Source: Reprinted with permission, from Chaos College (ID: hundun-university).
Responsible Editor | Nie Min
7640th in-depth good article: 8050 words | 17 minutes to read
.AI
AI can be defined as a transformative outlet.
But AI is very complex, it is difficult to express it with an identity, it can be a model, an algorithm, an APP application on a mobile phone, or a specific robot. So after the rise of AI, there are a lot of philosophical things that have become very subtle and interesting.
For venture capital institutions, we define a concept, or define a vent, whether it is temporary or transformative, its concept is whether it is an isolated track, or can drive an entire industrial chain.
For example, smart electric vehicles have driven power batteries, automotive chips, automatic driving and other directions, so we call it a transformative outlet.
AI has the formal characteristics of a transformative outlet: it includes the infrastructure layer, the model layer and the application layer, and there are disruptive opportunities in each direction.
As a senior investor, in the face of the rapid iteration of AI, I will also feel some anxiety in learning.
I was very impressed by the fact that in the third week of March this year, GPT-4, Baidu Wenxin Yiyan, Microsoft Copilot, Stanford University's Alpaca 7B, Tsinghua University's ChatGLM-6B... During that time, you and I should be the same, our circle of friends was flooded by AI.
With passive anxiety and active desire, I have had an intensive chat with more than 60 entrepreneurs in the field of AI in recent months, more or less thinking and judgment, and I share it with you today.
Although it may not be right after three years, I still hope to give you some inspiration.
First, some judgments on the development of AI
1. Attitude towards AI development, not FOMO, nor JOMO.
The emergence of many capabilities of artificial intelligence today did not start with the emergence of ChatGPT at the end of last year, but a new round of data and computing power revolution that began in 2017-2018, which is the quantitative change to qualitative change accumulated by technology.
Let's start with our attitude, we are neither too conservative because of FOMO (Fear Of Missing Out) nor too conservative because of JOMO (Joy Of Missing Out).
Investment in the technology industry is like this, too JOMO will miss the sector opportunities of the big wave of technology, too FOMO may be eaten back by the bubble.
Compared with the model of OpenAI + Microsoft + NVIDIA in the United States, China's ultimately successful large-model companies, regardless of alliances, business models, C/B revenue contributions, etc., will be very different, and future development will take time.
I have always stressed internally that it is necessary to be "calm in the midst of noise, and continuous radical in the calm", and it is necessary to distinguish what stage is the outlet and what stage is the value.
Overall, we feel that this is a long-term road, if we compare the golden decade of the mobile Internet (2010-2020), the best companies were actually founded in 3-4 years after 2010, and now AI has just begun.
2. Potentially successful AI companies should build their own data flywheels.
A classic division of the AI industry chain is the infrastructure layer, model layer, middle layer and application layer, and the horizontal division of application scenarios can be mainly divided into ToC and ToB.
At present, we believe that a new AI company can really emerge in the future to challenge the existing giants, and it must dare to seek a breakthrough from the ToC scenario, because the data flywheel effect that the C end can bring may be the key to winning in the early stage of AI.
I don't know if you have noticed that the top AI companies such as OpenAI and Character.AI are both models and products, which is very different from the mobile Internet, LBS, 4G/5G and other basic technologies of the mobile Internet, and upper applications such as Taobao and Didi are still relatively separate.
This is not an accidental phenomenon, it is caused by the stage of market development.
AI is indeed still in the early stage of technological innovation, and the typical feature of this stage is the need to use technology to drive products, such as ChatGPT's innovation is to use a chat interface to directly reach users with large models.
This is in the early stage of the technology penetration S curve, and it is necessary to continuously improve the technical effect to be able to do a good job in the product, and gradually approach the critical point of a significant increase in penetration rate. At such a stage, the data flywheel becomes extremely important.
I spoke with Moonshot AI founder Yang Zhilin the other day, and he believes that the best case scenario is that the model keeps serving the user, and then the user keeps generating new data for the model.
One of the most successful aspects of Midjourney is that it embeds user feedback in its core process, because each user must be forced to choose the one that best meets their expectations among the four images generated by AI, which is a huge data flywheel.
For ChatGPT, although there is such a feedback mechanism, it is not in its core process.
Therefore, in product design, how to form a feedback closed loop is very critical, the data flywheel will continue to optimize the model capability, this gap will ultimately determine how much value can be provided to the user.
