news

Image source: Visual China

Contributing author | Zhou Zihao

Editor | Wang Weikai

Editor’s Note:

sinceOpenAIHaving opened the Pandora's box of generative AI, big models will still be the hottest trend in 2024. As the birthplace of technology, Silicon Valley is swarming with talents. Any excellent organization must have found a method and corresponding tools to maximize the creativity of each individual. But more importantly, it is to seize those fleeting genius inspirations. "AI Light Years" records the changes in big model companies at home and abroad by visiting some AI company employees: Where are the talents who will master the future? This is the second article.

It’s nearly 40 miles from Sunnyvale to downtown San Francisco, and Zhao Ming, a master’s student at Stanford University, is going to attend an AI venture capital event. This is Zhao Ming’s activity trajectory on a weekend evening in June.

The event was hosted by a partner of a local US dollar fund in Silicon Valley. The attendees included founders of dozens of AI startups and nearly 50 investors representing different capitals. During the event, entrepreneurs went on stage to sell their ideas to investors, while AI engineers watched the "market" from the audience.

"Such scenes happen every week in Silicon Valley," Zhao Ming told the author of AI Light Years. He got a ticket to the field thanks to a friend's introduction. Zhao Ming majored in computer science at Stanford University and is now determined to enter the AI ​​industry.

He expected to meet more technology practitioners in this event to get cutting-edge information about the industry, but he was surprised that half of the more than 400 participants were Chinese.

As the glasses flowed, the identities of these people gradually became clear: there were venture capitalists who flew from Beijing specifically to seek investment opportunities; there were also Chinese software engineers who had worked for many years at Google, Microsoft and other companies; and there were also some Chinese students studying AI at top American universities.

They are all microcosms of the Chinese in Silicon Valley. A report by the Macropolo think tank in the United States (2022) showed that among the top artificial intelligence researchers working in the United States at that time, 38% were from China and 37% were Americans. In 2019, these two figures were 27% and 31%.

Chinese people have always been the backbone of Silicon Valley’s technology circle. Now, under the new AI wave, what kind of ups and downs are young Chinese people in Silicon Valley experiencing?

“Go to Silicon Valley, the place closest to new productivity.”

"Many of the technicians working on AI in Silicon Valley are Chinese," Qin Tian told the author.

He is a machine learning engineer (MLE) at a well-known technology company in Silicon Valley. After graduating with a master's degree in computer science from a top 30 university in the United States in 2019, he began to engage in AI-related algorithm research in Silicon Valley. In his team, Chinese engineers account for nearly 40%, "the others are mainly from India, and some are Eastern Europeans from Romania and other countries."

According to the Silicon Valley Index report, the population of the Silicon Valley area is mainly composed of Asians and whites, and among the Asian population, the Chinese account for the largest proportion. In 2022, the Chinese accounted for 31% of the Asian population, becoming one of the largest Asian groups in Silicon Valley. Previously, many of Qin Tian’s Chinese friends were software engineers.

But at the beginning of 2022,OpenAIThe release of the latest versions of GPT3 and DALL-E 2 broke this situation and many people began to transition to AI.

Take Qin Tian as an example. Earlier, his business line was developing a mobile camera application with AI functions. In June 2022, the company suddenly announced a strategic adjustment, cutting off the product line with a history of nearly five years and turning to the field of large models.

In fact, this kind of adjustment is not uncommon in 2022. That year, many Silicon Valley technology giants such as Meta, Adobe and Google made major business adjustments and developed the field of AI big models one after another, following the footsteps of OpenAI. This shift also prompted a large number of software engineers to flow into the artificial intelligence business sector, including many Chinese.

Zhang Yu, who graduated from Carnegie Mellon University, is currently working at Google's parent companyAutopilotcompanyWaymoZhang Yu introduced thatChatGPT-4After the emergence of the virus, Waymo experienced a wave of resignations, and some Chinese colleagues also chose to leave. "The resignation rate is uncertain, and some groups may lose one person every week," said Zhang Yu.

