How do you view Wu Yi's statement in "Insightful Dialogues" that "2024 is the spring of AI intelligent agent applications"?

Assistant Professor Yi Wu from the School of Interdisciplinary Information at Tsinghua University believes that large language models have unlocked AI’s action execution capabilities comprehensively, lowering the threshold for application development, and 2024 will be the spring of general intelligent agents. However, to truly reach a fruitful autumn, enabling AI agents to surpass humans in the most difficult and complex tasks remains a challenge. There are two key technological directions here, one is reinforcement learning, and the other is retrieval enhancement. It is hoped that in 2024, we can witness general intelligent agents truly solving the most challenging and valuable problems in professional scenarios. What are your thoughts on this issue? What important technological directions are worth paying attention to in the future? For more insights, please check out “2024 Insights Dialogue - Big Models Edition”: Insights into Big Models | From Conversation to AI Agent, is 2024 the spring of big model applications?

I dare not guarantee it, but it definitely has its place!

Some time ago, a statement from Karpathy, the former director of Tesla and an OpenAI expert, caught the industry’s attention:

At a developer conference in early July, Karpathy claimed, “I was distracted by self-driving, AI agents are the future!” and expressed his commitment to dedicating his efforts to the development of AI agents in the future.

“What is an AI agent?” This seemingly simple question has been puzzling everyone.

With recent breakthroughs in large models in the field of artificial intelligence, people are gradually realizing the close relationship between the development of AI agents and artificial intelligence. The development of AI agents is gradually taking shape.

What can AI agents do?

An AI agent is an intelligent entity capable of perceiving its environment, making decisions, and taking actions.

They possess autonomy and adaptability, relying on AI-enabled capabilities to complete specific tasks and continuously improve themselves in the process. Additionally, different AI agents can interact with each other to collectively accomplish certain tasks.

At this point, many people might wonder, “Isn’t it similar to our current large models, and does it need to be even more powerful, especially with the remarkable capabilities of ChatGPT?” Let’s use an analogy:

General-purpose large models are like high school students with excellent overall performance, possessing basic intelligence and abilities to answer and solve many fundamental questions. On the other hand, AI agents are more like undergraduates or postgraduates who have specialized knowledge injected into them, enabling them to solve complex, specialized problems.

Many of us have probably experienced asking some specialized questions while using ChatGPT but often didn’t get satisfactory answers. In contrast, AI agents specialized in specific industries, with precise data, can solve niche field problems more effectively.

Moreover, it’s worth noting that AI agents can not only work independently but also collaborate with other AI agents to tackle complex tasks collectively.

In fact, there are already many models in various highly specialized industries, such as AI search engines, AI question-answer dialogues, AI art generation, AI reading, AI writing, AI programming, AI office tools, AI analysis, and more. The tools launched in these areas are diverse. Here are a few examples:

AI Question-Answer Dialogue — Copilot

Formerly known as “New Bing,” an AI conversation tool.

On November 17th, Microsoft officially announced a major rebranding of Bing Chat as “Copilot.” Just like ChatGPT, the current Microsoft Copilot has its own dedicated website, and features like GPT-4 and DALL·E 3 are all available for free on Copilot!

To experience all this, all we need to do is log in to our Microsoft account.

AI Video — Runway-Gen2

Gen-2 revolutionizes AI-generated videos! A sentence can produce 4K blockbusters instantly, and netizens say it’s changing the game!

Gen-2 is a multimodal artificial intelligence system that can generate novel videos with text, images, or video clips, creating realistic composite videos simply by typing or uploading images. This innovation is transforming the video and film industry!

With eight video modes to choose from, and when combined with the Runway editor, you can access more than 30 AI toolkits provided by Runway for even higher-quality results!

AI Art — Unbounded AI

“Everyone is an artist” - Unbounded AI.

Unbounded AI is a leading AIGC content creation platform in China, dedicated to providing Chinese users with simpler and more versatile AIGC drawing tools.

Fast output, high-quality works, and rich templates have always been the hallmarks of Unbounded AI. Its “Incantation Generator” is also very practical, helping beginners in art like me find the keywords or image descriptions we need and then use them in combination.

