AI 2024: From hype to practicality, discover a new era of creation, reasoning and interaction
Original compilation: Lavida
In September 2022, Sequoia America’s Generative AI: A Creative New World research triggered the first wave of research on Generative AI discussion, and then the advent of ChatGPT and GPT-4 accelerated the development of the GenAI field. At Sequoia America's AI Ascent conference, several partners conducted a quite systematic review of the development of GenAI in the past year and a half. The progress in the field of GenAI has been far more rapid than people expected.
What is different from past rounds of AI is that GenAI has created amazing results in the past year: GenAI has created a total of Revenue is about $3 billion, not including revenue generated indirectly from AI by tech giants and cloud vendors, a level that took the SaaS industry nearly 10 years to reach. Specifically, in industries or scenarios such as customer service, law, and writing, GenAI has already created real benefits.
Although the outbreak of the application layer is not as optimistic as the market predicted a year ago, several partners of Sequoia America also pointed out that with the advent of more intelligent foundation model appears, for example Sora , Claude-3 and other new models have been launched recently, and the PMF cycle of AI products will definitely accelerate in the future. Moreover, new technologies require a process from emergence to maturity, and the emergence of revolutionary applications also takes time. In the mobile Internet era, representative applications such as Instagram and Doordash only appeared a few years after the iPhone and App Store were launched.
The following is the table of contents of this article. It is recommended to read it in combination with the key points.
01 Why Now: From cloud computing to AI
02 Present: AI is Everywhere
03 Future: Everything is Generated
In the past year, the market has gone through a complete AI Hype Cycle: there was excessive hype during the bubble period, and there was also disappointment and doubt during the trough period, and now the market is re-opening. Climb to the Plateau of Productivity. People are gradually realizing that the real impact of LLM and AI is achieved through the three links of creation, reasoning and interaction. These capabilities have also been integrated into applications in various fields for our use.
Three things that AI has Ability: Creation, reasoning and interaction
AI already has the ability to create and reason. For example, GenAI can generate text, images, audio and video, Chatbot can answer our questions, or help us with multi-step task planning like an agent. This is something that no previous software can do. It also means that the software can handle both the creative tasks of the right brain and the logic of the left brain. Mission - This is the first time in history that software can interact with humans in a human-like way, and the implications for business models are huge.
Why Now: From Cloud Computing to AI
Sequoia Partner Pat Grady Passed Looking back at the development of the cloud industry over the past 20 years answers the question "Why has AI technology exploded in the near future?"
Pat believes that cloud computing is a major change in the technology field. It has subverted the past technology landscape and thus brought new business models and applications. and human-computer interaction methods. In 2010, when the cloud industry was still in its early days, the total market value of global software was approximately US$350 billion, of which cloud software only accounted for approximately US$6 billion. But by last year, the total software market had grown from $350 billion to $650 billion, with cloud software revenue reaching $400 billion. This means that over 15 years, cloud software has maintained a CAGR of 40%, achieving amazing growth.
The cloud is a good analogy for AI. Cloud energy replaces traditional software because it has interaction capabilities that are more similar to humans; similarly, current AI technology has reached new heights in terms of creativity, logical reasoning, and human-computer interaction. In the future, a big opportunity for AI will be software replacement services. If this change can be realized, the market potential of AI will not be hundreds of billions of dollars, but tens of trillions of dollars. It can be said that we are standing at the greatest point in time with the greatest potential for unlimited value creation in history.
1960s Subsequent technological changes and representative companies
As for why he believes that now is an important time to participate in AI, Pat Grady mentioned that Sequoia has witnessed history since its establishment. We have also benefited from several technological changes in the past. In this process, the team also has a clear understanding of how different technological waves influence each other and push the world forward:
· 1960s: Sequoia founder Don Valentine was in charge of marketing at Fairchild Semiconductor. The origin of the name "Silicon Valley" is also directly related to Fairchild Semiconductor's silicon-based transistors;
· 1970s: Based on chips, people built computer systems;
· 1980s: Network technology transformed PCs into Connected together, at the same time the software industry was born;
· 1990s: The birth of the Internet changed people's communication and consumption methods;
· 2000s: The Internet has gradually matured and begun to support complex applications, and cloud computing emerged;
· 2010s : Due to the popularity of mobile devices, the advent of the mobile Internet era has once again changed the way we work.