Most of these colleagues have switched to technology companies engaged in the development of large language models, or returned to Google to be responsible for the AI ​​development of other products. Zhang Yu lamented: "In Silicon Valley,GPTIt has become a more popular field than autonomous driving.”

Those joining this AI wave include not only "old engineers" who have worked in Silicon Valley for many years, but also "newcomers" from China who have come to catch up with the trend.

Shiqi, 25, is one of them. In May this year, he flew from China to San Francisco and decided to join a startup focused on AI. Shiqi is not sure how many people have come here from China like him this year, but he believes: "Talent and capital will always flow to places with the most advanced productivity."

Shi Qi's decision to come to Silicon Valley was not an impulse one. Before that, he was a software engineer. After graduating from college in 2021, he worked in a large domestic Internet company. At first, Shi Qi, who had just entered the workplace, hoped to "show his skills" as much as possible in the business line. But later, he felt more and more that the actual work was different from his expectations.

Around 2022, the product line Shiqi worked on fell into a growth stagnation: no matter how much product experience optimization was done, the number of users no longer had a breakthrough growth. "The stock era seems to have really come," Shiqi felt that she had entered the "retirement state" early, maintaining the system every day, making some code modifications according to needs, and then it was over.

Working from 10am to 6pm, getting off work on time, and no longer working 996, Shiqi feels that she is surrounded by dangers. Although she is "free", it also means that "it's the same with or without you".

This situation prompted him to think about working part-time. By chance, Shiqi passed the interview of an AI startup in Michigan, USA, and began to work remotely for it. In the spring of 2022, the boss of the company suddenly sent Shiqi a message saying that the company had obtained a new model, which was OpenAI’s latest InstructGPT, and invited him to participate in the internal test.

"This is completely different from the AI ​​I've seen before," Shiqi exclaimed. Although GPT-3 was released as early as 2020, InstructGPT uses reinforcement learning from human feedback (RLHF) to generate responses that are more in line with human preferences and intentions. Shiqi had a hunch: "It might bring something different."

Since then, Shiqi's company has started to use this language model on a large scale to develop new products. This also made Shiqi feel the gap between China and the United States in the promotion of AI: at that time, few software engineers around Shiqi had ever come into contact with GPT, and no one in the office talked about papers about transformers. But in the United States, OpenAI has become a hot topic among major technology companies.

Shiqi recalled that the earliest batch of engineers in China who paid attention to GPT loved to visit GitHub (open source community). At that time, some mainland engineers began to access GPT's API and make some shell applications. But it was not until the release of GPT-3.5 at the end of 2022 that domestic large models began to be sought after by capital. In 2023, it was also called by the media as the "Hundred Model War".

At the end of 2023, Shiqi chose to resign. With his previous experience of remote work, he was invited by his current company. "Go to Silicon Valley, the place closest to the new productivity," he said.

Big companies compete on models, while startups compete on the B-side

Since the company went all in on AI, Qin Tian has felt increased work pressure.

Before the summer of 2022, the working atmosphere was relatively relaxed. According to Qin Tian's understanding, "this is due to the competitive barriers left over from the previous information age." As the person in charge of back-end development, he gets off work on time every day, and the pressure to release new products is not great. "Old Internet products maintain revenue, and it is normal for new products to be released after 2-3 years."

Nowadays, major technology companies in Silicon Valley dare not slack off in the AI ​​competition. The adjustment of business lines has made Qin Tian's daily life become tight. Getting off work at 11 pm has become the norm, and working overtime on weekends has become inevitable.

OpenAI's progress on AIGC was too fast, which put a lot of pressure on other Silicon Valley companies. In February 2024, after OpenAI released Sora, the stock price of Qin Tian's company was hit hard. In just one spring, it fell by more than 30%. In the previous year, the capital market was still confident in the company's strategic adjustments, and for a period of time, the stock price even rose by more than 60%.