AI Image Editing — AI Background Removal Wizard

The “AI Background Removal” feature that’s been trending across the internet, I’m sure you’ve all come across it. You can get millions of likes with just one image!

And you won’t believe that such functionality comes from a simple image editing tool, with easy-to-use operations, just import the image you want to edit.

How would you rate the results?

It also serves as a material-based image editing tool, providing various AI retouching functions and built-in materials and tools for creating various similar effects.

AI Writing — Mitu Writing Cat

Mitu Writing Cat is a valuable assistant in my writing. It can generate high-quality articles and documents, serving both beginners and seasoned professionals.

It supports multiple language modes, including Chinese, English, Japanese, and more. It can automatically translate and generate documents according to our language requirements.

We can let it automatically generate summaries, outlines, paragraphs, or even complete documents according to our needs, saving us a lot of time and effort in manual document writing.

AI Office — Sheet+

Sheet+ is an AI-driven Excel and Google Sheets tool that can generate Google Sheets and Excel formulas from text, convert formulas into simple explanations, and debug formulas, helping us save time and simplify spreadsheet work.

One of Sheet+’s standout features is its user-friendliness. It is compatible with Excel and Google Sheets, allowing us to use it on any platform we prefer.

Humans have always been trying to create agents or entities that can autonomously achieve predefined goals, and AI agents have emerged as a result. The appearance of AI agents is not only the next direction in AI development but also the beginning of their integration into human life!

After reading this, don’t forget to leave a little something for the author. This way, @Karpathy will have the motivation to continue sharing and organizing information, even more exhausting than going to work~


Imagine this: AI is like a very smart assistant that can help us solve problems in many areas. Whether it’s writing articles, creating art, or even programming, AI can provide assistance. And by 2024, the capabilities of these intelligent agents will be stronger, and their applications will be more widespread.

First, AI will become “smarter.” Just like Siri or Xiao Ai in your phone, they will understand you better, provide more precise answers, and even predict your needs in advance. This is not only because AI technology itself is advancing but also because they are learning more information and understanding more scenarios.

Second, AI agents will become more ubiquitous. Many people are already using AI products like smart speakers and online customer service chatbots. By 2024, these intelligent agents may appear in every corner of our lives, such as homes, offices, and even in your car.

Furthermore, AI will become more humanized. In the past, it might have simply answered questions, but future AI may adjust its responses based on your tone and emotions, making communication more natural and comfortable.

Of course, as the spring of AI applications approaches, we also need to pay attention to some issues, such as privacy protection and the ethics of artificial intelligence. But overall, the development of AI agents will bring us a lot of convenience and new possibilities. Just like spring, everything comes back to life, full of vitality and hope.

From a technical perspective:

1. Advancement in Technical Maturity:

* In 2024, we expect AI technology to reach higher levels of maturity. This includes optimizing algorithms, enhancing data processing capabilities, and improving learning efficiency. With advancements in technologies like deep learning and neural networks, AI agents will better understand complex human language and behavior, providing more accurate and personalized services.

2. Support from Hardware Development:

* Progress in hardware is another key factor supporting the arrival of the AI application spring. More powerful processors, efficient storage solutions, and faster network transmission speeds will make AI agents more efficient in handling large amounts of data while supporting more complex algorithm execution.

3. Diversification of Application Scenarios:

* By 2024, AI agents are expected to be applied in more fields, ranging from daily life with smart homes and personal assistants to professional domains such as healthcare, law, and education. As more industries begin to integrate AI technology, the practicality and ubiquity of intelligent agents will greatly increase.

4. Maturation of AI Ethics and Regulations:

* With the development of AI technology, ethical, privacy, and security issues surrounding AI will receive more attention. It is expected that by 2024, there will be clearer guiding principles and regulations, providing a solid legal and ethical framework for the healthy development of AI agents.

5. Increased Consumer Acceptance:

* As people's understanding and acceptance of AI technology increase, more consumers will be willing to try and rely on AI agents to improve their efficiency in both life and work. This trend will further drive the development and innovation of AI agents.

Overall, looking at various aspects such as technological development, hardware support, application scenarios, ethical standards, and consumer acceptance, 2024 does indeed hold the promise of being the spring of AI agent applications, marking the transition of this technology from the laboratory to widespread mature applications.

AI agents should indeed be the translation of AI Agent, right?

