AI News

Fei-Fei Li Unveils Unified World Model Atlas at CoRL 2026

36Kr · AIUpdated Nov 12 15:24

AI summary

At CoRL 2026, Fei-Fei Li gave a talk introducing World Labs' new model Atlas, a unified world model for generation and reconstruction. Atlas adopts a multimodal autoregressive diffusion Transformer architecture, predicting the next view conditioned on camera pose, and can switch among three modes: dense multi-image reconstruction, sparse reconstruction from few views, and pure text generation. She disclosed that Atlas ranked first in roughly 500,000 human judgments on the RapidData camera motion benchmark; after distillation, the model can run in real time at 15 frames per second, and its value in robot simulation and zero-shot real-machine deployment is being tested.

Why read this

AI score 75

At CoRL 2026, Fei-Fei Li publicly introduced Atlas's unified generation and reconstruction architecture for the first time, presenting quantitative comparison results for camera-controlled generation and depth reconstruction.

Topics

Tags

Body · AI translation

Robot Frontier reported on November 12 that today, at CoRL 2026, Fei-Fei Li delivered a talk titled "Atlas: A Unified World Model for Generation and Reconstruction," and then held a fireside chat with AMD Chair and CEO Lisa Su. This was the first time the two appeared together in public since AMD announced its acquisition of World Labs.

In September of this year, AMD announced that it had reached a definitive acquisition agreement with World Labs, with the transaction valued at approximately $8.2 billion (about RMB 54.7 billion), expected to be completed before the end of 2026.

In her talk, Fei-Fei Li mentioned the following key points:

1. In robotics, how to understand and represent the environment, predict the possible consequences of actions, and supplement real-world data through simulation are the three core problems that robots need to solve.

2. Atlas has three modes: dense reconstruction from multi-image input, pure generation under zero-pixel conditions, and sparse reconstruction from a small number of views.

3. Technically, Atlas adopts a multimodal autoregressive diffusion Transformer architecture, whose core is to predict the next view through camera-conditioned control.

It can perform dense reconstruction from a large number of images, sparse reconstruction from only a small number of images, and even generate scenes with no image input and only a text prompt.

4. Fei-Fei Li also disclosed that the practical value of Atlas in robot simulation is currently being tested. Atlas can not only generate and reconstruct 3D environments, but can also be used for robot training and navigation, connecting the real world with the simulated world, and then applying simulation results to real robots.

At the same time, Fei-Fei Li also demonstrated related explorations in her talk: robots can be trained entirely in a virtual simulation environment and then deployed to the real world in a zero-shot manner.

5. Finally, Fei-Fei Li also emphasized that Atlas ranks first on the Rapidata camera motion benchmark.

During the roundtable discussion, at the beginning, Lisa Su made a prediction about the future of the robotics field, saying that physical AI, especially robotics, is an opportunity worth hundreds of billions of dollars in the coming years, and revealed that AMD already has a fairly sizable robotics research team, although the related efforts are still at an early stage.

Fei-Fei Li put forward the view that "AI is computing, and computing is AI," and believes that frontier LLMs/VLMs and specialized models each have their own strengths.

Below is the transcript of the talk and roundtable discussion:

Today I want to share with you a model recently developed by the World Labs research team, called Atlas. It is a unified world model for generation and reconstruction (A Unified World Model for Generation and Reconstruction).

Since we are here at the CoRL conference today, let me start with the needs of the robotics field.

I know that many changes are taking place in the robotics field, and many advances have been made. Just a day ago, we saw new research results. No matter what robotics technical route we adopt, whether or not we use policy learning, and whether or not we use agents, there is a fundamental problem that the robotics field is trying to solve:

Given the current state of the world, the task goal, and the robot itself, at every moment, what should the robot do next?

Overall, robots need to solve several basic problems.

First, robots need the ability to understand and represent the state of the world and the surrounding environment. We can discuss whether this representation should be implicit or explicit, but understanding the environment itself is very important.

Second, robots need to be able to predict the possible consequences of actions and, in some way, plan those actions.

