Author |Wang Xin.
Edit |Liu Jingfeng.
The bubble in the field of artificial intelligence is always in a cycle of continuous production and bursting. ”
Melanie Marshall in her book AI30", which describes the cycle of 5 to 10 years in the field of artificial intelligence.
In 2016, after defeating Go world champion Lee Sedol, AlphaGo briefly set off a wave of artificial intelligence such as face recognition and autonomous driving. In 2023, the emergence of ChatGPT will make large models a well-deserved "top stream" in the AI industry. After the bubble burst and capital calmed down, the tide of irrationality finally receded. Nowadays, the AI industry rarely talks about face recognition again, and autonomous driving is difficult to land, and in China, large models have slowly become an area that investors cannot afford to invest in.
The group of people who study artificial intelligence is already familiar with this model: in the "spring of artificial intelligence", investment institutions over-promise and over-hype, and then usher in the cold winter of "artificial intelligence".
The world is not hot or cold. In the United States, the VC community is enthusiastic about investing in artificial intelligence. In Southeast Asia, where AI is embraced, China and the United States are competing for AI technology and investment – from 2020 to 2021, investors from the United States and China participated in 267 investment deals in Southeast Asian AI companies, accounting for 40% of the total investment.
A gratifying phenomenon is that in Southeast Asia, Chinese technology companies such as iFLYTEK, Huawei, and Hikvision are inextricably linked with the local area in Southeast AsiaChinese technology companies are growing into the mainstay of Southeast Asia's AI community.
Once, Pulitzer Prize-winning Thomas Friedman said in "The World is Flat" that the arena of the world has become flatter, and the flattened world allows every individual and region to stand under the same level.
Looking around the global AI community, we find that the world is not always flat.
What are the different directions of AI in different regions around the world?When AI companies go abroad, what opportunities and challenges will they face?Why wasn't ChatGPT born out of Tencent, Google and other big manufacturers with huge data sets?What is the inspiration and significance for AI start-ups?
Google Bay Area Headquarters. Source: Provided by the interviewee Qiu Zhen.
Now there is almost a consensus in the entire domestic investment community, that is, ".The investment heat for large models is cooling".
As the hottest track in the venture capital circle this year, large models have been hot, and countless large technology companies and AI startups have set off waves of large models one after another. Nowadays, the competition of global technology companies for large models is entering the era of stock.
According to IT orange data, as of the end of November 2023, there were 580 total financings in the primary market of the domestic artificial intelligence track, a decrease of 26% from 2022, and the total financing amount was 63 billion yuan, a decrease of 38% from the same period last year.
On the other side of the ocean, large-scale model financing in the United States is still in full swing. According to Crunchbase data, 11% of the investment in the VC stage went to the AI track last year, and as of the second half of this year, the proportion has increased by 26% in 2023, and 26% of the investment in the VC stage has gone to AIThe enthusiasm of the U.S. investment community for large-scale models has increased.
In the ocean and the east, AI financing is facing different situations.
Behind this, how are the differences in different financing environments caused?
First, in the United States, large models have reached a scale that can generate huge economic benefits and, to some extent, affect the macroeconomy. Some studies have found that this time the wind of large models blowing from Silicon Valley has contributed one percentage point to the growth of GDP in the United States. In China, although the war of 100 models is intensifying, the monetization of large models is still in the initial stage of exploration, and the problem of commercialization has always been an old problem in the last few rounds of AI waves, and it is difficult to find the best solution.
Second, there is no full-stack AI company with an application layer like OpenAI in China, which can go from the grassroots large model to the middle layer to the application layer. The field of large models, which relies heavily on the aesthetics of large computing power violence, is too money-burning and too high a threshold for startups, which can be called an arms race. Large models cannot be invested, but there are not many companies in the application layer, and the investment circle begins to think calmly, and everyone is in a wait-and-see state.
Foothill Park in Palo Alto, the heart of Silicon Valley, is adjacent to Stanford University and Steve Jobs' House. Source: Provided by the interviewee Qiu Zhen.
Qiu Chen, an overseas partner of Huaying Capital, believes that this reflects the difference in the original ability of technology between China and the United States.