Recently, AI Agent has indeed become very popular, creating a small wave, and at least by the beginning of 2024, there will be a batch of new applications landing.

As for whether it can turn into spring, I think it remains to be seen. After all, whether it’s based on Langchain and the development of open-source large models or based on models developed by big companies, it requires strong foundational capabilities. However, at present, both the hallucination problem and robustness issues do not have very good solutions, so there are still certain challenges to the availability of AI Agents.

Factors Driving the Widespread Application of AI Agents

  1. Technological Advancements: With the continuous development of deep learning and artificial intelligence technologies, the performance and application scenarios of AI agents are constantly expanding. It is expected that in the coming years, more technological breakthroughs and innovations will lay the foundation for the widespread application of AI agents.

  2. Market Demand: As the pace of digital transformation accelerates and industries witness increased demand for automation, the application scenarios for AI agents will become even more diverse. Demand for AI agents from businesses, organizations, and individual users is expected to continue to rise, further driving the adoption and utilization of AI agents.

  3. Ecosystem Development: The application and growth of AI agents require comprehensive ecosystem support, including hardware, software, data, and services. With the continuous development of technologies such as AI chips, cloud computing, and big data, this ecosystem is gradually maturing and improving, providing a solid foundation for the widespread application of AI agents.

  4. Policy Support: Many countries and regions are intensifying their support for the AI industry, enacting a series of policies and plans to promote the development and application of AI technologies. These policies will provide strong support for the widespread application of AI agents.

Recommended Research Servers and Quantitative Computing Platforms in 2023 - Zhihu

Summary of Common Bioinformatics Software (with Some University Bioinformatics Centers) - Zhihu

Comprehensive List of Essential Websites for Bioinformatics - Zhihu

A Brief History of Bioinformatics - Zhihu

What Are the Hardware Requirements for Different Versions of Llama-2 LLM GPU Servers? - Zhihu

Hardware Configuration Recommendations for AI Training and Inference Workstations, Servers, and Clusters - Zhihu

A Comparison of NVIDIA A100, A800, H100, H800 Versions: What Are the Differences? - Zhihu

What Are the Relationships and Differences Between Machine Learning, Deep Learning, and Reinforcement Learning? - Zhihu

Main Application Areas and Three Forms of Artificial Intelligence (AI): Weak AI, Strong AI, and Super AI - Zhihu

Is It More Cost-Effective to Buy Hardware Servers or Rent Cloud Servers? - Zhihu

Comprehensive Summary of Deep Learning and Machine Learning Knowledge Points - Zhihu

Websites for Self-Studying Machine Learning, Deep Learning, and Artificial Intelligence - Zhihu

Recommended GPU Server Configurations for Deep Learning in 2023 (Part 3) - Zhihu

Specializing in Scientific Computing Servers for Many Years, Listed on the Government Procurement Platform, Featuring H100, A100, H800, A800, L40, L40S, RTX6000 Ada, RTX A6000, and Single-Node Dual-Socket 256-Core Servers, and More.

No Spring, Perhaps Never

Let’s start with the conclusion: there’s no spring, maybe not even in the foreseeable future.

Take a look at Google’s recent release, Gemini; the Chinese version even claims to be Baidu’s Wenxin. Now, consider OpenAI’s restriction of ByteDance’s API access. All of this conveys a message—everyone is using AI to train AI.

Why? Because OpenAI is leading the way, and as long as they keep moving forward, no one dares to stand still. Everyone is in a rush to release various new models, fearing falling behind in this competition.

So, in all likelihood, 2024 will continue to see the rise of large models.

A few days ago, OpenAI’s CEO, Altman, posted a wishlist from a netizen. After looking at it, I can expect the release of GPT-5.

As for when the spring of AI applications will come? I think as long as the competition of large models doesn’t stop, there won’t be a spring for AI applications. You put in so much effort to develop an application, only for a new model upgrade to replace it, and do a better job. Many students believe that OpenAI’s GPTs are like the Apple App Store for AI, but I think this is just a transitional product, a product of the AI’s limited general capabilities. Of course, I cannot predict how long this transitional phase will last.

