This article is from a sharing on artificial intelligence investment made by the author and a group of investment analysts on April 14, 2023, with some abridgements. In addition, you can refer to the author’s article last month
Wang Chuan: Thoughts on chatGPT (1)
(1)
Do some risk warning first. First, what I share with you today is all from publicly available information that can be searched online. Second, here is just talking about my experience in the US stock market. I don’t know much about the Asian and Chinese markets. Third, I only predict the relatively long-term trend of more than three or four years, and I am only interested in this kind of research. That is, after you vote, you can ignore everything and lie flat. You don’t have to rack your brains to speculate on whether the next quarter’s revenue will be good, and you don’t have to worry too much about competitors coming, and so on. Fourth, my point of view will also be constantly adjusted with new information and new changes. Therefore, what I share with you today is just my opinion at this moment, not investment advice. Investment is risky, and you must be cautious when entering the market.
(2)
The emergence of the artificial intelligence large language model (LLM) represented by chatgpt will have a greater impact on the world than steam engines, printing, writing, and fire. Its speed of evolution and spread is unprecedented. Only by immersing in it all day can you understand its power. The structure and workings of the world will change completely and irreversibly.
Jason Wei, who was still working at Google last year, wrote a paper, “emergent abilities of large languages models” (emergent abilities of large language models), which revealed one of the core principles of the accelerated development of future technology. For many large language models, when the calculation training volume exceeds 10^23 FLOPs, the accuracy suddenly starts to leap.
The emergence of chatgpt is an “emergent” phenomenon and a mathematical inevitability. Once you understand this mechanism, you will realize that there will be a long list of powerful new AI functions that will emerge below, and this is also a mathematical inevitability. (Emergence can be defined as: a certain ability of a certain system suddenly begins to grow rapidly after the parameters of a certain dimension exceed a certain critical point. This ability does not exist before breaking through the critical point)
The key point of LLM is that the size of the model and the amount of data have reached a critical value. In the past, the reasoning ability that was considered impossible for machines suddenly has. Large models ushered in an artificial intelligence aha moment. People spend their whole life building models and making analysis and judgments about various situations. Now artificial intelligence has its own reasoning ability, and it will continue to scale up, automate, and reduce costs, which is equivalent to touching the essence of human activities. The impact of the future is unlimited.
The T in GPT is the abbreviation of Transformer. The core improvement of this technology for natural language processing is to establish correlations between relatively distant words when training AI comprehension. For example, the sentence “I am allergic to peanuts, every time I eat it, I feel uncomfortable”. Transformer can use the so-called “attention” mechanism to establish a correlation between the words “peanut” and “it”. Compared with traditional AI language models, its comprehension is greatly improved. Extending this concept, in order to improve our understanding of the world, we should also constantly train ourselves to quickly extract various relationships between things that seem to be far away but are actually strongly related.
(3)
An important factor in the breakthrough of artificial intelligence technology is the improvement of large-scale parallel computing capabilities.
The core of artificial intelligence technology is the neural network. The core algorithm of the neural network is the forward propagation and back propagation calculation when adjusting the network parameters, which is essentially matrix multiplication. For matrix multiplication calculation, the calculation of each element can be carried out in parallel independently, without interfering with the calculation of other elements, so it is especially suitable for the large-scale parallel computing capability of GPU to realize.
The primary visual cortex in the human brain is said to have 140 million neurons. These neurons perform complex parallel computing and information transmission in the background, giving the human brain fast visual perception. At the bottom level, the mechanism is similar to that of GPU’s parallel computing to obtain image recognition capabilities. Animals with visual ability can quickly gain an overwhelming advantage in the competition for survival in most environments, and slowly eliminate other animals without visual perception.