Finally, data is always a problem. The robotics field has a very long long-tail distribution, and for robots to be deployed successfully, they also face a large number of long-tail scenarios. Therefore, in many cases, if different environments and states can be generated as much as possible through simulation to supplement real-world data, it would be very valuable. This helps both model training and model evaluation.

No matter what method we adopt, these three problems still exist in the robotics field.

This also makes me think of world models. In fact, the above three problems correspond respectively to the three problems that world models try to solve.

For example, understanding and representing the environment corresponds to the rendering problem in world models and vision models; predicting the possible consequences of actions corresponds to the planning problem; and simulating different states in the world corresponds to the simulation problem.

Of course, I do not want to absolutize these problems, nor do I think they are completely independent of one another. In fact, in many cases, we can unify them, and these problems also intertwine with one another. These are all very exciting frontier directions in this research field.

01. Atlas attempts to unify generation and 3D reconstruction

Today, I want to focus on Atlas. This is a new model we have just proposed.

Overall, Atlas is mainly devoted to solving rendering and simulation problems, especially the simulation of world geometry and 3D and 4D dynamic processes.

In simple terms, Atlas is an autoregressive foundation model trained from scratch. It is a full-modal model that can receive multimodal inputs, including text, images, video, depth signals, grayboxing, and 3D structural signals.

Its output is pixel frames. But Atlas can not only generate high-quality image frames in a spatiotemporally consistent manner, but also synthesize new viewpoints, thereby achieving precise, high-fidelity 3D reconstruction.

Next, I will show some experimental results and specific details. Here I will first introduce Atlas's overall capabilities.

Let us first review the development of related research.

There are two very important problems in the field of computer vision. Of course, these are not the only two problems in computer vision, but they are indeed very important. Researchers have spent many years, even generations, studying these problems.

One of them is the problem of 3D reconstruction. The research I show here includes both the classic paper "Building Rome in a Day" and recent state-of-the-art work, such as NeRF-related papers.

The fundamental problem we are trying to solve is: how can we construct a 3D scene that conforms to the real world from images taken from different camera positions and poses?

Most traditional methods require a large number of images with camera pose information.

If these camera views do not cover all surface areas of the scene, reconstruction becomes very difficult, and the model also struggles to accurately estimate the uncovered regions. Therefore, the final 3D reconstruction results often have considerable uncertainty.

This is the classic 3D reconstruction problem.

But in recent years, generative AI has made great progress. Now, we already have pure generative models that can directly generate high-quality pixel content.

At the same time, diffusion-based backbone architectures have also provided new directions for exploration in multi-view generation research, further narrowing the gap between generation and reconstruction.

Along this line of research, some very excellent works have emerged. They attempt to generate previously unobserved viewpoints and maintain spatial consistency by jointly denoising multiple views.

However, overall, the results of these methods are still limited. They still struggle to maintain spatial consistency, and each denoising process can usually target only one fixed target viewpoint. If one wants to change the viewpoint or achieve real-time generation, the computational cost becomes very high.

Atlas continues to explore along the direction of a series of studies over the past few years, attempting to unify generation and reconstruction.

Our goal is to use the same model architecture and the same training scheme to solve both problems at the same time.

Specifically, Atlas can switch between the two directions of generation and reconstruction.

In one extreme case, if Atlas receives multiple input images and the number of images is sufficiently large, it can perform the reconstruction task. The model can perform dense reconstruction and output scenes and world structures of relatively high quality.

In the other extreme case, if there is no pixel-conditioned input at all and only a text prompt is provided, Atlas can perform a pure generation task. It can also generate multiple consecutive frames with viewpoint consistency.

More interestingly, Atlas can also complete a task that connects generation and reconstruction, which we call sparse reconstruction.

In this case, Atlas receives a very limited number of viewpoints. In the extreme case, the input is even only a single image.

Atlas can estimate camera poses, then further generate multiple viewpoints within the model's capabilities, and ultimately reconstruct and generate a complete world.

Therefore, Atlas can switch from dense reconstruction to sparse reconstruction and then to pure generation within the same model framework.

02. With only a few images, it can also reconstruct a complete world

Before introducing the technical details, I will first show you an example. Some friends present here should be able to recognize this as the Stanford University campus.