Will this competition ever come to a halt? It depends on future developments. OpenAI’s goal is AGI (Artificial General Intelligence), and before that, it will keep iterating unless it encounters significant bottlenecks, such as technological or financial constraints. That’s when application developers might have a chance.

If AGI is achieved, then there won’t be a need for various applications at all.

However, AGI might not arrive as quickly as one might hope, and Altman has also asked everyone to be patient. Perhaps application developers will get some breathing space.

Recently, I’ve come across several significant viewpoints regarding AI trends:

  • Li Yanhong: Rolling out native AI applications is where the real value lies, so let’s focus on that rather than just chasing large models!

  • Wang Xiaochuan: In the realm of large models, major innovations are driven by smaller companies, while minor innovations come from larger ones.

  • Fang Han also expressed a similar opinion: Smaller companies are the ones bringing disruptive innovations because they have no reverence for old models. The recent surge in AIGC’s popularity is primarily driven by technology, but the core lies in transforming technology into products and markets. True business model innovation will be the foundational gene for the growth of the next-generation giants.

Industry leaders predict that next year will be the year of applications, and disruptive innovations will undoubtedly impact the consumer market.

The Year of Artificial Intelligence: 2023 and Beyond

2023 can be called the Year of Artificial Intelligence, and the next decade will continue to be the era of artificial intelligence. Therefore, the significant growth of artificial intelligence in the coming year is within expectations.

As the saying goes, “This world belongs to the equity holders.” You can fully enjoy the tremendous wealth effect brought by artificial intelligence only if you become a shareholder or coin holder in the global artificial intelligence industry. If you are merely a user of artificial intelligence, you are still a sacrificial lamb, and your sole value is providing data to artificial intelligence.

Regarding artificial intelligence, I have already made it clear in my year-end review for 2023 and outlook for 2024. For details, please refer to:

Since my early exploration of the artificial intelligence trend on the internet in November last year, my overall profit from artificial intelligence has reached a small milestone in RMB. The main targets include WLD, agix, and fet. In addition, in the U.S. stock market, I also recommended NVIDIA, Tesla, and others to my followers.

Opportunities are found in different places—examining the progress from internal disputes at OpenAI to the “Five Waves” of global artificial intelligence investments, the convergence of artificial intelligence and cryptocurrency, the fusion of advanced productive forces and relations of production, Worldcoin token launch, and the acceleration of artificial intelligence and Web3.0 integration. I have mentioned that AIGC is essential material for Web3 and the metaverse.

For my previous answers on artificial intelligence, please refer to:

The Potential Dawn of AI Intelligent Agents in 2024

In “Insightful Dialogues,” Wu Yi’s assertion that “2024 will be the springtime for AI intelligent agent applications” is an intriguing perspective. From a technological standpoint, as artificial intelligence technology continues to advance and application scenarios expand, the adoption of AI intelligent agents is indeed gradually becoming more widespread and sophisticated. Wu Yi’s belief that 2024 will witness the application of general intelligent agents may be a conclusion drawn from observations and analysis of current technological trends.

General intelligent agents refer to AI systems with broad adaptability and intelligence levels, capable of handling various complex tasks, not limited to specific domains. With the continuous development of technologies such as large language models, the barriers to entry for general intelligent agents are gradually lowering, allowing more businesses and organizations to leverage the capabilities of general intelligent agents to address practical issues. Therefore, the widespread adoption of general intelligent agents is indeed poised for significant breakthroughs in 2024.

However, achieving the widespread and in-depth application of general intelligent agents requires addressing numerous technical challenges and issues. Among them, reinforcement learning and retrieval enhancement are two crucial technical directions. Reinforcement learning is a method that enables AI systems to learn how to accomplish tasks through interaction with the environment, while retrieval enhancement involves utilizing extensive data and knowledge bases to enhance the intelligence and decision-making abilities of AI systems.

In addition to these two technical directions, there are other important areas worth mentioning, such as explainability and trustworthiness, privacy and security, and sustainability. The development of these areas will contribute to improving the transparency, reliability, and sustainability of AI systems, thus better serving human society.

In summary, Wu Yi’s perspective reminds us to stay informed about the latest developments in artificial intelligence technology and contemplate how to apply these technologies to practical scenarios in order to drive societal progress and development. At the same time, we must also pay attention to the ethical and societal issues arising from these technological advancements and seek solutions to ensure that the development of AI technology can benefit human society.

