Author: Melissa
Original source: Atom Capital
This time I stayed in Silicon Valley for six weeks. I came in midsummer and left just after the beginning of autumn. The sunshine in California is always bright, and here at the forefront of technology, there is a wave of AI. Because I hope to have a deeper understanding of the development and direction of AI, I met many people (including friends from big factories, entrepreneurs, and investors), participated in online and offline activities, and really felt the surging tide at the beginning of the tide. Pick a few waves here and share them with you.
01 After the epidemic: manpower shortage and remote work
The epidemic has become a thing of the past, but maybe because I am new here, the traces left by the three years are very clear. Among the most felt are the impact of manpower shortage and remote work.
shorthanded
There is an obvious shortage of manpower in Silicon Valley, coupled with the recent inflation, the result is that manpower is very expensive. Once I used Uber Eat to order a subway takeaway. The sandwich itself was 8 yuan, but the takeaway costs add up to 17 yuan in total, which is more than doubled! I lived in Seattle for many years before, and I always knew that the labor force in the United States is not as good as that in China, but I was surprised to see this. After understanding, one of the main reasons is that during the epidemic, many people left their jobs or retired early due to fear of infection. Coupled with the government's money distribution in the past two years, the number of people working has decreased a lot. Talking to an entrepreneur doing AI education, he said that the shortage of teachers is very serious. This is a problem facing the entire United States, and I don't know how it will be solved.
work remotely
The remote work started during the epidemic has a greater impact, especially for newly graduated college students. There are two friends who started their own businesses, and they both mentioned this problem to me separately. The epidemic requires isolation, and college students do not have the opportunity to practice in companies while they are in school. After graduation, I work remotely and have no experience of getting along with colleagues. These lead to them not knowing how to work as a team, and it is difficult for the lead to guide remotely. They have recruited graduates from very good schools (including Stanford), and the potential of the students is very good. It is a pity that they have to let them go because they cannot cooperate.
Now big factories have begun to ask employees to come to work in the company one after another, but they have not recovered to the pre-epidemic situation. A graduate I recruited on the Expedia team many years ago is now the founder of an AI company. He feels that working remotely affects efficiency. During the epidemic, he dared not ask employees to come, fearing that he would ask them to leave. At present, he is watching the rhythm of the big factory, and when the big factory makes a clear request, he will keep up. According to the actual situation I have seen in large factories and start-up companies, the number of people who come to work in the office is still limited. Discussing with friends, everyone's attitude towards this is not completely consistent. In general, the larger the team managed by one person, the more dissatisfied with the remote working method. Everyone feels that it will gradually return to the past, but it is unlikely to happen overnight.
Talk about an interesting observation. Big companies such as Google and Meta are located in Palo Alto and Mountain View, resulting in expensive housing prices nearby, and much cheaper houses in distant places. Because remote work does not come to the company, as a result, housing prices in distant places have also risen very well in the past two years
02 The tide of AI: the pattern is initially determined, very early
My focus is on AI. To sum up, some observations and judgments about AI in Silicon Valley over the past month.
Large Models and GPUs
The industry structure of the large model itself is initially determined. Different from China's Hundred Models and Thousands of Models Competition, Silicon Valley has several big models that won. Closed sources are mainly OpenAI and Google, Anthropic can also be counted, open source is Meta Llama-2 and so on. Because the investment in the general-purpose large-scale model is extremely high, requiring a lot of manpower, computing power and capital, the pattern seems to be basically determined, and there are no new entrants.
There is still a shortage of GPUs, whether it is a big factory or a start-up company. Everyone is looking for GPUs . NVIDIA's junior gave me a popular science on the GPU production process, starting with the preparation of ore sand. Hardware is not my focus and my knowledge is limited. It sounds like due to the long production cycle, there will be a shortage of GPUs in the short term, and it should be ok in the long run.
AI is very early
Speaking of the status quo in the AI field, an investor friend described it vividly. He said it's like it's not yet dawn, and everyone is shining around with a flashlight, looking for directions. The real rise of the mobile Internet has not yet arrived. I have talked with many friends, including large model developers, large and small companies that use large models, and startups that provide product services around large models Infra/tools. The overall judgment is that the application of large models is still very early.
One example is very representative. I have a friend who used to be the VP of Engineering of a very well-known listed company. In the past few years, he has started his own business to do eCommerce platform-related work. He has more than 100 employees and is invested by several well-known funds in the United States. Her business needs large models. Recently, she is exploring how to do it and made two attempts. One is to fine tune the private data on the MosaicML model. One is to use GPT-4, put the private data in the vector database, and use search-retrieval to put the corresponding information in the prompt. Comparing the two, to her surprise, the result of GPT-4 is better than that of fine tune. She was very confused and didn't know how to do fine tune to have an effect. What kind of data is needed, how much data is used, and how to finetune are not very clear. Moreover, the large model is a black box, and she feels that the person who made the large model may not understand it very well. In addition, she said that the experience of using MosaicML is not good, but there are no other tools to choose from. Although GPT-4 works well, her private data cannot be made public. It can be used for testing, but not for official products. She feels that the existing team's technical strength in this area is limited, and plans to recruit AI engineers to solve this problem next.