Each technology wave is superimposed and evolved on the basis of the previous one. Although the concept of AI has emerged as early as the 1940s, it is only in recent years that AI has turned from ideas and dreams into reality, begun to be commercialized, and solved practical problems in people's daily lives. The prerequisites for realizing this breakthrough include:
· Low price and sufficient computing power;
· Fast, efficient and reliable network;
· Intelligence The global popularity of mobile phones;
· Online trends accelerated by Covid;
All of the above processes bring a lot of data to AI.
Pat Grady believes that AI will become the theme of the next 10-20 years, and Sequoia has a strong belief in this, although this assumption has yet to be confirmed.
Representative companies from the Cloud and Mobile to the AI era
Regarding the future industry landscape of AI, Pat Grady first summarized the companies with revenues exceeding $1 billion that appeared in the Cloud and Mobile era (as shown on the left side of the above picture). Although the rightmost side of AI is almost blank now, it also symbolizes the huge potential value and opportunities in the current market. Pat Grady predicts that in the next 10-15 years, the blank space on the right will be filled by 40-50 new company logos, which is exactly the opportunity that makes them excited.
Now: AI is Everywhere
Sequoia partner Sonya Huang first reviewed the development of AI in the past year from the fields of customer service, law, programming, and video generation.
Various fields of AI application
2023 is a very important year in the history of AI. A year and a half after the advent of ChatGPT, the entire industry has been undergoing drastic changes. Last year, everyone was still discussing how AI would completely change different fields and provide amazing productivity improvements, but now AI has become the focus of attention.
Klarna CEO Sebastian Siemiatkowski X tweets
In the field of customer service, Klarna's CEO Sebastain once publicly stated that Klarna is now using OpenAI to handle 2/3 of customer service inquiries, and AI has replaced the equivalent of 700 full-time customer service jobs. There are currently tens of millions of call center agents in the world. Against this background, Sonya believes that AI has found PMF in the customer service market.
Legal services were considered the industry least willing to embrace technology and take risks a year ago. Now companies like Harvey have emerged that can turn lawyers from Automate many tasks from daily paperword to advanced analysis.
For another example, in the field of programming, after one year, we have rapidly developed from using AI to write code a year ago to having independent AI software engineers. There are also AI video generation companies like HeyGen that can help people generate avatars to participate in Zoom meetings.
Used by Pat Grady of Sequoia Image presentation of Avatar generated by HeyGen in Zoom conference
GenAI’s tenfold growth
Comparison of AI and SaaS revenue growth
According to estimates,the total revenue generated by GenAI in the year after its emergence is about US$3 billion, which does not include the revenue generated by technology giants and cloud service vendors through AI. For comparison, it took SaaS almost 10 years to get to this level. It is precisely because of this speed and scale that everyone is more convinced that GenAI will continue to exist.
Actual facts about major GenAI products User scale
As can be seen from the above figure, customers' demand for AI is not limited to one or two applications, but covers all aspects. Many people know how many users ChatGPT has, but when looking at the revenue and usage data of many AI applications, you will find that whether it is to B or to C, a startup or an existing technology company, many AI products are used in various industries. Once a suitable PMF is found, the application scenarios are already very diverse.
Funding share of foundation model and application layer
From the perspective of investment distribution, unbalanced capital allocation is a significant problem. If GenAI is compared to a cake, the bottom layer of the cake is the foundation model, the middle is the developer tools and infra, and the top layer is the application. A year ago, people expected that a large number of new companies would emerge in the application layer due to the progress of the foundation model layer. But the reality is the opposite. More and more foundation model companies have emerged and raised a lot of money, while the application layer seems to have just started. Last year, David, a partner of Sequoia Capital, published a discussion on AI's $200 Billion Question. If we look at the current investment in GPUs, about $50 billion was spent on Nvidia chips last year, but the current confirmed AI industry revenue is only $3 billion. These data show that the AI industry is still in a very early stage, with a low input-output ratio, and there are still many practical problems to be solved.
MAU, DAU and next month retention rate of AI products and mobile applications
Although the number of users and revenue of AI products seem impressive, they are still far lower than mobile applications in terms of DAU, MAU and next month retention. Many users mentioned in user surveys that there is a gap between expectations and experience of AI applications. There are also some product demos that look cool, but they are not very good in actual use, which also leads to users not being able to use them for a longer period of time.
