Foreword: Sequoia recently organized their AI Ascent 2024 themed event. In addition to the opening speech entitled "The Opportunity of Artificial Intelligence", the live speeches of Andrej Karpathy and Ng Enda, the two AI gods that have been shared, are not to be missed. Bilibili: Opportunities for Artificial Intelligence: Sequoia Capital's 2024 AI Ascent Opening Speech [Chinese-English]-Bilibili] https://b23.tv/sxo67H5Andrej Karpathy's Speech Q&A:
P站传送:【精校】大神Andrej Carpathy最新红杉I活动演讲问答【中英】-哔哩哔哩】 HTTPS://B23.TV/APES吴恩达演讲:
Station B transmission: [Jingxiao] Andrew Ng's latest Sequoia AI event speech: the next step of AI intelligent agent workflow [Chinese-English] - Bilibili] https://b23.tv/shjLCQA friends who like the graphic version, the following is the full text of Sequoia's opening speech:
Together with all of our partners at Sequoia Capital, we welcome you to AI Ascent. In the world of artificial intelligence, a lot is happening. Our goal here today is to learn something new, meet someone who will help us on our journey, and hopefully we can have fun. So, what is this opportunity? A year ago, we thought this magic box could do something wonderful. I think we've gone through this cycle of hype in this simplified form over the last 12 months. Our expectations peaked, then we experienced a trough of disillusionment, and now we are climbing back to the peak of productivity. I think we've realized that what AI is really bringing to us today are three different capabilities that can be integrated into a wide variety of amazing applications.
The first is the ability to create, hence the name generative AI. You can create images, you can create text, you can create videos, you can create audio, you can create all sorts of things. The previous software couldn't do that, so that's pretty cool. The second is inference, which can be one-time or multi-step, agent-type inference. But again, software couldn't do that before. Because it can create, because it can reason, we've covered both the right and left brains, which means that software can also interact with human-like capabilities for the first time. This is huge because this has far-reaching business model implications, which we'll mention in the next slide. So, how should we look at this? A lot of times, when we see something new, we try to reason by analogy. In this case, the best analogy we can think of is cloud transformation, and while there are a million reasons why this analogy isn't perfect, it's still useful. Over the past 20 years or so, there has been a major structural shift in the technology landscape, bringing with it new business models, new applications, and new ways for people to interact with technology.
If we look back at the early days of cloud transformation, around 2010, the total market capitalization of the entire global software was about $350 billion, of which this small piece, only $6 billion, was cloud software. Fast forward to last year, and the total market capitalization has grown from about $350 billion to $650 billion, but this part of the revenue has reached $400 billion. That's a 40% CAGR over 15 years. That's tremendous growth. Now, if we're going to do analogical reasoning, the cloud is replacing software with software. As I mentioned the ability to interact in a human-like way, one of the big opportunities for AI is to replace services with software. If that's the total market capitalization we're after, then the starting point isn't hundreds of billions. The starting point could be tens of trillions. So you can really dream. About what this opportunity turns into. We're going to assume, it's a hypothesis, like everything we're talking about today, we're going to assume that we're standing on the precipice of the greatest value-creation opportunity humanity has ever seen. One of the benefits of being part of Sequoia is that we have a long history where we've started to look at different waves of technology, understand how they interact with each other, and understand how they're leading us to the present. We'll take a quick look back at the memory. So, in the 60s of the 20th century, our partner Don Valentine founded Sequoia Capital, who was actually the head of the listed company Fairchild Semiconductor, which got its name for Silicon Valley with its silicon-based transistors. We must pay attention to such a phenomenon. In the 70s of the 20th century, we saw how systems were built on these chips. By the 80s of the 20th century, these systems were connected via a network, with personal computers as endpoints, and software packages appeared. In the 90s of the 20th century, these networks were opened to the public in the form of the Internet, which changed the way we communicate and consume. By the 2000s, the internet had matured to support complex applications. In the 2010s, all of these apps came to our pockets in the form of mobile devices, changing the way we work. So why bother with this little build? The key here is that each new wave of technology builds on the previous ones. The concept of artificial intelligence is not new, its history dates back to the 40s of the 20th century, and I think it was also in the 40s that neural networks first became an idea. But what are the ingredients to turn AI from an idea, a dream, a reality, to solve real-world problems in a unique and compelling way, and to build a lasting business around it? We finally have cheap and ample computing power, and we have a fast, efficient, and reliable network. Seven billion of the 8 billion people on the planet have a supercomputer in their pocket. Thanks in part to the pandemic, everything was forced online, and as a result, the data needed to support all of these enjoyable experiences became easily accessible. So now is the moment when AI is the theme for the next 10 or even 20 years. Therefore, we have the strongest possible belief in an assumption that has not yet been proven, that the next few decades will be the era of artificial intelligence.
