Alex Ren, Founder of Fellows Fund: AI Value Creation from the Perspective of Silicon Valley

Source: Lei Feng Net

Author: Huang Nan

On August 14, 2023, the 7th GAIR Global Conference on Artificial Intelligence and Robotics, co-hosted by GAIR Research Institute, Leifeng.com, World Science and Technology Press, and Kotler Consulting Group, kicked off at the Orchard Hotel in Singapore.

In the era of the explosion of AI entrepreneurship, as an international AI forum, this conference has attracted many entrepreneurs and investors from Asia. The conference opened a total of 10 themed forums, focusing on the transformation and innovation of popular fields such as AIGC, Infra, life sciences, education, and SaaS in the era of big models. On the first day of the session “Outstanding Contributors in the GPT Era”, Alex Ren, a Silicon Valley pioneer investor and founding partner of Fellows Fund, shared a keynote speech on “AI Value Creation from the Perspective of Silicon Valley”.

In the past six months, AI entrepreneurship has been in full swing. For businesses and consumers, AI means better decisions, better actions, better outcomes, and better experiences. Compared with the past few years, some of the current AI companies have begun to make profits, and the prospects are very promising!

Alex Ren believes that the current investment in AI can start from four dimensions: one is the release of productivity, that is, AI-driven tools automatically perform tasks and provide output; the other is changes to the industry, that is, the use of artificial intelligence to optimize processes to improve efficiency, Reduce costs and improve results; the third is the AI middle layer, which refers to the AI middle layer connecting LLM to build scalable and customized AI applications; the fourth is AI Agent (AI intelligent body), where AI replaces humans and machines to interact and learn.

During the dialogue with Qiu Zhun, an overseas partner of China Shadow Capital, the two also discussed how AI can subvert the production method of social media content and the commercialization path of start-up companies.

The following is the content of Alex Ren’s live speech, which Leifeng.com has edited and organized without changing the original intention:

01. Four directions of AI investment

We are a venture capital fund located in Silicon Valley. Different from other VC companies, we look at projects from three points of view: space, that is, where is the space for start-up companies; A large number of commercialization stages; dimensions, what is the core competence of the company, and the methodology of competing with large companies.

Starting from these three dimensions, we propose four directions for AI investment.

The first direction is unleashed productivity, where AI-powered tools automate tasks and provide output. Today, users can generally feel the efficiency improvement brought by AI tools, such as using ChatGPT to generate text, write a song, or write code, etc., and it will be completed soon. Among the companies we invest in, several AI companies such as Gamma.app, Taskade, CodeComplete, Opus Clip, etc., their technical models provide AI capabilities through the combination of AI tools, so as to improve the model’s ability to deal with problems in the field and improve work efficiency. productivity.

The second direction is the mode transformation of the industry. Many industries will have their own data, such as biology, insurance, home services, etc. By embedding AI into the workflow of the industry for optimization, efficiency can be improved, costs can be reduced, and results can be improved. For example, we invest in Diffuse Bio in the biological field, Kyber in the insurance field, and LiveX AI in the life service field.

The third direction is the middle layer of AI. If the bottom layer of AI technology is a large model, then between the technology and the final application, we will need more middleware, such as LangChain, LlamaIndex or other middleware tools for a certain domain or architecture. Such as Anarchy AI we invested in.

The fourth, and the most popular direction in the past one or two months, is AI Agent (AI agent). AI Agent was proposed a long time ago, but until the emergence of GPT, it caused everyone to think about AI Agent. Currently There are many engineers in Silicon Valley who are doing entrepreneurship in this area. Anothermind.ai, which we invested in, is a new type of AI Agent start-up company.

A brief review of the past three stages of AI: the earliest period was the Classical ML period, and many statistical methods for learning from structured data and predefined features emerged. Then came the deep learning stage, where neural networks can learn from unstructured data such as images, text, and audio. After Transformer, we have also entered the stage of generative artificial intelligence and basic models. Based on the GPT method, various texts, images, codes, or chip design algorithms can be generated. It can be said that generative AI is the next round of breakthroughs.