But in hearing and other word processing tasks, the parallel computing ability of the human brain is actually very poor. Even if you read ten lines at a glance, you can only read more than two hundred characters at the same time. Now Gpt-4 can accept more than 30,000 words at a time, which is two orders of magnitude higher than human ability, and it will continue to increase rapidly. This ever-increasing high degree of parallelism means that its comprehension and penetration of words and the world will continue to reach a level that is difficult for ordinary people to understand. So be sure to work hard to learn to use this tool to improve your understanding of the world.
What needs to be understood is that gpt, as a tool with a natural language interface, can tolerate certain ambiguities in input information, so it can be expanded almost infinitely to a larger amount of data training. Many previous tools require specific input The interface and grammar are completely incomprehensible within a few letters, and they are not scalable at all, and they are not of the same nature as gpt.
Efficient tools, when evolved to a certain extent, will connect and synthesize various elements to become an unprecedented entity with higher and higher efficiency, even from a distant observer, it looks like a living body with a clear goal. Pay attention to observe the connection of Gpt, what kind of brand-new entities with completely different structures will be created.
Large language models will be the super glue for smart compositional overlays. Essentially, it has a strong language comprehension ability, which reduces the accuracy requirements of the interface (there are many types of software interfaces in general, and the definitions of each company are different, the format is strict and accurate, and compatibility must also be considered, and it is not easy to achieve versatility), which greatly improves The versatility of the interface lowers the threshold for technology combinations in various subfields
To extend this idea, when allocating time and attention, the most important thing is not your ability to master a specific technology, but to maintain an ecological connection (not cut off) with the most intelligent large language model at all times ability. Always immerse yourself in this ecology, and you will win most of the battle; building a car behind closed doors in isolation, and fantasizing about any competitive advantage that can be created, at best it will be short-lived, and there is a high probability that it will be a waste of time.
In the future, the distinction between super-individuals and companies will become increasingly blurred. The bandwidth of communication and interaction between people is relatively narrow, it is easy to make mistakes and lose the chain, and you often have to wait. There is no upper limit to the bandwidth of LLM and its plug-ins, and it should become more and more stable, fast and accurate.
Large language models such as Gpt are a kind of meta tool (meta tool). A typical feature is that it can correct its own errors, and after asking it to reflect on its own output, it can immediately output new and higher-quality content. This ability to be used in a myriad of scenarios is a key to what makes this type of tool fundamentally different from previous innovations. This is why the idea that “ordinary people always feel that they have some specialties that cannot be replaced by AI” is extremely naive.
The technology of Gpt is equivalent to enabling very few people to have the ability to “process unstructured data information on a large scale”. It used to be that machines couldn’t handle “unstructured data”, but now it’s different. A large number of labor-intensive and tedious tasks in software development (such as writing glue codes) will be easily replaced.
Some people arrogantly think that if a single skill is stronger than AI, it will not be surpassed and replaced. But after LLM and various plug-ins are linked, it is very easy to acquire new skills. Then all sorts of unprecedented superhuman abilities will naturally emerge.
AI will eventually automate all the work that everyone can do now, and the marginal cost will be infinitely close to zero. The entire social structure will undergo unimaginable changes, and many traditional concepts will either disappear or be completely reconstructed. And this change is irreversible.
Huang Renxun of Nvidia predicts that AI computing power can increase by one million times in ten years, which is six orders of magnitude, 2^20. The reason for the increase in computing power (relative to Chatgpt) mainly comes from new chips, parallel connections between chips, parallel connections between systems, new operating systems, new algorithms, and so on.
Technology is advancing so fast that the best strategy for ordinary people now, I am afraid the default is to “lie down, take care of your body, and wait”. Otherwise, the little resources you have worked so hard to accumulate may be worthless due to technical factors in a few years, but your health will be truly lost.
It will probably not be until after 2002 that the Internet will truly accelerate the productivity of the real economy. For example, in the past, the document processing procedures of the US banking industry were relatively backward, relying heavily on faxes or overnight couriers to transmit documents. Faxes often fail to be sent out, and you must print the fax yourself to confirm the sent information, lest the recipient cannot find it or deny it. In fact, everyone started to use broadband Internet to transfer files when doing business, and it started slowly after 2002. That’s nine years after the advent of the Internet browser. It will not be fully popularized until after 2009.