Again, we're going to draw an analogy between the cloud transition and the mobile transition. These logos are located on the left side of the page, and these are the majority of companies that have been born as a result of these transformations, and these companies have revenues of more than a billion dollars. The list is not exhaustive, but this is probably about 80% of the companies that have generated more than $1 billion in revenue as part of these transformations, not valuations, revenues. The most interesting thing about this slide is on the right. It's not something that exists, it's something that doesn't exist. Wide field of view. The opportunities are enormous. We thought, if in 10 or 15 years we were standing here, there would be 40 or 50 logos on the right side. Chances are, there are a bunch of company logos in this room. That's the opportunity, and that's why we're excited. So, I'll give it to Sonia. Thanks, Pat.
What a year. ChatGPT came out a year and a half ago. I think it's a whirlwind for everyone here. It may feel like we're all moving on and off, the ground beneath our feet constantly changing. So let's pause, narrow it down, and take stock of what's happened so far. Last year, we discussed how AI will revolutionize all of these different areas and deliver amazing productivity gains. A year later, it began to be in the spotlight.
Anyone here who has seen this tweet from Klarna's Sebastian? Raise your hand. It's incredible. Klarna now uses OpenAI to handle two-thirds of customer service queries. They have automated the equivalent of 700 full-time agents. We believe that there are tens of millions of call center agents around the world, and we've seen AI find its place in one of the most exciting areas of the customer support market.
The same is true for legal services. A year ago, the law was considered one of the least technologically advanced industries and one of the least likely to take risks. Now, however, companies like Harvey are automating much of what lawyers do, from the daily drudgery and drudgery all the way to more advanced analytics.
Another example is software engineering. I'm sure a lot of you here have seen some of the demos circulating on Twitter lately. It's worth noting that we've gone from theoretically writing code by AI a year ago to being a completely independent AI software engineer. I think it's really exciting. There will be more software in the future. Artificial intelligence isn't just revolutionizing work, it's already improving the quality of our lives. The other day, Pat and I were on Zoom and I noticed that he looked a little suspicious. There was no speech during the whole process. After some more reflection, I'm pretty sure he did send his virtual AI avatar and was actually going to the gym, which can explain a lot of things.
This is provided by HeyGen. It's amazing. That's how far technology has come in a year. It's just that it's scary to think about. It's both scary and exciting to think about how this will all play out over the next decade. "Two years ago, when we thought generative AI might be on the next great technological shift, we didn't know what was going to happen. Will real companies come up with it? Will real companies come out of it? Will real revenues be realized? I think the sheer scale of user pull and revenue growth has surprised almost everybody.
In our view, generative AI currently has a total revenue of about $3 billion, and that's not counting all the incremental revenue generated by F.A.N.G. companies and cloud AI providers. At $3 billion, for example, it took the SaaS market nearly a decade to reach this level of revenue. Generative AI achieved this in its first year. So the speed and magnitude of the dramatic change makes it very clear to us that generative AI is here to stay.
Customer attraction to AI isn't limited to one or two applications. It's everywhere. I'm sure everybody knows how many users ChatGPT has, but when you look at the revenue and usage of a lot of AI applications, whether it's consumer companies or enterprise companies, startups or established companies, a lot of AI products are actually resonating with customers and starting to look for products that fit across industry markets. As a result, we're starting to see the diversity of use cases starting to get really exciting.