Whether it is a ToC or ToB company, creating a feedback closed loop and forming a data flywheel is also a natural choice. For some small companies starting from scratch, this thing can be difficult, but because you have no baggage, there are opportunities.
3. Vertical models with specialized barriers may be where the opportunity lies.
Former Google founder Eric Schmidt has a view that the future will be multiple vertical models or multiple vertical assistants, including a variety of high-value, specialized AI systems.
That's because many high-value, domain-specific workflows rely specifically and necessarily on rich, proprietary datasets. For example, Bloomberg recently launched Bloomberg GPT, Bloomberg is to make the model small, the number of parameters is about 50B, compared to GPT-3 175B is much smaller, although it weakens the versatility, but in the financial field is stronger.
In the rhythm of domestic AI and industrial integration, there will also be some unique opportunities, especially in the real economy, advanced manufacturing, intelligent driving and other fields, domestic development may be faster, there will be some more innovative models, application scenarios, and the support of high-probability policies will be very obvious.
The AI era may upend many ideas in the SaaS era. We will look at how much of an AI application is the ability of a large model such as GPT, and how much is its own ability. If the barriers are too low, many products may not survive an iterative upgrade of GPT. We are also constantly thinking about what are the new observation points.
In this wave of AI, we cannot overemphasize the importance of data. Because, globally, data is becoming more scarce.
According to a joint study, Will we run out of data? Data native to humans may become increasingly scarce in the future, and high-quality natural language data may be exhausted by large language models as soon as 2026.
This means that the existing Internet data is limited, for domestic related companies, we may first reach a standard level, but then how to improve, depends on how to continue to obtain legal compliance, business logic data sources, the real value will become sustainable high-quality data.
Based on this, the cooperation and binding of domestic large model companies and the original data source head companies in various industries may be deeper and more equal, and even the data source in some fields will be stronger, in this context, there is a high probability that there will be large model companies with different angles, positions, and industries, and model + computing power + data + scenario will be the four most essential dimensions of successful companies.
4. Two directions of large model products: personalization & scene.
If we look a little further, the next step for large model products could be in two directions:
(1) Personalization: Equip it with "memory".
One of the missing things about the previous big language model was that it lacked memory updates, and every time you reopened ChatGPT, it didn't remember your last conversation. Some AI companies are seeking breakthroughs in this direction, such as Character AI, which is valued at $1 billion, and Rewind, which is valued at $350 million.
Among the AI startup teams we contacted, many teams wanted products to have memory capabilities and bring personalization to users.
Because it contains human emotions, understanding your needs, and personalized satisfaction of you, this is a further innovation of AI than the previous mobile Internet era, so that AI can truly become a human work assistant or companion, which also brings AI Infra such as the opportunity of vector database.
(2) Scene: Equip it with "hands" and "eyes".
If you think that ChatGPT can only ask some questions and can't do much, then you need to try OpenAI Plugin, which is a new app store launched by OpenAI, hundreds of plug-ins cover various daily needs such as clothing, food, housing, social, work and study, etc., which can be said to put "hands" on large models.
For example, a plugin called Klarna Shopping, its function is price comparison, you just need to enter the question: "Please compare the prices of Sony SLR cameras on different shopping sites", and ChatGPT will give you the answer.
Another example is KAYAK, which searches for flight, accommodation, and rental car information in real time and provides travel recommendations based on your budget.
For example, to book a hotel, you just have to ask: "Please find a hotel near the Museum of Modern Art in New York with a budget of $300 per night." ”
OpenAI is following a model similar to Apple's "hardware + App Store" to move towards a higher strategic system status, seeing OpenAI Plugin's hundreds of flowers of apps, is it a bit similar to the feeling of mobile Internet entrepreneurship.
The "eyes" are multimodal (text, pictures, images, etc.), and we do not only communicate through pure language (text) in our daily life, but also obtain a very high proportion of information through the eyes. Advanced AI like the Marvel movie Justin (J.A.R.V.I.S.) and Cortana in the "Halo" game requires multimodal intervention, which is an important development direction.
In order to install "memory", "hands" and "eyes", it is inseparable from the decline of the cost structure of large models. We have seen that after April this year, the cost of training + inference has been rapidly decreasing.
Chinese entrepreneurs have always played well in the application layer, and in the next six months to a year, there will be more application innovation on a monthly basis, and we will also pay attention to the leading team in technology and products.