Stanford has a very famous Main Quad.

If the model is given only two limited input images, it will generate, or rather complete, a Main Quad.

If you are familiar with Stanford University, you will find that this is not our real Main Quad. It looks like it has the feel of the Main Quad, but in fact it is not the real Main Quad.

The model performs conditioned generation based on features such as the arcades in the input images, so the generated scene retains some characteristics of the arcades and also has arches, but overall it is still a quad generated from the input.

If the model is given eight images, the result will be much better.

Considering that the space itself is very large, the information provided by eight images is still very sparse. But if you look carefully, the model has basically reconstructed the entire quad.

At first glance, it is almost perfect. However, people familiar with Stanford University will notice that the famous Memorial Church is missing. This is because the input images did not include the Memorial Church.

Finally, if 33 images are input, although these images are still very sparse relative to the scale of the entire space and many areas are not covered, one of the images does include the Memorial Church.

Therefore, Atlas has basically been able to reconstruct Stanford University's beautiful Main Quad.

I also want to emphasize one point in particular: all 33 of these images were taken at ground level, that is, they are just ordinary images captured by a person holding a camera.

But the model is able to generate different viewing angles, including this aerial view I am showing now.

So what is the key technology behind Atlas's ability to do this?

03. Viewpoint prediction is Atlas's core technology

For large language models (LLMs), we are already familiar with the objective function of Next Token Prediction.

For Atlas, our core idea is: Viewpoint Is All You Need.

Through precise camera pose conditional control, we let the model predict the next viewpoint.

Specifically, Atlas adopts a multimodal autoregressive diffusion Transformer architecture.

As mentioned earlier, Atlas can receive text input and generate images based on camera pose.

It can also receive an image, estimate the corresponding camera pose, and then, conditioned on this input image, generate more images and more new camera viewpoints.

Here is a very cute little robot in a garden. We can see that the model is able to generate frames from different viewpoints along a camera orbit path.

In addition, as the earlier example of Stanford's Main Quad showed, Atlas can also receive images from multiple viewpoints and interpolate and extrapolate along different camera motion paths, thereby recovering a 3D scene.

It can also receive depth images or gray-box inputs and output RGB pixel images. This capability is very useful in many robotics and gaming application scenarios.

Next, let us look at Atlas's performance using some quantitative results. I want to show its results in generation and reconstruction separately.

First, camera-controlled generation.

These results were actually obtained just a few days ago from an independent third-party organization, RapidData.

Without notifying us in advance, they benchmarked multiple state-of-the-art video models, evaluating camera-controlled generation capability, with approximately 500,000 human judgments in the test.

Ultimately, Atlas ranked first in this test, surpassing multiple current state-of-the-art video models.

As can be seen from the figure, there is a fairly clear gap between Atlas and the second-ranked model.

Atlas can generate coherent videos longer than 60 seconds. In fact, the duration it can generate now already far exceeds 60 seconds.

Here are our latest results from 2026. I especially want to compare them with the early model we tried about a year and a half ago.

As you can see, the range the early model could generate was very limited, and its geometric consistency was also relatively poor. At the time, we even received criticism like this: what you are building is not a world model, but a "Closet Model."

And now, we have gone from a closet to a much larger world.

Here is another very interesting generation case.

Atlas can receive two completely different and unrelated objects, then estimate a reasonable camera pose, further complete other parts of the scene, and ultimately generate a fantasy world that combines these two objects.

This capability is very suitable for various creative applications.

The last example is gray-box input.

Atlas can receive these gray-box structures and generate a complete, exquisite world from them.

This capability is useful not only for robotics, but also for design and creative application scenarios.

04. From depth prediction to 3D scene reconstruction

Next, let us look at the results in reconstruction.

We compared Atlas with some models specifically designed for 3D reconstruction.

First, we tested the depth prediction task on seven different datasets. Here, the lower the metric, the better, because it measures the error rate.

Atlas outperformed some recent state-of-the-art 3D reconstruction and depth prediction models.

In addition, Atlas itself does not need to use Gaussian Splatting as the scene representation.

