entrepreneurship model has been terminated, while the technology entrepreneurship model with the "card track" model has just been opened.
In the new round of technology entrepreneurship, the advantages of Internet manufacturers are facing the impact of the previous future. In fact, we found that the big manufacturers AI bulls have left, while the startup AI big bulls are relatively stable. Some TO B Chinese technology companies that can benchmark salesforce, Oracle, etc. in the United States may rise.