Source: Synced

Image Source: Generated by Wujie AI

Li Feifei and Andrew Ng discuss the 2024 AI trends.

In the history of artificial intelligence, 2023 has become a very memorable year. In this year, the AI wave led by OpenAI swept the entire technology field, putting practical AI tools into everyone's hands. However, at the same time, the development of artificial intelligence has also sparked widespread discussion and controversy, especially in its commercial applications and future prospects.

Renowned AI expert Rodney Brooks predicted at the beginning of 2024 that AI may be on the verge of a new "AI winter," with the bursting of the bubble posing serious challenges to the industry. His remarks have sparked widespread discussion in the industry: In the new year, will the AI field see more hype or enter a downturn? At the same time, many people are also concerned, will AI replace more jobs?

At the recent Consumer Electronics Show (CES), two renowned AI scientists and Stanford University professors, Li Feifei and Andrew Ng, took the opportunity to discuss these issues. In addition, they also discussed: In the new year, what new breakthroughs will AI technologies achieve? What does the rise of AI agents mean? How can enterprises with concerns about the accuracy of generative AI make good use of this technology?

Click to view original video

For the convenience of everyone to collect specific content, Synced has also organized the video into text:

Host: First, let's discuss the first question, which is related to the current state of artificial intelligence. Last year was undoubtedly a very crazy year for the hype of artificial intelligence.

Our good friend Rodney Brooks tweeted on January 1st: "It's time to put on thick clothes, another AI winter may be coming, and it will be very cold." So, in the new year, will the AI field see more hype or a downturn?

Andrew Ng: I think the media will do things in their own way. But we will not enter a winter, because the commercial foundation of artificial intelligence is stronger than ever, even before the wave of generative AI. Last year, artificial intelligence really took off. It may have driven a market of tens of billions, or even hundreds of billions of dollars, I'm not sure. At least for companies like Google, it's tens of billions of dollars, because showing more relevant ads can bring in a lot of revenue. So the commercial foundation of AI exists.

In fact, the most difficult thing to understand about artificial intelligence is that it is a general-purpose technology. This means it is not only useful for one thing. It's a bit like electricity, if I ask you what electricity is used for, you will find it difficult to answer because its uses are too broad. Artificial intelligence is the same. At the stage we are in today, even if artificial intelligence cannot make any technological breakthroughs, of course it will definitely make progress. Even if not, there are many use cases waiting to be discovered and built around the world. So I am very confident that its commercial foundation will continue to grow.

Li Feifei: Yes, I generally agree with Andrew's point of view. What we are seeing is another turning point in artificial intelligence, which began with the development of large language models, especially the first release of ChatGPT and the subsequent advancement of models.

I agree with Andrew's view that AI is a deepening horizontal technology. As a deepening horizontal technology, it is becoming an important driving force for the next digital or industrial revolution. The fluctuation in public media reports is not the most critical. What is important is that this technology has already taken root in our lives. It is penetrating into all vertical industries and consumer experiences, gradually changing our social, economic, and political structures. This is an indisputable fact, and this change is becoming more and more apparent.

Host: Let's jump to your predictions for the major breakthroughs in 2024. What do you think will be the biggest breakthrough for AI this year?

Li Feifei: Predicting the future is always very dangerous, because what I say wrong will be quoted. In terms of computer vision and the so-called pixel-centric AI field, I do think we are on the verge of very exciting technological advances. We have been focusing on generative AI, and we have been studying diffusion models, some people may have heard of Gaussian splatting or related content. I think in the current technological wave, there are many achievements that are about to break through. I'm not sure if it will mature like the large language models more than a year ago, but I see it more and more. I am very excited about this.

Host: Andrew, what do you think?

Andrew Ng: Yes, I also have a few predictions. First, we have seen the breakthrough of large language models. I agree with Feifei's view on images. So what I see is the transition from large language models to large visual models. Many advances are not only in image generation, but also in image analysis, which can improve the computer's visual capabilities, and this has a great impact on fields such as autonomous driving, anywhere with a camera. That's the first point.