Silicon Valley is still a driving force with original technology as the core and as the base, which is still very important, China's advantage lies in optimization, not necessarily original, that is, people have me, maybe people have me, and then scale, the last wave of deep learning-driven machine vision, it can be said that it is AI10, now the large language model is considered AI20, The current participation of these two waves in China is mainly in terms of optimization and scale. ”
After graduating from Peking University in 1997, Qiu Zhen went to the United States to study artificial intelligence, and after graduating from the Institute of Information Science of the University of Southern California, he joined Cisco, a legendary technology company in Silicon Valley. He continues to focus on AI venture capital in China, the United States, Southeast Asia and other places. Back in the millennium, Qiu Zhen was still working on technology in Silicon Valley, and it was the cold winter of AI at that time, but the entire Silicon Valley was still accumulating, patiently waiting for the next breakthrough.
But if we go to the other side of the ocean, we will see that generally speaking, after the next breakthrough in the United States, we will invest in it and do things that are optimized and scaled, so there will be a certain delay. ”
This pre-breakthrough accumulation process may actually require some patience, and after this stage has passed, we are really to the Internet. com that wave, is the middle layer equivalent to an interface can appear, this is the time when the application layer emerges in large numbers, in China may appear a large number of Didi and a bunch of other Internet companies, this time may still take a certain amount of time, may also need a certain amount of patience. ”
San Francisco Bay. Source: Provided by the interviewee Qiu Zhen.
Outside of China and the United States,Southeast Asia has also become the main battlefield for the confrontation between China and the United States.
In Southeast Asia, AI is still in the ascendant, and although the pace of implementation is relatively slow, there is still hope. Countries across Southeast Asia are embracing AI. The two AI whirlwinds of China and the United States are converging in Southeast Asia.
AI is not an emerging industry in Southeast Asia, and before the wave of large models, Southeast Asia focused on AI-related applications such as intelligent customer service and manual annotation due to factors such as low labor costs.
However, after coming to Southeast Asia this year, Zhou Chuanfu, vice president of iFLYTEK Cloud Platform Business Group, obviously felt that in addition to the original part of the traditional projects, now more departments and industries are now discussing and embracing new technologies such as AIGC and large models.
Although the most commonly used models in Southeast Asia are still large models from European and American AI companies such as OpenAI, in the Southeast Asian AI industry, there are also Chinese technology companies such as iFLYTEK, Huawei, and Hikvision. iFLYTEK and Alibaba have successively launched large models in Southeast Asia, the Southeast Asian national team has also joined the large model competition, and Singapore** has invested 52 million US dollars to support the AI multimodal large model development program (NMLP).
Zhou Chuanfu told Xiaguang Agency: "Although Southeast Asia is a region, when you open it up, many countries are very different. Singapore is the only developed country in Southeast Asia, and its AI implementation will be faster, and there are many applications in departments or industries such as education. Relatively speaking, Singapore (AI landing) is relatively advanced and more active, but countries such as Malaysia, Indonesia, Thailand and other countries can also deeply feel more enthusiastic than the original, of course, the whole landing pace is relatively slow, but there is still hope. We will continue to work in these places for a long time. ”
Vietnam also has the possibility of overtaking in corners in the competition of artificial intelligence. JPMorgan Chase has analyzed that Vietnam is at the "forefront" of the development of artificial intelligence in emerging Southeast Asia. As early as January 26, 2021, the "National Strategy for Artificial Intelligence Research, Development and Application by 2030" approved by the Prime Minister of Vietnam** made it clear that artificial intelligence will be developed into a pillar industrial industry.
In December 2023, Nvidia CEO Jensen Huang met with Vietnam's ** Prime Minister Pham Minh Chinh and promised to build a semiconductor base in Vietnam to make Vietnam a second home for NVIDIA.
Southeast Asia has become the first stop and strategic center of iFLYTEK's overseas business. In June 2023, iFLYTEK will hold a product launch conference and iFLYTEK AI TechDay Singapore Station in Singapore with the Xinghuo cognitive model and C-end intelligent hardware.
Because iFLYTEK has always focused on innovation and research of relatively underlying technologiesTherefore, if these technologies are used in more scenarios and more devices, it is not possible to rely on iFLYTEK alone. So the wholeDeveloper ecosystemIt took a lot of effort. Zhou Chuanfu told Xiaguang that in terms of the developer ecology in Southeast Asia, iFLYTEK has built an international station of iFLYTEK open platform centered on Singapore.