I was a little surprised to hear that. Because she is very experienced, the entire entrepreneurial team has a bright background and good technical strength. If she still doesn't know how to fine tune effectively, then other companies can imagine. Her result comparison (fine tune is not as good as GPT-4 search-retrieval), is not a special case, and I have heard many similar examples. Another friend of mine's start-up company provides AI tools to serve large corporate customers. He said that large-scale models are a brand-new technology for large enterprises, and his clients are just beginning to think about them, and they are particularly concerned about the accuracy, speed, data quality and privacy issues that need to be prepared for the model. Customers are also exploring which business problems to use AI to solve. He judged that it would take at least 6-12 months for a company to be able to use it internally.
It can be seen that this round of AI is still in a very early stage. There is no killer app (except ChatGPT) on the C side, and it will take time for the B side to land. AI infra and tool layers should still have a lot of room for development. For example, Databricks spent US$1.3 billion to acquire MosaicML, hoping to establish AI capabilities to empower customers as soon as possible.
Here I see two positive messages:
- It is precisely because it is still in the early stage, the tools are not perfect, and large companies do not have ready-made technologies available, which leaves room for startups. If big companies can use it immediately, they have their own data and scenarios, and there are fewer opportunities for startups. This is a point of view raised by Howie Xu, a teacher in Silicon Valley, which I resonate with.
- Big companies are eager to use AI, at least with a sense of crisis. I learned that for this round of GenAI, many companies have set up special budgets internally. Now that the money is ready, even if the development is slower in the early stage, the prospect of AI is still very bright, and it is not easy to be cold.
Why do you feel that the development of AI has slowed down in the past two months?
I don’t know how you feel. Compared with the beginning of the year, I feel that the rhythm of the AI field has slowed down significantly in the past two or three months. why is it like this? Observation, roughly as follows:
- It is related to OpenAI's strategy. This wave of rhythm is mainly led by OpenAI. It has been holding back its big moves before, releasing the achievements of the past two or three years (such as GPT-3) in two or three months from the end of last year, which makes people feel dizzy. After this period of catching up, Google has become a strong opponent, and now OpenAI dare not launch products that are not ready, otherwise the gains outweigh the losses. Therefore, there has not been a particularly big change recently, and it may feel slower than before. In fact, I think this is the rhythm that technology should have, which is not so fast in the first place.
- Entrepreneurs are working hard to build. I gave a lecture in the Silicon Valley AI community on this issue. Feedback from the community. At the beginning of the year, entrepreneurs were busy participating in various conferences/lectures/meetup learning discussions, trying to figure out what GenAI is all about. Recently, everyone has a basic understanding of large-scale model technology, and they are busy spending time building their own products. From the outside, it seems that it is not as lively as before.
- In the field of research, paper after paper is still being published, and it has not slowed down.
The primary market has indeed slowed down
The pace of investment in the overall primary market seems to have slowed down. Mainly related to the environment. People feel that the future economic trend is uncertain, and the Russo-Ukrainian war has increased uncertainty, affecting people's confidence in investment. In addition, during the epidemic, the government's large-scale release of water has caused the valuation of many entrepreneurial projects to rise very high, and the valuation is still in the process of correction. Against this background, the primary market in the AI field is actually relatively good. However, because it is still very early, I have observed that in addition to the competitive projects that really make big models (including character.ai is actually making big models) and get a lot of money, it is not easy for other AI entrepreneurial projects to raise funds now. Many investors are waiting and watching.
03 Exploring Big Factory: OpenAI, Google, NVIDIA
In this wave of AI tide, OpenAI & Microsoft, Google and NVIDIA have become the trendsetters of the times. Three of them are headquartered in Silicon Valley. I went to find out and summarize the information that can be shared.
OpenAI
OpenAI is very concerned about information protection, and employees are also very sensitive about it. I don't know much, but a few points are relatively impressive.
People who work with OpenAI mention that their staff is very capable and efficient. Its system performance and monitoring are particularly good, and its engineering capabilities are strong. Perhaps, Infra's engineering capabilities - how to use hardware more efficiently, improve performance, etc., is one of its core barriers.
OpenAI is obsessed with AGI, and I can only really understand this after talking about it in detail. They judge the priority of work internally, and they will see if this can help the development of AGI. If you can better train the model and help the model learn, then you will do it; otherwise, you will not spend effort. For example, they had worked on robots before, and they felt that they were greatly constrained by the actual physical world and could not help AGI, so they stopped. According to this inference, there is a high probability that it will not do vertical fields.
Before ChatGPT appeared, users had no perception of the effect of LLM. It is very important for users to perceive. In addition to AGI, ChatGPT and API are also the focus of OpenAI.
In the past, Google was relatively slow in advancing AI. In addition to conflicting with the advertising business, it was also related to two things. One is that a researcher felt that the large model was conscious and was fired. Before that, a black female employee sued Google because she was rejected for publishing a paper. These made Google very cautious about AI and slowed down the progress of AI.