Our team has been paying attention to AI since 2016 and 2017. We can find that the core difference between current AI companies and previous start-up companies is that in the past few years, the profitability of scene applications represented by autonomous driving has not met expectations, but Many AI companies today are highly profitable, and we can already see some AI companies making money.

So what value can AI generate? We sum it up as better decisions, better actions, better outcomes, and better experiences.

The first is to use AI to make better decisions, such as using it to analyze credit scores and analyze financial risks in marketing scenarios.

Second, better actions, that is, to infer from user actions and provide better recommended personalized services.

Third, better results, that is, to obtain better output results through optimization.

The last is to provide users with a better experience. For example, if you call a certain bank or airline in the United States and often have to wait for a long time, the user service experience is very poor, but with AI customer service to optimize the internal process, the user experience can also be greatly improved promote.

In specific applications, the AI experience mentioned above needs to be completed with a Workflow.

For example, an article written by the media needs to go through manuscript annotation, editing and other processes before it can be published. If AI is used to assist from the initial draft, article revision, summary, etc., the value of using AI and not using AI Curves will vary.

In the absence of AI assistance, limited by personal ability or speed efficiency, the limit will soon be reached. But with the help of AI, even a writer who has no experience in writing legal documents can use a large model to complete the writing of relevant documents and effectively supplement the content. ability. In this process, AI makes inferences and executions by understanding people’s needs, and finally assists people in the writing process.

Here is a brief introduction of how the big model uses Agent to improve its own ability.

Because large models are trained on historical data, they cannot understand current events. Suppose we let the big model rank the weather and temperature of all cities in Southeast Asia today. At this time, the big model needs some tools to be able to communicate with the outside world to obtain this information, and then complete reasoning based on the information. This is the concept of Agent. In other words, the Agent is the eyes and ears of the large model, enabling the large model to understand the environment and thus be able to process our current information.

02, large model innovation Agent paradigm

The next question I want to discuss is, what can AI do for us?

It mainly includes three aspects, namely automation and auxiliary AI, that is, how AI can automate the process in Workflow; release creativity; and better human-computer interaction.

In automated and assistive AI, we can distinguish their different values along two dimensions. In the face of problems with low complexity and a large amount of tasks, by embedding AI into workflow automation processes, the production efficiency of enterprises can be greatly improved; when faced with more complex tasks, AI is a human tool, Play a supporting role. For example, in the fields of drug development and material design, the tasks themselves are more complex and require higher professional knowledge. Therefore, the AI capabilities it uses are often not what an AI that can draw can solve.

Let me give you a case of using AI for project management that we invested in before. In Taskade, an instruction of a marketing plan is sent to AI, and various user roles are generated, and various packaged PDF files are handed over to AI for text analysis and processing. This is task-driven AI.

It is an integrated task management tool, coupled with AI embedding, can realize the whole process management of the completed production process. A very important point in this tool is the research on user behavior, using content to interact with users, the production of content can also be generated directly based on user behavior, and feedback from the user side can be used to feed back the model, and at the same time, personalization can also be used Recommend pushing product advertisements to users.

Therefore, we believe that this should be the general trend of the next generation of e-commerce and retail development.

It should be noted that the release of productivity based on large models is also being done by big companies such as Microsoft today. How should start-ups compete?

We found that many large companies often have the problem of not being able to reach the “last mile” in terms of AI product experience. It will do relatively poorly, which may become a gap for start-up companies to compete with large companies-iterate fast enough to give users a better experience.

For example, Opus.pro, an AI video generation company we invested in before, has gained a large number of loyal customers in just two months since its launch. Users only need to enter a YouTube video link, and the platform can generate a dozen short videos within three to four minutes, and distribute the videos directly to TikTok, Instagram, and YouTube. Similar AI content generation capabilities have great potential in games, movies, and more.