The first thing that surprised me at least last year's funding environment was the imbalance in the share of financing. If you think of generative AI as a tiered pie, there's the base model at the bottom, developer tools and infrastructure on top, and then there's the application at the top. A year ago, we had predicted a Cambrian explosion at the application layer due to new enabling technologies at the base layer. On the contrary, we actually find that the formation of new companies and the formation of capital are opposites. More and more foundational models are emerging and raising a lot of money, and the application layer seems to be just getting started.
Our partner David was here last year and asked a thought-provoking question in his article "The $200 Billion Problem with Artificial Intelligence." If you look at how much companies are currently investing in GPUs, you'll see that we invested about $50 billion in NVIDIA GPUs last year. People often think that as long as you build something, people will use it naturally. Artificial intelligence is undoubtedly a field full of dreams. However, by now, what we need to remember is that in the last slide, we have confirmed about $3 billion in AI revenue, as well as changes from cloud providers. We've invested $50 billion in ground facilities, plus the cost of energy and data centers and so on. We have released three products. To me, this means that math hasn't quite become math yet. The funds required to build these facilities far exceed the investments made so far. As a result, we still have some practical issues that need to be addressed.
Although the usage, revenue, and number of users of AI may seem incredible, the usage data suggests that we are still in the early stages. So, if you look at the ratio of daily active users to monthly active users, or if you look at a one-month retention rate, generative AI apps are still far behind mobile apps. For me, this is both a problem and an opportunity. This is an opportunity because AI is still a weekly, monthly tinkering phenomenon for most people right now. But we have an opportunity to use artificial intelligence to create apps that people want to use every day.
When we interviewed users, one of the biggest reasons why they don't stick to AI apps is the gap between expectations and reality. So when you find out that the model isn't smart enough to reliably do what you're asking it to do, those seemingly magical Twitter demos become disappointing. The good news is that last year GPU spending exceeded $50 billion, and we can now build increasingly intelligent foundational models. Just last month, we saw Sora, we saw Claude 3, and we saw Grok over the weekend. Therefore, as the baseline level of intelligence increases, we should expect the market fit of AI products to accelerate.
So, unlike some markets where the future of the market is very uncertain, the benefit of AI is that you can define very clearly how these applications are going to get better and better. Let's remember that success takes time. We said this at last year's AI Ascent, and we'll say it again. If you look at the iPhone, you'll see that some of the first apps in App Store V1 were the Beer Drinking App, the Lightsaber App, the Flip Cup App, or the Flashlight, which were fun, lightweight demonstrations of new technology. These end up either being native apps like flashlights, etc., or utilities and gimmicks. The iPhone came out in 2007 and the App Store came out in 2008. It wasn't until 2010 that you saw Instagram and DoorDash in 2013. As a result, it will take time for companies to discover and take advantage of brand new features of the iPhone in creative ways that we have not yet been able to imagine. We think the same thing is happening in the field of artificial intelligence. We think we've seen what the next batch of legendary companies might look like. Here are some of the use cases that have caught our attention recently, but I think it's much broader than the set of use cases on this page. As I mentioned, we see customer support as one of the first use cases that really impacts the market fit of an enterprise's products.
As I mentioned in the Klarna story, I don't think it's the exception, it's the rule. I think that's the rule. For many of us, AI friendship is one of the most surprising applications. I think it took us a couple of months to think. But I think users and usage metrics in this category imply very strong user liking. Next, we will discuss the horizontal application of enterprise knowledge. Later, we'll hear more insights from Glynn and Dusk. We believe that enterprise knowledge is gradually being unlocked. Here are some of our predictions for what might happen in the year ahead.
First, we predict that by 2024, real applications will shift from Copilot (which advises and assists you) to agents that can truly take humans out of the loop entirely. AI will be more of a colleague than just a tool. We're seeing it start to play a role in software engineering, customer service, and more on this topic today. I understand that both Ng and Harrison Chase are planning to speak about it.
Second, we predict that one of the biggest problems for LLMs is that they seem to be repeating statistical patterns in the text without actually taking the time to reason and plan for the task at hand. However, this is starting to change with the emergence of many new studies, such as inference time calculations and iteration of game style values. What happens when you give the model time to actually think about what to do? We think that's a major focus for a lot of foundational model companies, and we expect it to make AI more capable of performing higher-level cognitive tasks like planning and reasoning over the next year. Later today, we'll hear more about it from OpenAI's Noam Brown.