Similarly, it will take time for the large language model LLM to be popularized in many real industry applications, but this time it should be faster. Maybe within a year or two, some companies will start to use LLM tools to improve efficiency when doing business with each other. Maybe within four years, by 2027, it will be fully available. Then the enterprises at that time looked at the current operation mode of the enterprise, just like the modern people see the old man using a fax machine to send dozens of pages of documents slowly, and find it ridiculous and backward.
(4)
From an investment point of view, the only companies worthy of long-term investment are companies that can control the ecology, have a strong monopoly, and have the ability to collect taxes; without the ability to collect taxes, competitors can continue to go around, or keep lowering your price, which is not good. investment object.
You can refer to the author’s old articles
Wang Chuan: Looking at Investment from the Evolutionary Mechanism of Power and Monopoly (1)
What is a “taxable niche”?
1. That is to say, most other players in the ecosystem have to use the products and services of a certain company;
2. And you can’t leave after using it, and you can’t exchange for other company’s products. (Because of the continuous deep integration of usage habits and functions, industry chain group inertia, or other reasons), and the longer it takes, the harder it is to leave;
3. Because it is difficult to be replaced by competitors, the price of this product service will not drop compared to the proportion of the entire ecological economy. It may even continue to rise.
When your investment object occupies the ecological niche of tax collection, you can hold it very firmly, and you will basically not be affected by all kinds of rumors and judgments. And in this way, you will be less interested in investing in companies that are not in the “tax collection niche”.
There are also strong and weak tax collection capabilities. For example, Apple’s tax collection capabilities are stronger than Facebook’s and Google’s.
Most of the entrepreneurial toss, if it does not reach the “tax-collecting ecological niche”, it is very hard and risky for investors. Reaching the “tax-collecting ecological niche” is a long-term process of hard work, and it also requires a certain amount of luck. But if you’re not convinced yet that something hits the “tax niche,” it’s not there yet.
AI is an infinite game. Only by building the largest, most open and richest ecology, allowing as many players as possible to join their own ecology and help themselves share the cost can they truly occupy the strategic commanding heights.
Netflix has more than 100 million paid subscribers worldwide. chatgpt plus is said to have close to 2 million paid subscribers, but as long as it keeps improving the service, there is no reason why its paid subscribers will be lower than Netflix.
Microsoft and Openai have a long-term profit-sharing agreement, about 75% of the profit in front of Openai, until it recovers its initial investment of US$13 billion in Openai. The subsequent profit sharing ratio is 49% for each. So basically Microsoft and openai can be regarded as one.
The real barrier to competition lies not in data, but in the construction of the ecology, especially when a large number of third-party developers spontaneously participate in the construction of the ecology.
For developers, Microsoft is the master of github. Developers use the portal by default. Once you use github, you will use the programming tool copilot provided by github. Once you get used to it, it is difficult to change. Then you use github, the path of least resistance is to continue to use other tools and facilities in the Microsoft ecosystem, because it is convenient and cheap, such as the chatgpt plugin plugin, or participate in related plugins yourself, because the openai ecosystem The most potential customers. After you do it, whether you train a large model by yourself or deploy the service on the cloud, it will be difficult for you to escape the track of Microsoft’s azure cloud service, or because of convenience. And Microsoft is gradually integrating all these Microsoft AI tools into office, bing, edge browsers, and windows. Windows still occupies more than 70% of the desktop computer market. Azure accounts for more than 20% of the market share, and Microsoft’s office software package, including excel, powerpoint, word, accounts for about 85-90% of the entire related office software market. These tools will be integrated by Microsoft and its AI software. So once you come into contact with any of Microsoft’s tools, you will naturally be attracted to his AI ecosystem, getting deeper and deeper.