In terms of user interaction and communication, AI can also solve problems in natural language understanding, such as the interaction between humans and machines, communication in different regional languages, etc., which is different from the past interaction methods that mainly focused on API calls. The emergence of large models The interaction method has also undergone tremendous changes. This mode of understanding through large language models, human-computer and human interaction has formed a new interaction paradigm centered on Agent. For example, in translation, search and other scenarios, many jobs are being redefined by AI.

But at the same time, we must also see the limitations of large models. Such as the illusion of search, information lag and other issues. For example, when we search Google for someone’s resignation news, because the model training uses a large amount of past data for training, when the information is not updated in time, the answer generated by the large model will produce wrong results. It is necessary to further comprehensively process and iterate the problem through human correction or guidance, and finally a correct conclusion can be drawn. This is what we are currently emphasizing on large-scale model innovation in search scenarios.

In the application of automated testing, AI is currently also used in drug screening, new material design, etc. For example, in the field of drug screening design, the development cycle of a drug element in the past was as long as 7 to 12 years, which can be effectively shortened by AI assistance. drug development cycle. For example, Diffuse Bio and Persist AI, which we invested in, apply AI to drug screening and drug packaging respectively.

The AI ecosystem and investment structure have also changed a lot today. If the bottom layer is defined as the AI operating system, this operating system includes various frameworks, such as TensorFlow, PyTorch, computer hardware, and open domain data. Models in some development fields can be processed based on open field data, such as various common GPT models, Diffusion models, etc. This is our definition of the latest operating system.

On this basis, add vertical field data, which are data that companies such as OpenAI or Google do not have, and train proprietary large-scale models based on specific fields, supplemented by tools to better train the model, and solve vector databases and data. Issues such as privatization, plus today’s hot Chat Agent assisting people to do some things, and some work on using AI to reshape user experience for specific application areas.

In the past, the Internet digitized us, whether it was people, scenes, objects, or behaviors, based on digital derivative search engines, various e-commerce platforms, etc., and managed them in the form of the Internet. But today, what we have to do has become how to make some tasks or AI-realizable problems into agents to realize automation and more efficient processing, which will become the focus of the next breakthrough of AI.

Through AGI (Artificial General Intelligence, general artificial intelligence), it will be possible for AI to surpass humans in the future and release humans to more creative and valuable things. We can switch to something more interesting, valuable, or more in need of “human” abilities. The development potential of AI is very great, but I am not worried. Among them, what we see is the convenience that each generation of technological innovation brings to human beings, enabling human beings to do more valuable things. Therefore, we also hope that in the future, we can better support innovators in various fields and help you create some great businesses in Silicon Valley and around the world!

03、Alex Ren talks with Qiu Zhun of Huaying Capital

**Qiu Zhun:**I have invested in Silicon Valley for many years. Recently, I have traveled a lot between China and the United States. I am more concerned about some directions of Chinese companies going overseas. AI is the most important part of this. I would like to ask Alex the first question. Just now you mentioned a lot about AI Value Creation (AI value creation). We often say that the part of Value Creation is AI. As a key factor in startups, can you share some cases how landed? Social media and e-commerce have entered a bottleneck period, can AI bring any disruptive changes? How to evaluate from an investment point of view?

Alex Ren: At present, we see a lot of discussions about AI, mainly around how to integrate into daily work and life to help people save time and improve efficiency. On the other hand, it is entertainment and interaction.

One of the great advantages of AI lies in its ability to tell stories. Whether it is a movie, a novel or a video on YouTube, traditional story content production is done by people. I often give an example, such as Cao Xueqin wrote “Dream of Red Mansions”, Jia Baoyu became a monk, and Lin Daiyu died. This story is established, but in the narrative logic of the new era, everyone should be able to experience the Dream of Red Mansions, and the result is not Same. For example, in my story, I can chat with Jia Baoyu and ask him why he wants to become a monk, and I can also change Lin Daiyu’s ending.