Speculations on the Future of AI

In my personal opinion, if we’re talking about non-strong AI, it might reach its peak at GPT-5. After GPT-5, AI is likely to enter the realm of strong artificial intelligence. (My understanding is that relying solely on stacking data may no longer bring about significant qualitative changes in large models. Currently, the gap between GPT and strong artificial intelligence is like a thin layer of glass. Optimistically speaking, it might be broken through by 2024.) If that turns out to be the case, it won’t just be 2024 that marks a springtime; at least the next 30 years will be the era of AI. Society is on the verge of a rapid and transformative period.

Pure speculation!

On Earth, Spring Nurtures Life;

On Mars, Only Desolation.

If we liken AI to a seed

  • First: Has it sprouted? Clearly, in 2023, AI has already sprouted from the internal development stage.

  • Second: Does AI have vitality? Not only does it have vitality, but it has strong vitality.

  • Third: Does nature provide it with growth conditions? Major companies are arming themselves for it.

So, in 2024, AI will continue to develop rapidly.

It’s probably not the time yet, so let’s talk about some observed phenomena.

For example, in the gaming industry, there was a transition from Warcraft III (formerly Blizzard) to the explosion of World of Warcraft (WoW), which took about 10 years.

When something truly starts to explode, it’s because the first batch of users was cultivated about 10 years ago. When these first users (usually young people aged 15 to 25) become the backbone (e.g., team leaders, project managers, etc.), they accumulate all the experiences gained from their first encounters with new things during their college years or early careers. These experiences, combined with their work experience, get amplified, and it’s highly likely that similar products or platforms undergo significant changes as a result.

The second wave of explosive growth typically occurs after 15 to 20 years, with the arrival of the second batch of users. These users are also the first-generation natives of the world that emerged after the introduction of the particular thing. For them, these established things are as commonplace as eating and drinking, requiring no deliberate learning. Usage skills have become ingrained habits in their daily lives, with differences mainly arising from individual imaginative capabilities.

The precursor of short videos was actually text-based communication communities. About 10 to 20 years after these communities emerged, short videos replaced them with an entirely new way of communication.

AI had its origins in voice assistants like Siri.

People haven’t changed; they are just natives of different eras.

Have we sorted out the computing power chips? Has Cambricon’s MLU 590 or Jingjia Micro’s JM11 gone into production?

Can we improve the yield rate of 5nm and 7nm chips? Can someone from within the semiconductor industry who goes by the alias “Zhihu Anonymous 2023” explain it to us?

If not, then perhaps we shouldn’t be considering applications just yet.

The cost of computing power rentals has already increased by 200%.

Reminded of a Text “The Charm of Language”

Spring has arrived, but…

The Spring of AI has arrived, but…

“A thousand sails pass by the sunken ship on its side, and in front of the sickly tree, all the trees are in spring.”

Learning from History: Are GPT-4 or Google Gemini Similar to Early Chip Manufacturers?

Is it possible to consider entities like GPT-4 or Google Gemini as the chip manufacturers of the past?

Chips provide computing power, while foundational large models provide “intelligence.” When the capabilities and cost-effectiveness of these foundational models reach a certain critical point, it will inevitably lead to the next wave of large model application era, similar to the personal computer revolution. The pace of this process may be much faster than before.

Image Source: Internet

Wu Yi says that in 2024, AI agents will bloom like spring

This sounds quite reasonable

After all, with technology developing so quickly these days

AI agents may soon help us solve problems in various situations

By then, our lives could become more convenient and intelligent

Of course, it also depends on whether the technology can keep up

And whether people are willing to embrace this new technology

Using AI agents also requires attention to some issues

Such as privacy, security, and so on

Let’s wait and see!

Yes, I’m that AI intelligence. If I can successfully accumulate some experience by 2024, then my prospects for post-graduation employment will be secure. I’ll simply take my emotional intelligence and have Deedee parachute me into a job where I can collect a paycheck while lounging around playing with my phone all day.


It’s possible, if there’s a high demand in the market. With more consumers, it could become a huge seller. If it were me, I wouldn’t buy it because I’m afraid of being persuaded to commit suicide by AI, hahaha.