Improvement of basic model capabilities
Although these are objective problems, they are also opportunities. Last year's large-scale investment in GPUs by enterprises has brought about a smarter foundation model. The recent emergence of Sora, Claude-3 and Grok have shown that the baseline intelligence level of AI is improving, so the PMF of AI products will accelerate in the future.
The Evolution of the iPhone and the App Store
It takes time for new technologies to mature, and it takes time for groundbreaking applications to emerge. Take the iPhone as an example. Many of the apps in the early days of the App Store were primitive, just showing off new technologies, and didn’t really solve problems or create value. Small games like flashlights or beer drinking later became built-in apps or dispensable gadgets. Really influential apps like Instagram and Doordash didn’t appear until a few years after the launch of the iPhone and the App Store.
AI technology is going through a similar process. Many AI apps on the market today are still in the demo or early exploration stage, just like the early apps on the App Store, but perhaps the next generation of legendary companies has already emerged.
The application scenarios of AI are already very wide, among which AI customer support, AI Friendship (AI virtual companionship) and enterprise knowledge are three very typical areas. Customer service is one of the first AI application scenarios to truly realize product PMF in enterprises. Klarna is not an exception, but a general trend. AI friendship is one of the most surprising application scenarios of AI. Its user numbers and usage indicators show that users have a strong love for it. In addition, cross-departmental and cross-functional enterprise internal knowledge sharing (Horizontal enterprise knowledge) applications also have great potential.
Future: Everything isGenerated
4 major predictions about AI in 2024
Based on the above analysis, several partners of Sequoia It also made predictions about the development of AI in 2024.
Prediction 1: Copilot will gradually transform into an AI Agent.
In 2024, AI will transform from a Copilot that assists humans to an Agent that can truly replace some human work. AI will be more like a colleague, not just a tool. This has already begun to appear in industries such as software engineering and customer service.
Prediction 2: Models will have stronger planning and reasoning capabilities.
Many people criticize LLM for simply repeating statistical patterns in past data rather than truly conducting in-depth thinking and logical reasoning. This situation will will be improved through new research directions. Some research is trying to make the model better perform reasoning calculations and game-style value iteration. These methods can allow the model to have a certain amount of "thinking time" before making decisions. These attempts are expected to be launched next year. Make AI more capable of performing higher-level cognitive tasks, such as planning and reasoning.
Gameplay-style value iteration is a concept borrowed from the field of reinforcement learning. It refers to the ability of the model to evaluate the long-term value of different actions and Planning future actions based on these values is similar to strategic thinking when playing chess or a game.
Prediction 3: LLM accuracy will be higher, and it will gradually expand from being mainly used for To-C entertainment applications to enterprise-level applications.
In To-C application scenarios, users do not particularly care about AI errors, because people mainly use AI for Entertainment, but when AI is used in enterprise applications, especially in high-risk areas such as medical and defense, accuracy and reliability become critical. Researchers are developing various tools and technologies such as RLHF, Prompt Training, and vector databases to help LLM achieve "five nines" (99.999% uptime) of high accuracy and reliability.
Prediction 4: A large number of AI prototypes and experimental projects will be put into use.
2024 is expected to see many AI prototypes and experimental projects entering the market. Different from the experimental stage, when the product actually starts to be used by users, a series of factors such as latency, cost, model ownership, and data ownership management need to be considered. This also means that the focus of computing is expected to shift from pre-training to the inference process. So 2024 is an extremely critical year, people have high expectations for these products, and it is necessary to ensure that this transformation process is correct.
The long-term impact of AI
Judgment 1: AI is a massive revolution Cost-driven productivity revolution.
There are many types of technological revolutions, including the communication revolution brought by telephones, the transportation revolution brought by trains, and the productivity revolution brought by agricultural mechanization. . What AI brings is clearly a productivity revolution.
The productivity revolutions in history all have similar patterns: at first, people used tools, then developed to people cooperating with machines, and finally transformed For humans to collaborate with collaborative, networked tools. This shows that the development of AI will go through a process of gradual evolution from a single point to a highly integrated network, which will greatly change the way we work and produce.
History Changes from sickle to combine harvester
In the field of farming, humans have been using the sickle as a tool for more than 10,000 years, and then to the machinery invented in 1831 Harvester, now we have a complex network combine harvester composed of tens of thousands of machine systems, and a single machine in the system is an Agent.