The second point I want to make is, you ask ChatGPT a question, and it will respond immediately. But what I am more interested in is the rise of automated AI agents. When you give an AI system a command, such as "Dear AI system, help me conduct market research and competitive analysis on this company," the agent will not immediately respond, but will plan a series of actions, such as conducting web searches, downloading web pages, and summarizing information. It will work for half an hour, an hour, or a day, and then come back with the answer. Automated AI agents can plan and execute a series of actions. They are just getting started now, but there has been a lot of progress in research and commercialization. I look forward to new breakthroughs in the coming months.

Finally, which may also apply to the CS field, I am very interested in edge AI. I often run large language models on my laptop. I often use GPT-4, and I also often use Bard. But what many people don't know is that it is actually quite feasible to run large language models on your laptop. Although these models are not as large as GPT-4, they are sufficient. I think this is significant. Wouldn't all PC manufacturers want to sell consumers a more powerful PC to use the latest AI technology? Graphics cards are often a motivation for people to upgrade their computers. I think running edge AI on your laptop, PC, or industrial PC is actually much more capable than most people think. For the CS field, I think this will drive the sales of many devices.

Li Feifei: I have a slightly different opinion. You mentioned "autonomous AI agents," but I prefer to call them "assisted AI agents." We have observed in the current large language models and these massive foundational models that dealing with long-tail problems is still a challenge, whether it's in hallucinations or other aspects. In many work environments, focusing on the long-tail distribution is very important to provide high-quality services and products. Therefore, I think it is more likely to see human-machine interactive collaboration. In this mode, assisted AI agents include both autonomous work and collaboration with humans. This is more likely to become a reality than complete automation.

Andrew Ng: Indeed (high five). I actually somewhat agree with your view. I'd like to share my experience. The term "autonomous AI agents" may be problematic, but what I see in the business domain is that most people are more willing to have AI assist humans rather than replace them, especially considering the issue of unemployment, which is a real challenge. I don't want to overlook the suffering of those who are unemployed.

What I have observed is that the decision to use AI to automate or replace human labor is often more based on business economics rather than ethics. Perhaps it should be an ethical decision, but honestly, when I work with companies to build chatbots, I see that most companies conduct very rational economic evaluations. They care about making the best economic decision on how to do things, after all, their competitors are doing the same thing. I hope we can say "don't replace human jobs," but unfortunately…

Host: Regarding this type of AI agent, I'd like to quote Mira Murati. She said, "The concept of AI agents is not new, but now we are iterating towards the future, developing intelligent agents with intelligence and common sense, understanding why we do certain things."

Li Feifei: Alright, I'd like to add a point and comment on that view. I believe we need to carefully distinguish between replacing "jobs" and replacing "tasks". I believe you and I have seen the same reports. Reports indicate that every human job is actually composed of multiple tasks. I have spent a lot of time researching healthcare. A nurse has to complete hundreds of tasks during an 8-hour shift. AI agents do play an assisting and enhancing role in many tasks, but we need to be cautious with the use of the word "jobs". I do believe that economic business decisions and ethical social decisions are not mutually exclusive. This is a deeper discussion, and I know you and I are in agreement on this.

Returning to your question about these agents having understanding capabilities. I think this is a very subtle term, especially when focusing on the business aspect. What is understanding? There is understanding of patterns in the data, understanding of the decisions you make, and understanding of the intent behind any human task. So, in reality, I wouldn't simply use a broad term like "understanding" to describe today's AI agents.

Host: When do you think AI agents will reach a certain level?

Li Feifei: I believe the best we are currently doing is understanding patterns in the data, especially when we have a large amount of training data, we are doing very well. I'm referring to large language models, using transformer-based algorithms for sequence-to-sequence, we are really good at extracting patterns from the data to create powerful predictive models. So I think this may be the forefront. Again, I emphasize that "understanding" is much more subtle. You all come from the business field and know the subtle differences in it, I think there is still more work to be done. As for intent understanding, I think we have only scratched the surface.