Therefore, Gen AI is actually a very good opportunity to generate content narrative logic in the new era. Each generation of social media produces content in a new way, and this new way also produces a new medium. Therefore, the new medium is actually a highly personalized narrative tool. The same is true for e-commerce. Can the future Amazon e-commerce platform use the Peer to Peer model and use the model behind it to understand people’s needs, match them, and even produce. Therefore, the form of this traditional large platform may be completely subverted in the future development of AI. What we need to do is to find and amplify these opportunities to generate new-generation social media and new e-commerce platforms. Its methods may be completely Unlike the way we use it today.

Qiu Zhun: This is very interesting and reminds me that the earliest form of e-commerce is actually P2P. Indeed, from the perspective of Unit Economics (unit economic benefits), this model with too much manual operation intervention is difficult to achieve, but if AI is used, especially in the form of large models, this may become a very core direction . The second question, if you look at the commercialization path, can you give some direction to large-scale startups?

**Alex Ren:**From the perspective of a start-up company, we need to be clear about our own limitations, that is, what can we do? What are you good at?

As I said earlier, when startups compete with large companies, their advantage is not in training models or doing Infra, but in realizing customer value. Chinese entrepreneurs have a special advantage, that is, in the Internet age, they have learned how to iterate quickly and find the pain points of users. The consumer Internet is not a model of “you tell me what you need, and I will make it for you”, but a model of “we all say a word, throw a lot of mud on the wall, and see which piece can stick, and we will choose that thing” , This model is also applicable to AI today. In addition, we should pay attention to the adoption (landing) speed. There are three different customers in the market, one is To C, one is Prosumer (professional consumer), and the third is enterprise. As far as today’s market is concerned, it is clear that To C and Prosumer will be implemented much faster, but the enterprise scene will be slower. Therefore, in this process, start-up companies need to know your landing scenario, who are your customers, and what are their pain points? How fast is your adoption? If the landing speed is relatively slow, it will be more painful for the company’s development, and what the company can build is the technical threshold. Of course, this can also be done, but it is relatively slow.

**Qiu Zhun:**Let me also add something. In fact, today we see a lot of projects related to commercialization. We still pay more attention to the entrepreneur himself, his industrial background, and his understanding of application scenarios. From the perspective of large-scale models, there are two main differences in the paths of Chinese and American startups: one is the path of realization, and the other is the status of adoption. From your observations, are there any differences between the two?

**Alex Ren: **Today, whether entrepreneurs in China or Singapore generally have advantages in To C and To SMB, but To Enterprise is more difficult, because the marketization of corporate customers in the United States is also a big problem. Every American To B company needs to build a large sales team and marketing team. This is not only a challenge for Chinese companies, but also for all overseas entrepreneurs.

The difference between the two is that most startups in Silicon Valley seldom talk about macro strategy before the A round, and focus more on how the product solves the pain points of users. This is the most critical, and the other is the development trend, so it will be more grounded. Silicon Valley often emphasizes PLG, Power Lead Growth, to attract users through the improvement of product functions. This is a core shortcut that pays attention to each user’s feedback on the product, and then performs rapid iterations.

Qiu Zhun: At present, there are a lot of AI companies in China, whether it is the application layer, the bottom layer including the Infra layer, everyone is very active. In contrast, how is the adoption progress of American startups measured? At what stage?

**Alex Ren: **The more mature ones at this stage are text and Vincent diagrams. From the earliest company Jasper AI, to ChatGPT, and Midjourney in the field of Wenshengtu, in fact, they all landed very quickly. But what they have in common is that they are all oriented to C-side consumers or Prosumers.

There is a very interesting phenomenon in American companies, that is, for example, Midjourney is invoked through a platform such as Discord, that is, the network effect we talked about in the Internet era. See, I can also see that I can also generate other people’s cases. This network effect allows users to learn from each other. Therefore, Discord is also a very important product distribution platform, especially for Prosumers. , everyone can quickly form a community.

This model is very important for early-stage companies. The network effect formed among users can accelerate product dissemination, causing the user group to explode quickly. Of course, as far as the present is concerned, the current explosive applications are still mainly text and two-dimensional image generation and processing, and other technologies are not ready enough and are still under development.