Faced with competition from large companies such as OpenAI, Qin Tian's company has accelerated the pace of product updates and shortened the release cycle to at least once every six months to ensure that the company remains competitive in this wave of technological innovation and avoids being left behind.

"Product iterations need to be accelerated, and at the same time, more cutting-edge papers need to be studied." Qin Tian said that a new technology arms race is taking place in Silicon Valley. Qin Tian is mainly responsible for the development of multimodal large models (MLLMs). Simply put, it allows the model to accept inputs of language, video, audio and other data at the same time, so as to understand and generate content more comprehensively.

The multimodal field is a new mine waiting to be explored by AI engineers. An industry paper published in May showed that reducing resource consumption while expanding its applicability and minimizing performance degradation is the main research and development goal of MLLMs.

Recently, China Intelligence and a number of domestic universities proposed a multi-task long video understanding evaluation benchmark, MLVU, to evaluate 20 of the latest popular multimodal large models and found that the single-choice accuracy rate of the top-ranked GPT-4 was less than 65%, and existing models still face huge challenges in long video understanding.

“We need to keep an eye on whether new technologies emerge to deal with this new field,” Qin Tian said. For example, when the DIT (Diffusion Transformer) algorithm architecture used by Open AI generates better videos, Qin Tian and his colleagues will immediately track it to determine the direction of product iteration.

Large companies are forging ahead in the big model race, and the densely populated start-ups in Silicon Valley are also setting up their own AI camps.

According to Shiqi's observation, many Silicon Valley startups have B-end product lines, which may be related to the background of the bosses of these companies in large companies in the past. "They can use the connections they have accumulated at work to open up business channels more quickly," Shiqi said.

Today, Shiqi and his team are developing an end-to-end developer tool for enterprises to improve programmers' work efficiency through AI. Last September, Microsoft also launched a similar application "copilot". They hope to achieve such a scenario: developers can ask these programming assistants when writing code to complete their work more efficiently.

"Improving technology efficiency is the direction of competition for many startups in the United States," said Shiqi, and end-to-end means creating an automated workflow. The entire process is completely implemented by AI, and customers only need to pay attention to whether the final result meets expectations. "If we want to overtake others, we have to think more about differentiated competition with these big companies," said Shiqi.

However, some technical challenges still face Silicon Valley engineers. In Shiqi's view, improving the accuracy of the product is a technical point that they urgently need to break through, and different user scenarios have different accuracy standards.

For C-end products that focus on user emotional companionship, 70% accuracy may be enough. But for enterprise-level needs, 80% is their goal. This also forces them to find more accurate knowledge base data for model training.

In order to compete, it is inevitable to burn money. For Shiqi's company, every time they serve a company, even a small project, the monthly inference cost is thousands of dollars. "Such a single project cost is equivalent to the monthly salary of a full-time algorithm engineer in a developing country."

Faced with high costs, Shiqi frankly stated that the company has not yet achieved profitability, but recently several large-scale companies have signed service contracts with them, and they expect to achieve profitability in a year. "At least payment is a positive signal." Shiqi said.

Layoffs, entrepreneurship, and returning to China: Silicon Valley’s trio

According to the Silicon Valley Index, from 2019 to 2023, nearly 3,000 startups were born in Silicon Valley and San Francisco, reaching a record high. Among them, many Chinese CEOs can be seen. Currently, four AI companies founded by Chinese - Scale AI, Cognition AI, Imbue, and Cresta have joined the ranks of unicorns, with a total valuation of more than US$18 billion.

In Silicon Valley, there is no shortage of legends of people getting rich quickly through entrepreneurship, but entrepreneurship also means risks under fierce competition. Among Qin Tian’s friends, Chinese engineers who leave large companies to start their own businesses are still a minority, and more people choose to wait and see.