Andrew Ng: I'd like to return to the topic of tasks, because I think it's important. My team works for many companies, and occasionally I receive calls from CEOs. They say, "Hey, Andrew, I read about AI agents, what should I do?" The result shows that there is a way to help companies determine which tasks should be attempted to be enhanced or automated with AI. As Feifei mentioned, most jobs are composed of many different tasks. Take radiologists for example, radiologists read X-ray images, they need to obtain patient histories, operate machines, maintain machines, and mentor young doctors, and so on. So radiology is an example involving many different tasks. The approach I have seen companies take is to observe your team, understand what tasks all your employees are actually doing, not at the job level, but at the task level analysis. To what extent this task is suitable for enhancement or automation with AI and what is the business return on investment. Every time I do this work with companies, we always come up with many ideas, but we don't have time to implement them. So there are many opportunities for enhancement or automation with AI.

The second point I have learned is that the tasks with the highest return on investment are often not what people initially think. Take radiologists for example, people usually think, oh, radiologists just look at X-rays. That's the mental image you have of this profession. But when we break down this job, we find that there are many other tasks, such as collecting patient histories and so on. These tasks may be easier to implement and have a higher return on investment. So I find that systematically doing this exercise often helps companies identify valuable opportunities, and then make decisions on building and buying, and executing AI projects.

Host: This is exactly the next topic I wanted to discuss. Let's talk about some more practical issues. In working with Fortune 500 companies, what commonalities have you found in applications that have clear, demonstrable, and achievable return on investment? For example, what applications do you think most people in this room should focus on?

Andrew Ng: If we broadly observe Fortune 500 companies, I think common applications are in customer operations or customer support. Many companies are trying to enhance or automate customer support. I think software engineering is also transforming. Our support for software engineers goes far beyond Github Copilot. Github Copilot is a nice tool, but we can do much more. I think sales operations have also been greatly impacted. But for specific companies, almost every company is doing customer operations, so you should also seriously consider this aspect. But what's more exciting is that I have talked with a very large agricultural company, and we have identified some tasks. These are not the harvesting tasks you would think of, but some strange tasks. We think maybe AI can save them a lot of time. So, I think niche areas that are suitable for your business and industry characteristics are usually more intriguing. They are not only interesting, but can also create industry-specific defensive flywheel effects. Because everyone will eventually buy some general sales operation tools, but those specific to your business, you should build internally, I find these things very exciting.

Li Feifei: There are indeed various customer support or operation solutions now. But one point I would like to add is that we can also find common opportunities in current technologies. I believe there is a conclusion that still holds true today, that is, where the data is richest, you can identify repeatable or effective patterns, whether it's human language patterns, structured data patterns, or image data patterns. Where the data is, and where the data patterns have been proven to be valuable and actionable in your business, that's where you should focus.

Host: Let's talk about the obstacles that Fortune 500 CEOs may face. We held a healthcare summit last Sunday, and several healthcare CEOs attended. We asked one of them, what are you most interested in digital innovation? He said artificial intelligence. Then I asked him, as a CEO, what are you most worried about? He said, inaccuracy.

Li Feifei: That is one of the most difficult industries.

Host: What would you say to those CEOs who elevate inaccuracy to a CEO-level concern? In your view, what other issues can rise to this level?

Li Feifei: As I mentioned earlier, it depends on your product, your service, and the stakes of the outcome. In many industries such as healthcare and financial forecasting, precision in the long tail is very important. You cannot afford loss of life or bodily harm, you cannot afford banking errors. So you need to understand your industry, understand your solutions and services, and see where AI can truly make a difference. This relates to what you mentioned about hype. When I talk to business executives, we should discard the hype, understand what this technology can do, and avoid investing in directions where AI is not yet proficient.

Host: So Feifei, for industries like healthcare that involve life and death and are highly regulated, they want to use generative AI but are concerned about accuracy issues, what would you say to them?