Google always felt that it was in the lead until ChatGPT appeared, which put a lot of pressure on Google. It is relatively rare to start code red (the highest priority) internally in December. Now the company attaches great importance to GPT. There is a dedicated team to do GPT (DeepMind and Google Brain merged), and other teams are encouraged to use AI as soon as possible. Many of my friends are at Google, and they have confidence in Google, and they feel that Google will not lag behind in this regard.
NVIDIA
In this LLM tide, NVIDIA has become the biggest winner. In fact, I have not paid much attention to this company, because my experience and interest are in software. This time I took a good look at it and found it very interesting, so I will share more here.
A person's Startup
NVIDIA's style, in a nutshell, is the startup company of Lao Huang Jason. The friends who work there admire Jason very much, and I feel that Jason is a superman😊. Jason has always believed in Compute. No matter what the stock price is, he has insisted on doing it since 2012 and has never hesitated. Jason has a deep understanding of technology, understands the actual situation of the project, and is approachable. If there is something that cannot be decided, everyone asks Jason, he makes decisions quickly and well.
Jason is very compassionate. For example, at the beginning of the epidemic, the company usually conducts employee evaluations in September, but he decided to do it in advance. As a result, in March, the whole company completed the evaluation, salary increase and bonus, so that everyone can get the money in advance. At the same time, Jason has insight and a sense of crisis, and is very popular among employees. Even when the stock price was not good before, employees thought of him very highly.
Emphasis on technology, flat organization
Its company culture is significantly different from other companies I know. As a company with nearly 30,000 people, NVIDIA has no people manager (manager who only cares about people). The company emphasizes technical capabilities, and managers, regardless of their level, are very technically strong.
Tissue is flat. It seems that only Jason has an assistant in the whole company, and no one else has one. I asked what to do about team building and the like. A friend said that the company has no team building, no Christmas dinner, only a company-wide conference. At the meeting, Jason gave a two-hour speech without script. He was a good joker. After the speech, many employees went up to take photos with him.
NVIDIA ecology
I heard that NVIDIA’s ecology has been doing well for a long time, so I specifically asked what it was referring to, and my friend made it very clear:
- Complete tools are provided. The chip is a deep stack from bottom to top, which requires various supporting tools, including compiler, debugger, profiler, etc. The needs of R&D personnel are different. For example, some people want to do in-depth optimization, so it is not enough to just encapsulate the function into an API.
- The speed and ease of use of the system.
- Internal and external company, horizontal communication is done very well. For example, the company has a team responsible for communicating with customers, and they also know internal technology very well. What customers have needs, they have discussed directly with the internal R & D team very early. The same goes for the interior. The software team works closely with the hardware department, instead of waiting for the hardware to be ready before developing the software, but interacting and cooperating in a timely manner during the process.
04Internationalization of Chinese enterprises
Changes in Sino-US relations are closely related to Silicon Valley. I noticed two noticeable changes this time. Entrepreneurs are more focused on market selection, either in the US market or in the Chinese market, and few people do both. Some good domestic entrepreneurs and funds are also looking for new opportunities here.
How to do a good job of internationalization for Chinese enterprises is a question of common concern. I participated in a closed-door salon on the weekend, and the topic of discussion was this. I think the guests are quite representative: there are CEOs of Chinese listed companies in the global market, there are fund partners who focus on investing in Chinese companies going overseas, and there are managers who manage both teams in China and the United States. entrepreneurs, and I am one of them. You all shared a lot of insight. China has advantages in R&D costs, a complete supply chain, Internet product operations, and diligence, but going overseas faces completely different challenges, involving marketing, products, team culture, and management. One point that the guests resonated with is that to do international business, the founder's thinking must first be internationalized.
My emotion is more outside the discussion. I am no stranger to the topic of internationalization. The topic discussed many years ago was how American companies expand to China. Now it is the other way around, discussing how Chinese companies can do the international market. The center of gravity of the world is changing. After these years of hard work, China has really become much stronger, which makes people proud.
05 Flowing Silicon Valley
I have always been very envious of Silicon Valley's talent resources and the atmosphere of free communication. There is a high density of talents here, and I often chatted with them and found out that they are Tsinghua alumni. There are 30 students in my undergraduate class, 6 of them are here. On weekends, I attended a barbecue party organized by a friend, chatted casually, and found that several of them talked well. One more question, it turns out that he is a successful person who hides his secrets
Because it's Silicon Valley, the entrepreneurial spirit has always prevailed. Followed by various lectures, forums and so on. When I first came, my friend gave me a Google doc, which densely listed AI offline activities in San Francisco, almost every day. It is not convenient for me to go to the city, so I only participated in a few times selectively. Later, I searched by myself, and found various online Webinar and community discussions on topics of interest. Later, I got familiar with it and found that there are many activities in the Bay Area. Whether online or offline, the quality of these activities is generally very good. There are core members of big companies or top startups, and young entrepreneurs. The information shared is dense and updated quickly, and the speakers are independent thinking attitudes and sincere technological frontiers. . I have always enjoyed learning new things and have enjoyed my time here.
Silicon Valley is fluid, with fluid talents, fluid information, and fluid capital. These flows bring dynamism and innovation, making it possible to change every day and feel forever young.