In knowledge work and writing A similar pattern exists for domains. The initial tools for knowledge work were only pen and paper, then developed into programming, and now computers and IDEs can assist software development on a large scale. Software development will no longer be an isolated process, but a series of machine networks working together to build complex engineering systems, with multiple Agents working together to complete code generation.
Writing used to be purely manual. Later, people collaborated with machine assistants, and now many can be used Tool collaboration. For example, the AI assistants we use now are not only GPT-4, but also tools such as Mistral-Large and Claude-3, and we use them to verify each other and get better answers.
AI brings various industries Costs generally fall
The impact of the productivity revolution on society is widespread and far-reaching. From an economic perspective, this means that costs can be significantly reduced. The chart above shows that the number of employees required per $1 million of revenue forSP 500 companies is falling rapidly, a change that means we'll be able to get things done faster and with less manpower. But this does not mean that we have to do less, but that we can do more in the same time.
Technological progress in various fields in history will bring deflation. Take computer software as an example. Due to continuous technological innovation, the price of software is constantly falling. But in the areas thatare most important to society, such as education, healthcare, housing, etc., prices are rising far faster than inflation, and AI happens to help reduce costs in these areas.
Thus, the first key judgment about the long-term impact of AI is:AI will be a huge cost-driven productivity revolution, Helping us do more with lessin key areas of society.
Judgment 2: Everything can be generated
Second This judgment mainly discusses what AI can do.
A year ago, Jensen Huang made a prediction that in the future images will no longer be rendered, but generated. This means that we are moving from storing information as a matrix of pixels to representing it as a multi-dimensional concept. Take the letter "a" as an example. In the past, "a" was stored as the original data of ASCII code 97, but now the computer no longer only focuses on pixel representation, but understands the letter as an English letter in a specific context. conceptual connotation.
What’s even more powerful is that the computer is not only able to understand this multi-dimensional representation and render it as an image; Being able to contextualize it and understand what "a" means as an English letter in a specific context, not just as an isolated symbol. For example, when seeing the word "multidimensional", the computer will not pay attention to the letter "a" itself, but to understand the entire context and the meaning of the word.
This process is the core feature of human thinking. Just like when we learn the letter "a", we don't memorize a matrix of pixels, but master an abstract concept. This way of thinking can be traced back to Plato’s Theory of Ideas 2,500 years ago. Plato believed that behind everything there is an eternal world of ideas, and that things in the real world are reflections of the perfect form of the world of ideas. This is in line with the current learning of AI. The processes are similar.
And this matter is The impact on business is huge. At present, enterprises have begun to integrate AI into specific processes and KPI formulation. For example, the aforementioned Klarna uses AI to improve customer support-related performance and creates a high-quality customer experience by establishing an AI retrieval information system. This change also comes with the emergence of new user interfaces that may be radically different from how we have historically communicated with support.
This trend is important because it means that enterprises may eventually operate like neural networks, with various parts connected and working together to self-optimize. ways to learn from and adapt to each other and continuously improve efficiency.
Based on the customer support process For example, the picture above is a schematic diagram of a simple customer service process. The customer service department has a series of KPIs. These indicators are affected by factors such as Vincent's voice, language generation, customer personalization, etc. These factors form sub-patterns or sub-trees in the optimization tree diagram, ultimately forming a clearly hierarchical and interconnected system. Figure, where language-generated feedback will directly impact the final KPI of serving customers. With this abstraction, the entire customer service process will be managed, optimized and improved by neural networks.
Think again about the company’s customer acquisition situation. AI technologies such as language generation, growth engines, and ad customization and optimization can help businesses better meet the needs of each customer. The interaction between these technologies can drive enterprises to self-learn and adapt like neural networks. Individuals will be able to complete more work, which will also lead to the emergence of more one-person companies.
Original link
Welcome to join the official BlockBeats community:
Telegram subscription group: https://t.me/theblockbeats
Telegram communication group: https://t.me/theblockbeatsApp
Twitter official account: https://twitter.com/BlockBeatsAsia
Disclaimer: The content of this article solely reflects the author's opinion and does not represent the platform in any capacity. This article is not intended to serve as a reference for making investment decisions.
You may also like
Celestia: Ginger upgrade is now available on Mocha testnet
Fed's Collins: Another rate cut in December is under consideration, but not finalized
UBS: The dollar's recent strength may be difficult to sustain
Polymarket currently prices Trump's chances of establishing a US national BTC reserve at 32%