Li Feifei: Andrew and I have both worked in the healthcare field for a long time. In fact, there are many AI applications in healthcare, starting from very upstream drug discovery, we can do a lot of things.

Host: What I meant was generative AI.

Li Feifei: This is an overloaded term. Today, every kind of AI is called generative AI. In the early research of Andrew and me, we had very specific mathematical definitions to define generative AI. But now, we…

Andrew Ng: We used to call it machine learning, but machine learning…

Li Feifei: We used to differentiate between generative AI and discriminative AI. Nowadays, this mathematical rigor has disappeared.

Andrew Ng: Yes, I think the mainstream media has taken over the technical terms. The tech industry is just adapting to it.

Li Feifei: Indeed. So, when you say generative AI, I assume you are referring to the kind of big data-driven models with a pre-training phase. Some people may include transformers and predictive models in this, but I'm not sure if that's always the case. But anyway, I think if there are indeed accuracy issues, we should examine several aspects, such as whether it's a model issue or a data quality issue. Is it because of introducing AI into the process that this problem arises? There are many business issues that lead to inaccurate results. You need to understand these issues and try to address them. Sometimes, for example, in certain medical diagnostic and treatment scenarios, you do need to recognize the limitations. If the risk is too high, we cannot push forward excessively.

Andrew Ng: Let me add, although we use the term "generative", AI is often used for analysis. My team has done many projects, such as using these large language models to read electronic health records, attempting to draw conclusions rather than writing text. Even if you are writing text, if you use summarization software carefully, you can still get good results. There may still be some errors, but I think in the healthcare environment, there are still many opportunities to deploy these technologies if a certain scenario is not too high-risk.

In terms of diagnosis, if you make a mistake, the result is very bad. But we have deployed a system that is still running in a hospital. This system is used to screen patients, read electronic health records, assess who has a higher risk of death, and recommend that they consider hospice care. But we cannot trust the decisions made by the system. So we hand it over to the doctors to review the cases we present and make the final decisions.

In fact, a friend of mine in Singapore is using a system from the National University of Singapore to view patients' electronic health records and try to estimate how long the patient will stay in the hospital. Sometimes, a doctor may think it's a simple case and the patient will be discharged within three days, but the AI says no, it will take 5 days. This has sparked further discussion. This is what is happening in Singapore. The doctor says, "Oh, maybe I need to take another look at this patient. Maybe I missed something that the AI found." So, these things are actually being deployed. But based on the capabilities of the system, we can usually design safety measures to ensure responsible deployment of AI.

In healthcare operations, if you use AI to schedule the use of MRI machines, even if you make a mistake that leads to a decrease in the efficiency of MRI machine usage, while it's bad, it may not be as bad as making a critical misdiagnosis. So, in reality, there are many deployment opportunities for AI in healthcare, I think almost all industries are the same.

Host: So, next we will discuss base models. This question was inspired by an article I read on a tech website. If 2023 is the highlight year for OpenAI. Then, among the leaders of base models, which company will be the protagonist in 2024? Will Apple launch the Ajax large model? Will we pay more attention to Gemini than GPT?

Andrew Ng: It turns out that whenever there is a wave of technological innovation, the media likes to talk about the technical aspects. So the media will focus on companies like OpenAI, Google, AWS, Microsoft, Meta, Nvidia, AMD, and so on. There's nothing wrong with that. But it turns out that for the success of the technical infrastructure layer, we also need to achieve another area, which is building applications on top of these technology providers. To be honest, we need to generate more revenue from applications built on these tools so that they can pay the bills for these tool builders. A good article shows the capital investment in GPUs. As a group, we need to create some applications to fill the gap of several billion dollars, which is the capital investment we are currently making, especially in GPUs.

So, I don't know what the media will do, they have their own hype cycle. But I think a lot of actual work exists not only in the base models, but in all these fields such as healthcare, financial services, education, e-commerce, to identify and execute projects.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。