Thirdly, we predict that we're seeing an evolution from interesting consumer applications or prosumer applications, where you don't really care if the AI says something wrong or crazy once in a while, to real enterprise applications where the stakes are very high, like hospitals and defense. The good news is that there are a variety of tools and techniques emerging that can help these LLMs sometimes reach the reliability range of five nines, from RLHF to just-in-time training to vector databases. I am sure that you will be able to have an exchange of views on this later today. I think a lot of people in this room are doing really interesting things to make LLMs more reliable in production.
Finally, we expect to see a large number of AI prototypes and experiments put into production in 2024. What happens when you do that? That means latency is important. This means that cost matters. This means that you care about model ownership. You care about data ownership. This means that we expect the balance of computing to start shifting from pre-training to inference. Therefore, 2024 will be an important year. When transitioning to production, some applications will face tremendous pressure and expectations. We have to do the right thing. Next, I will turn the topic to Konstantin, who will help us explore the possibilities of artificial intelligence in a longer time frame. Thank you, Sonia. Thank you all for joining us today.
So, what question did Pat just ask? Why is this question so important? Why are we all in this room? Sonia just introduced us to what to do now?What stage is AI at the moment?What are we going to talk about next? We're going to take a step back and think about what this means in the context of broader technology and the concept of society as a whole. Therefore, there are many types of technological revolutions. There has been a revolution in communications like the telephone. There has been a revolution in transportation like locomotives. Productivity revolutions are taking place, such as the mechanization of grain harvesting. We believe that AI is primarily a productivity revolution. These revolutions follow a pattern. It started as a human with tools and transformed into a human with a machine assistant. Eventually, this will be transferred to humans with machine networks. Both of the predictions we will discuss in this section are related to the concept of human use of machine networks.
Let's look at a historical example. The scythe has been used as a human tool for more than 10,000 years. The mechanical harvester is a single machine system invented in 1831 by humans and machine assistants and used by humans. Today, we live in an era of having a combine. Combine harvesters are made up of tens of thousands of machine systems, forming a complex network. We have started to use the language of artificial intelligence to describe certain things. For example, we can refer to a single machine participant in a system as an "agent". Today, we had an in-depth discussion on this topic, including the topology of agents and how information is transmitted between agents. We came to refer to this type of transmission as "inference". Essentially, we're building a very complex abstraction layer on top of AI primitives.
Today, I will give two examples, which are what we are experiencing in our knowledge work. First, let's take a look at the software. Software development started out as a very manual process. Ada Lovelace, who wrote logic programming with pen and paper, was able to complete calculations, but without the help of a machine. Now, we live in an era where we have important computer assistants, not only computers, but also integrated development environments and more and more technologies to accelerate software development. We are entering a new era in which these systems work together in a complex network of machines. What you see is a series of processes that work together to produce complex engineering systems. What you see here is agents working together to generate code, not one at a time, but actually in harmony and harmony.
The same pattern is common in writing. Writing is a human process, a human being, and a tool. Over time, this has evolved into a collaboration between humans and machine assistants. Now we have a guy who is actually using not an assistant, but a network of assistants. I'll tell you in my own personal workflow that now, whenever I call an AI assistant, I'm not just calling GPT-4, I'm calling Mistral-Large, I'm calling Claude-3, I'm calling them to work together and also against each other to get better answers. This is the future we see right before our eyes.