Second, the ice and fire of AI,
Experimentation is always more important than sitting and talking
A small question that I was asked about, but I thought it was necessary to bring it out and talk about it.
The question is: "Now AI seems to be very hot, but it seems that most of them are discussing, chatting, and sharing, and there are not many actual uses and applications, what do you think of this problem?" ”
I'd like to start with two interesting numerical comparisons. One figure is that executives at 110 S&P 500 companies that held earnings calls mentioned AI in March-May, three times more than in the past decade.
Another figure is a recent Morgan Stanley survey of more than 2,000 people, and the result was that 80% of people had never used ChatGPT or Google's Bard.
Have you found the comparison between these two numbers, is it interesting? Obviously, the group surveyed by Morgan Stanley is most likely the group of people who have discussed, shared, spoke, and mentioned AI in this round of enthusiasm.
And this is not an isolated case, the real-world situation should be similar to the body feeling of this set of numbers, on the one hand, everyone is talking about AI, and on the other hand, there are not many people who have really used AI products.
A possible reason is that the current stories of billions of dollars of valuation, thousands of GPUs, and a large number of professional words are too tall, which has distanced us from AI, making us dare to talk about it, but be afraid to practice.
So, is AI really that far away?
In fact, AI is playing a practical role in many places, and I especially want to give a few very down-to-earth and interesting examples to give you some intuitive feelings.
The significance and value of these examples is that all of us who are interested in participating in innovative entrepreneurs, or possibly individual and business users, can participate more actively in the practice of AI.
1. Gain real data growth by embracing AI: Notion and Character.AI
Notion is a company founded ten years ago, and I don't know if any of you have used it. It is a personal note-taking software that can be used for blogging, social media copywriting, meeting minutes, work emails, and much more.
After the first wave of access to ChatGPT at the end of last year, this product suddenly exploded, because after its above-mentioned functions were integrated with ChatGPT, the generated content only needed to be easily modified and could be used directly, and its application scenarios were greatly simplified, efficient and strengthened.
The direct result of this application scenario is that Notion achieved $10 million in ARR revenue in just one month.
Character.AI is also an interesting app, which is a chatbot based on a large AI model. Compared to ChatGPT and New Bing, the memory function is Character.AI strong feature, ChatGPT when you start a chat with it, you don't remember what you talked about last time, but you Character.AI remember.
So, you can create and train your own personalized artificial intelligence on it, you can also choose existing public virtual characters, such as Elon Musk, Socrates, etc., you can talk to Musk about rockets, you can talk to Socrates about philosophy.
The direct result of this application scenario is that it was downloaded 1.7 million times in less than a week on mobile, more than three times that of ChatGPT, and nearly 200 million visits in April.
Not only is the growth fast, but users also spend a long time on the Character.AI, reaching 25.4 minutes per visit.
In the future, this also points to virtual humans with a certain sense of role, although there may be some ethical problems, but it is not far from realization, through the combination of software and hardware, there will be a lot of imagination.
2. Specialized and vertical model scenarios: DoNotPay and legal applications
Since the large model is the most mature in text processing, it has a natural advantage in the text-rich legal industry, which has been verified overseas.
There is a popular company in the United States called DoNotPay, which mainly provides AI legal services for "small lawsuits". For many small things, people often do not go to court, because the cost performance is too low.
DoNotPay leverages AI to digitize the entire legal process, making it easy for individuals to operate and for a small fee. For example, if you want to stop the membership fee of the gym, there is a lack of manual customer service in many places in the United States, and it requires very cumbersome procedures to refund.
With DoNotPay, you can help you find the right contact information, write a complete reprimand email, and help you refund. And this scenario is just one of hundreds of legal scenarios in which DoNotPay helps people solve real pain points.
You might say that the U.S. is different from domestic, but let's take another example of a domestic company using AI in the legal field.
We have a food consumer portfolio company that uses AI in the legal process. The biggest headache for their legal department every day is to confirm whether a certain marketing technique can be said to the outside world, which is very dependent on personal experience, time-consuming, labor-intensive and inefficient.
They tried AI for 3-4 months, and the results were very good, even the most experienced people are difficult to achieve the quality and efficiency of AI judgment. AI can even tell you which legal provisions are involved, and what have been similar judgments in the past, which is very convenient.
These scenarios are actually very common, but we don't have a particularly convenient entrance to apply, but this is the opportunity.