There is another data, that is, Y combinator, an angel investment company in Silicon Valley, recently supported 280 projects, of which about 38 are for chatgpt-related projects, so these people are all in the end for Microsoft and openai Work part-time to help them grow their ecology.
Now by default Microsoft/openai is the leader in this space. If other competitors want to surpass, they must rely on a large amount of market data to prove themselves, rather than relying on a few press releases to fool investors and users.
Microsoft’s current share price is less than 290, and the PE ratio is about 30, which is not particularly high. This frenzy of artificial intelligence may cause its profits to double in the next five years. According to historical data, Microsoft’s revenue in 2022 will increase by 100% compared with 2017, and its profit will increase by 200%, which means it will triple. Therefore, it is not outrageous to expect Microsoft to double its profits in the next five years. This is one of the long-term win-win opportunities with relatively clear logic. (Not investment advice, investment is risky, be cautious when entering the market!)
Regarding AI hardware companies, the biggest risk is that, relatively speaking, at the low end of the value chain, revenue comes from corporate users, and the volatility may be greater. When the bubble bursts, the price drops very sharply. Even for Cisco, the giant of the network equipment company, in 2002, compared with the highest point in 2000, the price retracement reached nearly 90%. In the same period, Microsoft’s price retracement was only about 50%.
(5)
This is a possible development trend in the field of AI investment in my opinion in the next few years. History rarely repeats itself, but it often rhymes.
People saw the power of LLM and saw OpenAI making money;
People want to compete with OpenAI;
New competitors buy new GPUs;
GPU makers like Nvidia make money;
New competitors want to outdo Nvidia, or make specialized chips.
Chipmakers like TSMC make money;
Competitors who want to surpass TSMC to build better fabs need better lithography equipment;
ASML makes money;
OpenAI went public, and the early venture capital received more than 100 times the return.
Early-stage VCs go out and raise new mega-funds;
Institutional investors piled in, worried about missing out.
More AI-related companies get funded, especially infrastructure companies.
Valuation models increasingly become based on the fabrication of fears (fear of missing out, fear of being acquired by a competitor, fear of looking stupid) rather than rational calculations based on the realistic discounted value of cash flows.
Trend chasing is profitable for a while, and people who play it safe can look pretty stupid at first.
VCs can raise more money from naive investors by bragging about high IRR based on illusory or unsustainable valuations in the private market.
Then the Fed lowered interest rates, and more hot money poured in.
The tide rises, and all ships rise with it. For a while, everyone made a lot of money.
LLMs, chip makers, fabs, equipment suppliers, VCs, analysts, AI company employees. Anyone who hasn’t reinvested their investment earnings into AI-related stocks will feel pretty stupid compared to those who have.
At the peak of the bubble, hedge funds that shorted the stock bubble would be hit hard.
The analyst who had been wary of the bubble, but was wrong for five years, changed his view at the peak and became bullish.
Then the Fed tightened. Then everything fell apart.
(6)
In the face of bubbles, it is difficult for ordinary people to be alone. One is that I don’t have the resources to stick to a relatively mediocre strategy for a long time, the other is that I can’t bear the huge mental pressure of someone who looks stupid than me temporarily surpassing myself by a large margin, the third is that I think I can choose the time to leave early, and the fourth is that Thinking that you can choose the time and choose to take the initiative to short, and so on.
Even if you know the approximate evolution of the framework, you can’t actually change anything. Corporate funds enter the market because of the fear of competition, institutional funds enter the market because of the fear of missing out, and retail investors enter the market because they are afraid that the old man next door will make a lot more money than themselves. All the funds entering the market compete to raise prices first and reinforce each other. This fear-driven strategy has worked for quite some time, and people’s belief in this strategy has continued to strengthen and is difficult to change.
The valuation of a company ultimately depends on cash flow. However, the cash flow analysis of early high-tech companies is very difficult, and in many cases it is almost impossible. At this time, people tend to use some self-deceiving one-sided extraction of technical details and new technology terms to analyze and judge the company’s technical strength, and use it as cash. Alternatives to Stream Analytics. This method of analysis tends to appear effective in the early stages of bubble expansion and is therefore prone to (erroneous) reinforcement.