So, what does this type of revolution mean for everyone in this room? Frankly, it has an impact on everyone outside the room. From a cold economic point of view, this means significant cost reductions. This chart shows the number of employees required by S&P 500 companies to generate 1 million in revenue, a number that is rapidly declining. We are entering an era in which this will continue to decline. What does this mean? Faster, less. The good news is that it's not about us being able to do less, it's about allowing us to do more. We will talk about this in the next set of predictions. It is also fortunate that there has been deflation in all areas where we have made such progress in the past. I'll take computer software and accessories as an example. Since we're constantly building on each other, the processes of computer software actually cost less over time. Television is also here, but some of the most important things for our society, like education, college tuition, medical care, housing, are growing much faster than inflation. It is perhaps a very happy coincidence that AI is expected to help reduce costs in these and many other key areas. So this is the first conclusion about the long-term future of AI as a huge cost driver of the productivity revolution that will be able to help us do more with less resources in some of the most critical areas of society. The second conclusion has to do with what it's really doing? A year ago, we had Huang make a powerful prediction on stage. In the future, he said, pixels will no longer be rendered, but generated. Any given image or even information will be generated. What does he mean by that?
Well, as everyone here knows, historically, images have been stored as rote memories. Let's consider the letter A, ASCII character number 97. Okay, if we use very simple black and white, store it as a pixel matrix, whether it is present or not. The presence or absence of these pixels. Well, we are entering a period in which we are already representing concepts like the letter A, not as a storage for rote memorization, not as the presence or absence of pixels, but as a concept, a multidimensional point. What we need to consider is the conceptual image of the letter A, which can be extended to any given format of the letter A. In this multidimensional space, there are countless different fonts, and we sit in the center of this space. So, where do we go next? Amazingly, computers are now beginning to understand more than just the point of this multidimensional space, more than just how to take it, render it, and generate an image like Huang said. We are now at a stage where we are able to put this understanding in concrete context. The computer will understand the letter A, be able to render it, understand that it is a letter, understand that it is an English letter, and understand what it means in the broader context of this rendering. The computer will see the word "multidimensional" and will not even consider A, but understand the full context of why the word is proposed. It's amazing that this future is the way we think, the way we think. We no longer store rote memorization pixels in computer memory. That's not how we think. I'm not told that the letter A stands for the presence or absence of pixels on the page. Instead, we'll treat it as a concept. Powerfully, that's what we've been thinking about for thousands of years. 2,500 years ago, my fellow Greek man Plato said that the idea of Plato's form is something that we all agree on, and that we all strive for. In the example of the letter A, you have this concept, or the software engineering concept around which we were actually able to build a model.
So, now that we've talked about the second pattern, the idea that we're going to generalize inside the computation itself. What does this mean for each of us? It means a lot for company building. Today, we've integrated it into specific processes and KPIs. Just mentioned how Clarne leverages this to accelerate their KPIs around customer support. They know that they have certain KPIs that they can drive achievable, and they can have a system that can actually retrieve information, produce a great customer experience. Tomorrow, this is already happening with the new user interface. This could be a different interface for the actual support of the communication style.
And that's where I'm personally really excited because this future is going to present concepts, because this future is going to generate everything, and eventually the whole company might start working like a neural network. Let me explain with a concrete example. It's a cartoon. As with everything in this demo, in reality, everything is sequential. These are discrete. Here's a cartoon of the customer support process. Your customer service has specific KPIs. These are driven by text-to-speech, language generation, customer personalization, and more. This is fed into the sub-mode, subtree, that you are optimizing. Eventually, you'll actually get a fully connected diagram here. In fact, you'll get feedback from language generation to final customer service KPIs. In a way, this will become an abstraction layer where the neural network manages, optimizes, and improves customer support.
Now, let's consider unique customers, which is another part of the important job of building a business. Well, then again, you have AI primitives from language generation to growth engines to add customization and optimization. All this will merge with each other again. The strong conclusion here is that eventually these abstraction layers will become interoperable, enabling the entire company to function like a neural network. This is where the rise of one-man companies comes in. One person's company will allow us to not only do less, but do more. More problems can be solved by more people, creating a better society.
So, what's next? The reality is that the people in the room will decide what to do next. Ladies and gentlemen, you are the builders of the future. We are excited about the future because we believe that AI can help us reduce costs and increase productivity in some of the most critical areas of society, such as providing better education and producing a healthier, more productive population. That is what we are convening this group for today. We want each and every one of you to join the discussion about how we can use our technology to abstract away the complex and mundane details to truly build something more powerful for the future. Thanks again.
来 源 | Web3天空之城
☞ Business Cooperation: ☏ Please call 010-82306118 / ✐ or to [email protected]