3. A small, but significant example: applications in the medical field
In a more common but potentially more professional scenario, I recently read a public account article about the experience of an American emergency department doctor who treated a 96-year-old Alzheimer's patient at 3 a.m. who had difficulty breathing because she had fluid in her lungs.
The patient's three children, also in their 70s, were also in the emergency room, emotional and verbal, and they began to make constant requests based on their own experience, but these requests were actually very wrong for the 96-year-old patient.
Since there are still several patients to be treated at the same time, if the doctor goes to explain and argue with these emotionally unstable family members over 70, it will take a lot of time and affect the treatment.
The doctor gave ChatGPT-4 a directive "why you can't give intravenous fluids to people with severe pulmonary edema and breathing difficulties and explain them in compassionate language."
ChatGPT wrote a very good answer, the doctor asked the nurse to read this answer to the families, and after listening to this reasonable, logical, and empathetic explanation, the excited expressions of the families melted into calm agreement.
I'm very impressed by this example, at a time of urgency like the emergency room, if ChatGPT can save an average of 5 minutes per patient, and the number of emergency room visits per year exceeds 130 million, that means saving 10 million hours of time per year, even if my calculation is idealistic, even if the number is reduced tenfold, it is extremely meaningful.
This is actually a very small scenario, which does not involve payment and business processes, but this is the finest granularity of embracing AI.
4. Some examples of ToB scenarios
The examples just mentioned are basically ToC examples, which are close to everyone. In many ToB scenarios, AI is also being rapidly applied.
For example, the generation of marketing materials in e-commerce scenarios will not be significantly lower in quality than labor, but the cost will be reduced by 1-2 orders of magnitude;
In the consumer goods industry, AI can automatically generate multiple professional-grade product concepts for customers to choose and use in the early stage of product development and inspire early inspiration.
In the field of human resources, AI can efficiently screen all resumes for a position according to the needs of HR, greatly improving the efficiency of HR;
In the world of programming, 41% of the code on GitHub today is generated by AI, and the process took only 6 months...
These scenes may be relatively far away from some of you here, but they must also be very close to some of you.
So, going back to the beginning of this section, why do I say this issue is worth talking about? Because most of us are still in a state of sitting and talking, we are eager to discuss and learn, because we have FOMO emotions, but it may be more meaningful to do it ourselves, in life, in work, in products, to do some tentative practice, so that we can really not be abandoned by this new era of AI.
Third, a few suggestions on AI entrepreneurship
Finally, the host asked me at the conference, "After experiencing so many true and false outlets, can you give some advice to entrepreneurs who want to embrace AI?" ”
Recently, many friends have asked me similar questions, and I have also seriously thought about it, combined with the recent exchanges with dozens of entrepreneurs, I think it is not a suggestion, it is also what I am doing, for your reference.
1.AI learning and application, everyone must pay attention to the use and effective iteration is greater than everything.
This has actually been mentioned in the second part just now. To say more, a simple key point is to learn to write prompts, knowing how to ask questions is very critical, how to better interact with AI is also a science.
2. You can try to form the correct underlying work logic, or design the right AI to change the process of work life, determine the goal and review according to the rhythm.
For example, the company's business process combing, the use of appropriate AI tools, the most reasonable and appropriate entry point. In your own network, find the most suitable people, companies, products, and services, so that you (whether you understand technology or not) can truly iteratively recognize. The evolution of each technology, from the beginning of germination to the formation of the market, is cyclical, different points in time, different strategies, different products, different business line penetration. The founder's biggest fear is to get the wrong stage, step on the wrong time, and spend the wrong money.
At this stage of the 3.AI wave, it must be technology-driven and product-oriented.
Future applications and model capabilities are more closely integrated, so the understanding and gap of the model will determine the product and user experience, and small teams with technological innovation genes and capabilities must run hard.
4.AI entrepreneurship, we must not only be able to make good use of the AI tools on the market, but also organize the company structure from the perspective of AI efficiency and change.
Better use of AI tools will definitely bring more efficient per capita output, if the AI era is still backward organizational structure, talent density, it shows that the founder's iteration and real understanding of AI are not in place.
5. If you are making ToC products, the positive feedback effect brought by the data flywheel is very important.
This flywheel needs to be paid attention to from day one of product design, and better data can produce a real flywheel effect. Like many of the great products we've seen, whether it's Character.AI or Midjourney, their feedback mechanism is very well set up.