But if we use this representation in post-processing, we can compare Atlas's Gaussian Splatting reconstruction error with specialized reconstruction models on six different datasets.

Here we used two different error metrics. The results show that Atlas performs very prominently, surpassing most specialized models.

In addition, we gradually adjusted the number of viewpoints provided to Atlas to observe how the model transitions from sparse reconstruction to denser reconstruction.

In this experiment, we likewise compared Atlas with multiple advanced methods and tested on two different datasets.

In terms of error rate, Atlas once again achieved very good performance.

Here are some specific cases.

This is a very famous apartment scene in the field of 3D reconstruction. Given eight input images, Atlas is able to reconstruct this rather large apartment.

Now, we can sample the scene from various different viewpoints, and people can fully control these camera viewpoints to observe the entire scene.

Here is another example showing how content generated by Atlas from a single input image can be further converted into the Gaussian Splatting representation space.

As you can see, from just a single input image, the model generates a very beautiful imagined space.

Similarly, in the next example, we can generate not only the scene content but also depth information, allowing artists to precisely control the camera position and generate videos that match the intended effect.

Those familiar with visual effects may know a very popular effect called Bullet Time.

This effect was first showcased in the film The Matrix. It captures a scene with dynamic changes in space and time using hundreds of cameras simultaneously, then combines these frames to create a very beautiful bullet time effect.

Now, using only three iPhones set up temporarily in the office, we can use Atlas to reconstruct this spatiotemporally dynamic scene and present a beautiful bullet time effect.

Here is another example. The designer combined Atlas with an LLM to display semantic information in the scene.

In the world constructed by Atlas, we can carry out various design tasks and also hope to conduct robot simulation.

Another recent advance with Atlas is that we have distilled the model so that it can run in real time at 15 frames per second.

Finally, I want to talk about Atlas's potential applications in robot simulation.

We are testing the practical value of Atlas in robot simulation, including for robot training and navigation, as well as connecting the real world with the simulated world and then applying simulation results to real robot manipulation tasks.

As you can see, some of these recent research results show that robots can be trained in simulated virtual worlds and then deployed directly to the real world in a zero-shot manner.

We are further deepening this research by incorporating Atlas. We will release Atlas soon, so please stay tuned.

Finally, I would also like to invite everyone to pay attention to the excellent work done by my students.

One of them is the Behavior Challenge organized by the Behavior team at Stanford University. This is already the second year of the challenge, and we will announce the winners tomorrow.

In addition, many of my students have put a great deal of effort into participating in CoRL.

We have six papers accepted, and we welcome everyone to learn about our papers during or after this year's CoRL conference. The World Labs team and the Stanford student team have both put a great deal of effort into this research. Thank you very much!

The following is the transcript of the second roundtable discussion between Fei-Fei Li and Lisa Su.

05. Fei-Fei Li in conversation with Lisa Su, discussing opportunities in robotics development, AI compute demands, and future talent cultivation

1. In the robotics field, the market size may reach hundreds of billions of dollars in the coming years

Peter Stone: First, I'd like to start with Lisa and Fei-Fei. Everyone here today is a researcher in the robotics field.

We have heard that AMD is advancing its acquisition of World Labs, and Fei-Fei just shared her views on world models and future development opportunities.

Let me first ask Lisa. AMD has been investing across every layer of the AI technology stack, from data centers to edge computing and embedded systems. Looking ahead to robotics and physical AI, what do you see as the biggest opportunity for AMD? And how will World Labs fit into AMD's broader strategy?

Lisa Su: Thank you, Peter, and thank you everyone for inviting me here.

I think this is an extremely exciting time in the development of artificial intelligence. Looking back at all the progress we have made so far, it is clear that we are actually still at a very early stage of future possibilities.

The key here is that we must truly understand physical laws, models, and the entire environment. Therefore, we are very excited that Fei-Fei and the World Labs team are joining AMD.

Beyond the fact that Fei-Fei herself is a superstar, we can look at this from another angle.

In our business, if you consider that hardware optimization takes many years, then the deeper your understanding of the workload, the fuller your understanding of the future direction.

I believe the concept of co-optimization among hardware, software, and systems also applies to the robotics field.

