Martin Casado stated that artificial intelligence is achieving success in the three areas of creative creation, companionship, and programming. He mentioned that the current bottleneck in the development of artificial intelligence lies in the fact that human knowledge is limited. Artificial intelligence requires data for development, but it can only utilize all existing human knowledge and cannot capture future human knowledge
Although facing doubts about a "bubble," the valuation of artificial intelligence continues to soar. What does this mean for companies and investors hoping to get a piece of the pie? Where are the opportunities?
Recently, at the annual WSJ Tech Live summit, venture capitalist and a16z partner Martin Casado spoke with Jason Dean, global technology editor of The Wall Street Journal. Here are the key points from Martin Casado's perspective:
- Artificial intelligence could be a super cycle, which has several decades left, and is still in a very early stage.
- In addition to well-known large models like ChatGPT, there are many smaller models that handle tasks such as voice, music, and images, and their investors have also reaped significant rewards.
- Currently, there are three successful tracks in artificial intelligence: creative composition, companionship, and coding.
- The current bottleneck in AI development is that human knowledge is limited; AI development requires data, but it can only use all existing human knowledge and cannot capture future human knowledge.
- Another major limiting factor in AI development is electricity; to enter the next stage and enhance model capabilities, resources such as computing power, electricity, and data need to be increased tenfold.
- Even if all cutting-edge models stopped developing today, converting existing technology into practical applications could still yield substantial value.
- AI will not replace artists; AI is just a machine.
The full translation of the interview is as follows:
Q: What types of AI startups are worth paying attention to now, and are not too expensive or too late?
Martin Casado: It is meaningful to compare AI with the internet. Throughout industry history, we have periodically seen the marginal cost of certain things drop to zero— in computing, the marginal cost of computation has become zero. We used to manually calculate logarithm tables and then let computers do it, which created the computing revolution. Then came the internet, where the marginal cost of distribution dropped to zero.
Now, in artificial intelligence, the marginal costs of language, reasoning, and creation also seem to be approaching zero. If this is the case, it represents a super cycle, and if so, we have several decades left, so there is no notion of "too late"; in this sense, we are still in a very early stage.
Currently, the models that everyone has heard of from Google, OpenAI, and Anthropic are supported by large companies that will not collapse. Clearly, this has strategic value for them and marks the beginning of great things.
There are some well-known large models, such as Gemini and OpenAI, but there are also many smaller models that handle tasks like voice, music, or images. If viewed from an investor's perspective, these companies are actually very successful. Q: Some companies already have very high valuations, and the barriers to entering the artificial intelligence field seem to be rising. What is the most interesting thing right now?
Martin Casado: Currently, there are three application scenarios that are achieving success, and there may be more in the future.
The most effective application scenario right now may be creative creation, which can include images, music, etc. Many companies in this field are growing at the fastest rate we have ever seen.
Imagine the cost of creating a AAA video game. It's about $500 million to $1 billion. In fact, today's models can create every aspect of the game; you can create 3D models, stories, videos, textures, and the actual computational cost of generating this content is only about $10.
The second application scenario is quite interesting, which is companionship. As technologists, we have never solved the emotional issues of computers; obviously, they cannot express emotions. However, I can give you an example: my daughter is a "pandemic child," now 14 years old, and has spent a lot of time on Character.AI. She not only uses Character.AI herself but also brings her characters into conversations with friends. This phenomenon has integrated into social life, and we see this companionship feature being very widespread.
The third area that is achieving success is programming. For example, using the AI code editor Cursor, whether you are a professional programmer or a beginner, you can leverage the model for complex programming, and it works very well.
Q: Can artificial intelligence create works that are remembered by people?
Martin Casado: We have the largest creative model portfolio in the venture capital world, and I see many models capable of producing images, videos, etc. Then I find that many artists use these works. The biggest factor in judging whether the output work is excellent is whether the user has a formal artistic background.
These technologies will not replace artists; they are just machines. If you have a sensitivity to beauty, you can create something more beautiful. If you understand people's needs, you can get what they want.
Q: People often say that all problems can be solved by increasing computing power. Do you think things are more complex than that?
Martin Casado: If you haven't read "Bitter Lessons," please do; it's very short. This book says, when it comes to artificial intelligence, you have two choices: you can come up with some very clever algorithms, or you can find ways to utilize more computers. We can always acquire more computing power; as long as we keep increasing computational resources, things will get better. This is basically what we do.
However, artificial intelligence needs data. We are humans and have existed for a long time. For three thousand years, we have observed the universe, thought deeply, judged what objects are and how they interact, and we have done all this computation and recorded it.
**Artificial intelligence has simply found a way to utilize this existing work and output it. Once you exhaust these resources, things will slow down. We have created these amazing models that capture all human knowledge, but they cannot capture future human knowledge; artificial intelligence cannot reason autonomously **
This does not mean that the development in other areas will slow down. For the current version of the model, we may have reached the limits of general language improvement, and now we need to enter more specialized fields.
Q: Power is another major limiting factor.
Martin Casado: That's right, we have also reached the limits of our ability to build data centers. We need to find ways to construct larger buildings, provide more power to these buildings, and build network architectures.
However, to enter the next stage to elevate the model's level, we need to increase resources tenfold at each stage: ten times the computing power, ten times the electricity, ten times the data—this is exponential growth.
Q: Does this affect your decision-making as an investor? Do you need to consider that this might be a good idea, but I don't know if we will have enough computing power, data, and electricity in the next five to ten years to achieve it?
Martin Casado: If all frontier models stopped today, there would still be a lot of value to be gained by translating these technologies into practical applications. Again, everyone is focused on OpenAI, but in terms of value creation and integration, if you look at all those private companies and small firms, they are building their own small models, and they are among the fastest-growing companies we have seen in industry history