Actually, I'm very envious, and the point of envy isC-end cool products are easy to catch consumers, but the underlying technology may not know what you are doing for two days。So the whole ecological aspect is putTechnology application ecologyIt is our long-term layout, not relying on one or two years, at least 3-5 years to do a good job, which is the current state. Of course, we have also seen a lot of breakthroughs in Southeast Asia. ”
He believes that behind these breakthroughs, it is inseparableTechnological innovationwithLocalization
The complex historical and cultural background of Southeast Asia, and the transnational flow of immigrants have led to the formation of a multi-ethnic society with different dialect systems in Southeast Asia, and there are many dialects in various regions. For example, the official language of Indonesia is mainly spoken in Jakarta, but it is also found elsewhereVery many dialects。In order to do the core technological innovation of speech recognition and synthesis, it is necessary to improve the oral scenarios in different dialect regions in addition to improving the general ability. In addition, the identification of synthetic large models is also an important underlying technology for iFLYTEK to seek breakthroughs.
Localization is a challenge that Chinese enterprises must face when going overseas, and for iFLYTEK, this is also something that has to be done. Because iFLYTEK has to face many B-end user scenarios, there is no shortcut, and they have been ready to take deep roots for two or three years before they can do a thorough job in the local industry.
At present, the scale of iFLYTEK's Southeast Asian team is not particularly large, about 20 or 30 people, and native speakers account for about 40%. Most of the duties of local employees are in business development, marketing and other positions. In terms of core technology, it is still based on domestic headquarters and local technical support.
In-country teams and local teamsCollaborationprocess,Cultural differencesThe challenges are not to be underestimated. Zhou Chuanfu believes that in this case, we must put ourselves in the local country to consider the problem, "and we can't say how China is doing, how foreign countries are doing, this is actually a process of adaptation for us, it is very simple to say, but we must really put our own positioning in the local country." ”
In addition, all large-scale model enterprises must cross the two mountains of data and talent if they want to run through.
Let's go back to the question at the beginning:
Why wasn't ChatGPT born out of Tencent, Google and other big factories with huge data sets and talents?
Tencent and Microsoft data must be bigger than OpenAI, why does OpenAI have GPT, while Tencent, Microsoft and Google do not have GPT?This is because the effective use and collection of data is critical. This is especially true for startups, especially those who want to truly become an AI company, or just be an AI-enabled company.
Lounge area at Google headquarters in the Bay Area. Source: Provided by the interviewee Qiu Zhen.
Qiu Zhen believes that private property is not necessarily the source of value itself. "You have to have data to be able to actually use AI, but private data doesn't have to be valuable. This should be understood by startups. ”
To accumulate and collect data, be sure to consider algorithms. For example, building a data stack platform, but its threshold is actually very high, so the entry threshold for collecting and sorting out data alone will block many people from the outside. Even the data of the giants may not be all useful.
When it comes to the AI arms race under the wave of large models, volume computing power may be important, but the core algorithm talents are the most important resources in this round of competition.
For startups, Qiu Chen's advice is to start looking for some core talents for large model algorithms from now on: "As a startup, I may have to search for some real core talents today, which may sound a bit far away, but I think my suggestion is likely to be useful." For some startups, if you start today, no matter what you do, or even if you just do the application layer, you must pass the data level, but it is useless to have data alone, and in the end it is likely to depend on your algorithm, no matter what algorithm you are, you may not need to touch the base model, but even if you want to do fine-tuning, or even just to adjust the API, you will need a deep understanding of the training algorithm, and the most important armament is actually talent. ”
This is also because of the current shortage of domestic large-scale model talent reserve resources. Liu Chenghui, CEO of Doctor AI, once said to **At present, 90% of the talents who make pedestal models in China are from Tsinghua University, and there are no more than 200 people who can really tune and train models in China. "And the war of large models has also made the cost of employment higher,Vivo vice president said in an interview with **:"The vivo model now has an annual input cost of 2.03 billion yuan, and the total investment cost has exceeded 20 billion yuanTalent and data computing power account for half each, and the average cost of talent is 1 million yuan per person after tax. ”Finding core talent is especially critical for startups right now.
How to find their own position in globalization and give full play to their own advantages is a new topic that all AI companies need to face.