At the same time, although AI has triggered a new wave of enthusiasm in the capital market, Silicon Valley has experienced a continuous wave of layoffs in the past two years.

Aria, who works in Silicon Valley, told the author of AI Light Years that many of her Chinese colleagues were affected by the wave of layoffs. According to her, the laid-off engineers were mainly engaged in basic code writing or debugging. Obviously, this is a type of work that can be easily replaced by AI.

According to estimates by Layoffs.fyi, an employment tracking agency, more than 400,000 tech jobs have been cut since the beginning of 2022. Another data is that the total number of global layoffs by US technology companies exceeded 500,000 between 2022 and 2023. This is the largest batch of layoffs in the technology industry since the 2008 financial crisis.

Qin Tian also clearly felt that there would be far fewer job opportunities in Silicon Valley in 2024 than in 2019.

At the end of last year, the annualized inflation rate in the United States was once below 3%, which was almost halved compared to 6% at the beginning of 2023. Silicon Valley employment analysis company CGC explained that after the inflation rate fell, technology companies had little room to raise prices and could not improve performance through price increases, so layoffs and spending control became the most direct option.

"A Chinese engineer who was laid off by Google came to our company for an interview," Shiqi told the author of AI Light Years, but his company did not hire him in the end. According to Qin Tian, ​​in Silicon Valley, the average annual salary of an AI algorithm research engineer who has worked for about three years is about $450,000, which is a considerable expense for many startups.

"The wave of layoffs has made many Chinese people who have entrepreneurial dreams afraid to take action," said Qin Tian.

On the other hand, the strict visa regulations also make them lack the "freedom of choice." According to U.S. immigration laws, before obtaining a green card, unemployed people holding an OPT internship visa or an H-1B work visa must find a new job within 60 or 90 days, otherwise they will face the risk of being deported.

If the business fails, unemployment may make it impossible for them to stay in Silicon Valley. Aria believes that now is not a good time to start an AI business. Like many Chinese AI companies, in Silicon Valley, "how to commercialize AI products is still a problem."

Across the ocean, China's enthusiasm for AI seems to have just reached its climax. Not long ago, the World Artificial Intelligence Conference just ended in Shanghai. Being in a foreign country and limited by work energy, Qin Tian pays less attention to the new developments in the domestic AI industry, but he believes that domestic scholars are already very advanced in AI algorithm research.

In the world open source model rankings released by Hugging Face, the world's largest open source large model community, many open source models from China once occupied the top ranks.

But in the short term, Qin Tian has no plans to return to China for development. He believes that computing power is crucial for the training of large models, and the lack of advanced computing power has always been a problem that Chinese AI companies have to solve.

In August 2022, Nvidia received a notice from the United States requiring it to stop exporting A100 and H100 chips to China.

On July 17 this year, Bloomberg reported that the Biden administration is considering whether to introduce a strengthened regulation that allows controls on foreign products that use any U.S.-made technology. It will target the Chinese operations of companies such as Japan's Tokyo Electron and ASML, which produce chip manufacturing machinery that is critical to the industry.

Geopolitical factors have caused many AI technical talents in Silicon Valley to have doubts about their employment and development after returning to China.

"Even if the job competition is fierce, you can't give up going to Silicon Valley." Zhao Ming, a master's student at Stanford University, lamented to the author that many AI master's students in the United States choose to stay in the United States as their first choice. For this reason, many students will wait for the latest industry exchange activities on activity platforms such as "luma". In his opinion, finding a job through social networking is the most realistic shortcut in Silicon Valley at present.

In May this year, at the Yabuli Forum, Ding Jian, managing director and partner of Jinshajiang Venture Capital, said that in the past, more than 70% of Silicon Valley AI graduate students were able to return to China, but now there are almost none.

"Even if I can't secure a green card, having some work experience in Silicon Valley is definitely not a bad thing," said Zhao Ming.

(Zhao Ming, Qin Tian, ​​and Zhang Yushiqi are all pseudonyms)