Over the next five to ten years, this will become an extremely important field. Therefore, I am really looking forward to World Labs joining AMD.

More importantly, for AMD's ecosystem, I am also very much looking forward to further expanding the AMD ecosystem through this, and getting the entire robotics community here involved.

Peter Stone: Fei-Fei, next I'd like to ask you. From your perspective, how will joining AMD expand World Labs' capabilities?

You have already articulated your vision for world models. So what will the combination of the two allow you to build next? And what will this bring to the robotics field?

Fei-Fei Li: Actually, I have known Lisa for many years. Looking back over the past decade, from AMD to the entire computing field, and now to the era of the AI explosion, everything Lisa has done is very remarkable and amazing.

In the AI field, the relationship between models and computation has fundamentally changed.

This is true for both world models and language models.

Frankly, we are seeing a convergence among many different models, including unimodal and multimodal models. As Lisa just said, this requires co-evolution and co-design between hardware and software.

So this is a very interesting opportunity.

I also want to talk about the PhD students, master's students, and undergraduates here.

We are entering a very exciting era of Team Science.

Combining the respective strengths of the four labs with AMD's strengths, and contributing to the development of computing and AI through Team Science—this is very exciting for all of us.

Peter Stone: You just talked about the inseparable relationship between AI and computing, and also about the changes happening in the robotics field.

For the next question, I'd like to ask both of you to answer together. We see that large language models and other foundation models are playing a role in robotics, and at the same time, video generation models are also beginning to be applied in robotics.

What technology do you think will drive future robots? Do we still need to develop modules at different levels separately? Or can foundation models ultimately solve robotics problems in an end-to-end way?

Can they do everything independently and truly solve the problems in the robotics field?

Fei-Fei Li: Let me answer first. We still need to stay rational and honest.

The research results recently announced by frontier labs, as well as some of the results shown at this conference, are truly amazing.

We are at a very early stage, and at the same time we are seeing a clear turning point. So there are still a large number of questions that need to be explored further.

I think robots driven by vision-language models (VLMs) or large language models (LLMs) will play an increasingly important role.

But I don't think we have fully solved the system-level problems yet.

So in the future, we will continue to see various changes.

Next year's CoRL conference will be very different. Over the past two weeks, we have been reviewing the results of the Behavior Challenge. Seeing the speed of technological progress is truly incredible.

I think this change is far from over.

As a researcher, I personally have always remained highly vigilant. Because I sometimes feel a bit nervous: which research directions should we actually bet on?

But at the same time, this is incredibly exciting. Because now we have seen new signals from multiple different directions, and we also have the opportunity to combine these potential approaches.

Lisa Su: I'd like to add something. In many vertical application areas, what we see is that there is no single model or single development direction that can solve all problems.

Facing various real-world application scenarios, we actually need to cover a very wide range of use cases and use different types of models.

Fei-Fei, I agree with what you just said. In some scenarios, what we call frontier large language models will perform very well. But I can also imagine that in other scenarios, more specialized models will have more advantages.

3. Robots will drive the growth of computing demand in various forms at scale

Peter Stone: Lisa, robotics has become one of the important factors driving the growth in demand for chips, computing resources, and tokens.

Many students and academic researchers are here today, and they all very much hope to obtain more computing resources. As Fei-Fei just said, these resources are very important.

So how does AMD view the future demand for chips and computing resources in the robotics field?

Training models requires enormous computing resources. Of course, once robots are actually deployed in real-world environments, they also need to complete perception, reasoning, and action in real time, and are often subject to strict power constraints.

Does this mean that robotics is essentially a systems engineering problem?

For these workloads, how does AMD consider designing computing, memory, software, and other parts of the system together?

Lisa Su: What I most want to say is, is there anyone here who needs computing resources?

In the era we are now in, computing is foundational, and in many ways, computing is indeed directly related to intelligence.

If we look back at the stage of development we are currently in, I can say that we need various different types of computing capabilities. This is also why I am very excited about the entire computing ecosystem.

I have been working in the computing field for 30 years. Compared with everyone here, this makes me seem a bit older.