The misunderstanding that many speculators of high-tech companies tend to fall into is that when the entire valuation model collapses like a volcanic eruption, they are unable to quickly correct their thinking models and escape the disaster.
During the rising period of the bubble brewing, there may be a few years of “invest in a dazzling growth high-tech company, no matter how much profit, no matter whether it can be sustained, sell at a higher price” This thinking mode has always been effective, Gradually become deeply ingrained, taken for granted, and confident.
When the inevitable decline and industry reshuffle occurs: many customers of the company themselves go bankrupt, and the revenue from them disappears; the company that was originally profitable suddenly becomes negative cash flow; the original intention to pay 100 because of the fear of fomo Investors with double PE are now hard to protect themselves, and are only willing to pay 25 times PE; the financial fraud that can be demolished to compensate for the west, and there is no more room to maneuver to maintain it. At this time, the combination of expectations and reality can only be realized by the collapse of stock prices, and there is no other choice.
Investors who made a fortune before did not understand that this was a big collapse. The original rational reaction was to immediately switch to the “emergency mode of self-preservation”, and escape as far away as possible. But for people with a fixed thinking pattern, the more natural reaction with less resistance is to comfort themselves and say that all this is temporary, and it will rebound in a few months. I used to go through temporary periods like this Difficult!
When the volcano erupted, he continued to run in, constantly investing his precious resources to consume. This is the most effective way to use up your life savings within a few months. No matter how well you have done before and how much you have accumulated, before the inevitable periodic depression and industry reshuffle, staying away from the damage caused by this big collapse is one of the keys to success or failure.
When the industry reshuffles, even some leading companies with promising long-term fundamentals will inevitably encounter a price retracement when the stock price falls by more than 75%. The reason is simple: if the revenue is reduced by 20%, the profit may be cut by more than half, and if the PE drops from, for example, 60 to below 30, the stock price will naturally fall by more than 75%. As for non-leading companies, it is common for prices to retreat by more than 95% or completely return to zero. Price pullbacks cannot be completely avoided, but act with a clear understanding of the worst-case scenario and the resources and ability to withstand the shock.
Faced with the advent of the new wave of technology, entrepreneurs often make a mistake. When they do not understand what sustainable competitive advantage they really have, they follow the trend and rush forward, thinking that after they make a wishful product, everything It will be solved. More often, deep down in his heart, he may not really want to make a good product, but imagines that when he has done enough, he will definitely find another company to take over the business and make a small fortune. If he happens to meet an investor who is as naive and eager as he is, it is easy to hit it off and push this blind move forward.
In practice, once the market is tested, you will find that: potential customers do not buy the product; some customers are interested but unwilling to pay so much; customer service and operational support costs are too high to make a profit; Suddenly there is a new competitor launching a similar product, the price is lower and better, and the promotion channels are stronger. The effort I spent in the past is completely useless; there is no feedback for a long time, and the feeling of being ignored is really painful. I can only knock my teeth out. Riton. In the end, I really didn’t have the financial and material resources to continue, so I could only quietly close the stall in desperation, and then the painful lessons were drowned in the vast crowd, and continued to be repeated time and time again by various hopeful young people behind.
This article is reproduced from: https://chuan.us/archives/907?utm_source=rss&utm_medium=rss&utm_campaign=%25e7%258e%258b%25e5%25b7%259d-%25e4%25bb%258e-chatgpt-%25e7%259c%258b% 25E4%25ba%25ba%25E5%25B7%25A5%25E6%2599%25E8%258BD 25E7%259A%2584%25E6%2595%25E8%258B5%25E6%259C%25A%25E 25E 4% 25bc%259a%25e5%2592%258c%25e9%25a3%258e%25e9%2599%25a9
This site is only for collection, and the copyright belongs to the original author.