This feedback will make your data flywheel form a positive cycle, spiraling upwards. At the same time, it is necessary to pay attention to the memory and personalization of the product, which may be one of the biggest product features in the AI era.
6. In the current state of such a hot state related to AI, it also puts forward higher and more comprehensive requirements for founders.
Founders need to think comprehensively about technology, market conditions, investor expectations, and currency stock, all of which need to find a balance and make certain pre-judgments, be able to comprehensively grasp the speed of their own and the company's comprehensive growth, and have a forward-looking awareness of various risks.
7.AI entrepreneurship must grasp the rhythm of financing.
This needs to be expanded. The overall situation right now is that the U.S. is very hot, and China is just getting started.
According to PitchBook Data, the U.S. invested $12.7 billion in the first five months of 2023, up from $4.8 billion in all of last year.
The biggest recent funding round was Inflection, a $1.3 billion round valued at $4 billion, and Inflection is only a year and a half old. The biggest acquisition was also recently made, with Databricks acquiring MosaicML for $1.3 billion, a 6x premium to its valuation.
But for the country, everyone is now watching more and less hands, thunder is loud and rain is small, why is this?
The first point is that in today's environment, money has become more precious. For innovative companies, especially those that need a lot of tolerance for capital, the business model has not yet been verified, and the failure rate is relatively high, the money that can be given to these companies in the market today is 10 times less than 3 years ago.
The second point is that in terms of development direction, it is still relatively chaotic. There are not many large model companies in the market that have really obtained financing, and there are slightly more middle layers and more application layers.
From the perspective of timeline, after the emergence of ChatGPT, China also quickly followed, large factories, scientific research institutes, startups and other forces have quickly launched their own large models, everyone's goal is to do the Chinese version of OpenAI.
Later, the development of the open source model and the OpenAI open API interface changed a lot, open source led to a reduction in the cost of ownership of its own large model, and the open interface made many entrepreneurial opportunities in the middle layer and application layer in addition to the large model.
So, in the long run, I'm optimistic that there are opportunities in all directions, but it's not so clear yet, and of course the lack of clarity is an opportunity in itself.
For entrepreneurs, whether it is infrastructure, model layer, middle layer, front-end application, open source model or closed source model, each will have its own scenarios and advantages.
Of course, my personal view is still as mentioned earlier, in the end, the best AI companies in China are very likely to do both models and applications.
IV. Conclusion
All in all, new things often come out with two extremes, we don't want to deify AI, but we can't ignore it. What is useful will eventually remain and develop, not transferred by personal will.
Although most companies have not yet generated significant revenue through AI, and a large number of practices have not yet been born, everything is in full swing, and the next 6-8 months is an important window.
Many of Jingwei's portfolio companies are also actively trying, such as Ape Tutoring, which is trying to cut into AI tools in many business links; Another example is Power-Law Intelligence, they have cooperated with domestic large model companies to jointly train PowerLawGLM, a vertical large model in the legal field... We are in the early days of AI penetration into the business, and many products are still in development and internal testing.
In fact, some of our smart electric vehicle production companies have applied AI in multiple links including manufacturing, customer service, marketing, etc., which is a kind of penetration of productivity.
According to Sam Altman, founder of OpenAI, future model performance will triple every 18 months. Although AI can not replace people at present, we can see the participation of AI in various productivity links, and its ability to reduce costs and improve efficiency will gradually be brought into play, and we see that many entrepreneurs are extremely excited.
If we think that in ten years, AI is likely to disappear into the invisible, penetrate into all corners of our work and life, and change the world, and all companies will become AI companies.
Finally, I think of a very interesting topic, some people always think that if AI develops very rapidly, it will most likely become a disaster for human beings in the end.
But I have discussed this with many people before, and the premise of this judgment is actually to think about AI from some of our inherent concepts, thinking that it will inherit and be consistent with the deep-rooted and stubborn diseases of "domination and possession" of human beings.
But in fact, if we think about it from another angle, maybe AI thinks completely differently than we do, and it may not be interested in these at all.
Of course, I don't mean that in the process of AI development, we don't need any regulation or vigilance. Instead, I feel that in the initial stages, it may be more necessary to regulate and regulate the people who use AI.
With development, in the future, what we need may be to think outside the box, and maybe AI will eventually subvert our existing concepts. This is perhaps the most important thing for those of us involved in this new era today.
But anyway, for us who make investments, winning is so sweet because most of the time we lose.