But the fact is, there has never been a better time than now, and there has never been a more important time than now.

We need computing capabilities at every level. We need cloud computing resources to train these enormous models.

We also need enterprise on-premises computing resources, and we need computing capabilities in edge environments. As robots become more and more important, these resources will also become very critical.

In addition, we also need computing capabilities to enter the real physical environment.

My view is that we need computing capabilities of different scales and different forms, but they all need to follow some basic principles.

If models have different scales, then our computing capabilities also need to have different scales.

Fundamentally, this is a solutions problem.

We need to co-optimize hardware, software, and systems together, so that they can solve specific problems as efficiently as possible.

Fei-Fei Li: I'd like to add one more point. We're discussing this today at a robotics conference, and I'm not saying this just because I'm at CoRL—I genuinely believe that robotics will drive computing into its next stage of development.

I really like this assessment; it's a good thing.

If you look back at the history of robotics, you'll see that in the past its development lacked the driving force that large language models have provided, because robotics itself is a very difficult problem, and it also needs to interact with the physical world.

But the turning point we're seeing now is real.

I can't accurately predict the slope of future development, but what's certain is that we've already entered an upward trend.

Robotics will drive the growth of computing demand across various forms at scale, not only on the training side but also at the edge.

Peter Stone: Fei-Fei, you've experienced this issue from many different angles.

You were once a professor and have long advocated for resources for academia; you founded a venture-backed startup; and soon you will also serve as Chief Scientist of AMD.

We just talked about one challenge facing frontier research: it's hard for researchers to get the computing resources they need.

What do you think needs to be done to make more advanced robotics R&D more accessible to researchers, startups, and students?

And what role can companies like AMD play?

Fei-Fei Li: I'll leave that question to Lisa to answer.

But Peter, you and I have long been colleagues and collaborators. We've been pushing to get computing resources to the public sector.

As early as 2019, some of us pushed for a bill, the National AI Research Resource Bill (NAIRR).

In other words, we must provide resources to the public sector.

The public sector includes university researchers and also nonprofit organizations, because this is the first stage of innovation.

This country's innovation engine begins with basic research in the public sector, with open-ended, curiosity-driven research.

At the same time, we are also training the next generation of inventors and the future workforce.

The public sector must play a key role in this.

So the public sector plays a crucial role.

Peter Stone: I completely agree with you, so what role can AMD play? How can you help make this happen?

Lisa Su: I think a good approach is that everyone should go ask AMD's Chief Scientist exactly what should be done.

First, I one hundred percent agree with everything Fei-Fei just said.

For us, the opportunity is that we are building computing capacity on a very large scale, reaching the multi-gigawatt level.

Frankly, we are indeed at a stage where demand far exceeds supply.

But dedicating a portion of those resources to long-term research and providing them to universities, to ensure the next generation of innovators can access computing resources—that is absolutely a responsibility we should bear as an industry leader.

We very much want to spend time engaging with this community and make sure this actually happens.

I wasn't entirely joking just now.

There's a reason I'm so looking forward to Fei-Fei joining AMD as Chief Scientist.

But all of this is still at a very early stage.

Now, what we need to think about is: how do we ensure these technologies truly make the world better?

Robotics is one important area among them.

I think we are already very close to the goal, but we haven't yet truly demonstrated the practical value that robots can bring.

I hope AMD can be part of this. Of course, I believe other companies in the industry are equally motivated to help us turn this vision into reality.

4. Will AMD establish a robotics research team?

Peter Stone: Next, I'd like to turn to questions from Slido.

There are several questions with very high scores here, and their rankings keep changing. I'll pick the highest-ranked one for now.

Everyone is very interested in this question: Will AMD establish a robotics team?

Lisa Su: You're assuming we don't already have a robotics team.

If you look at it from the perspective of the computing stack, you'll see that many of the components robots need, we actually already have.

For example, we have embedded FPGAs, processors, and AI engines. These components can already control many robots, especially those you see in industrial environments.

But to go further into the robotics field that everyone is researching, I think we are still at a very early stage.

I very much hope that after Fei-Fei joins AMD, she can devote a great deal of energy to robotics research while strengthening our connections with the entire ecosystem.

Because I have always firmly believed that no single company can do all the work across the entire technology stack alone.

Ultimately, this is also why startups can play such an important role today.

We have industrial labs, universities, and large enterprises.

What we really need is to combine the strengths of these different domains.

In such a development environment, building an open ecosystem, providing powerful hardware, enabling software at different levels to work together, and letting different labs each conduct their own research—these are all important ways I believe can accelerate the speed and progress of robotics development.

We also need to truly bring robots into manufacturing and industrial applications.

5. No matter how advanced AI becomes, human intelligence remains very valuable.

Peter Stone: I think many people here are at the beginning of their careers, and they are seriously thinking about which are the most important and most interesting problems they should research as PhD students.

Conversely, are there some research directions that you would not recommend people choose, that are not suitable as PhD dissertation topics?

I would like to hear views on both of these aspects at the same time.

Which areas are very worthwhile as topics for PhD research or undergraduate research? Which directions should be chosen cautiously?

In addition, at this conference, many people have also asked me a question that I think everyone may also be interested in: in the coming years, what exactly will a PhD mean? Is pursuing a PhD still the right path into the research field?

I think, Fei-Fei, the answer you give may not be what everyone wants to hear, but I still want to hand this question to you.

Fei-Fei Li: First, I think this is an era of tremendous change.

And we entered this era only after we had become adults.

As for whether one should pursue a PhD, and what one should do during the PhD, my own answer is also changing. I want to tell everyone my views frankly.

I do not think a PhD is suitable for everyone.

Now, people can become researchers, become innovators, and produce real impact through more different paths.

This does not mean that a PhD no longer has value.

I think it is still very good. I myself loved the experience of doing a PhD, and I believe Peter and Lisa loved it too.

But more different paths have now emerged, and you need to actively find the direction that suits you.

So, what exactly does a PhD mean?

I still firmly believe that the core of a PhD lies in curiosity.

This is a fully protected period during the golden years of your life. During this time, you can pursue the most challenging, most curiosity-inspiring, most interesting problems that are expected to have important impact.

It may even be becoming scarcer, but it is still precious.

Now, we have a massive number of tokens competing with human brain activity.

But I think the civilization we are building still needs human intelligence, and it also needs bold human curiosity.

Therefore, for those who have a strong inner desire and curiosity, I do think that a PhD is still a very good path.

But you must have this belief, and pursue those important questions fearlessly.

Lisa Su: I would like to talk about it from my own experience.

When I was doing my PhD back then, I actually never really thought that I had to obtain a PhD.

My parents said to me at the time: "Come on, you are still too young. What are you going to do next? Just stay in school."

As it turned out, those were truly the best four years of my life.

The reason is exactly what Fei-Fei just said.

What matters is not the specific problem you research, because specific problems will change.

The specific problems we study today change very quickly.

But what is truly important about doing a PhD is that you have the opportunity to become one of the best people in your field, or one of the best researchers in this field, because you have the opportunity to explore.

So, for this reason, I think doing a PhD is a very wonderful thing.

If you have the opportunity, I would encourage you to try it.

Because life is long, and no one can take your PhD away from you.

Moreover, you will always have that experience and knowledge: you once chose a problem and studied it to the frontier of your field.

As for what Peter just asked about what everyone should research, I also have a very strong view.

Right now, the entire industry is discussing artificial general intelligence (AGI), and how far away we actually are from AGI.

Sometimes, I hear some peers and industry people say that we are already very close to AGI, as if it could be achieved very soon.

But I actually do not agree with this statement.

I do not think we are that close to AGI.

Looking back at the tremendous progress artificial intelligence has made, we have indeed achieved many incredible results.

But ultimately, you still need to use your own expertise and intuition.

I often say that when recruiting talent, you can tell which people truly have intuition for their field and which people merely have so-called book knowledge.

You need to find the problems you are truly interested in and cultivate your intuition in this field.

In this way, no matter how the tools change, it will not affect your continued advancement of your field.

All of us lack patience. We want products to launch as soon as possible, and we want the next model to be released as soon as possible.

And this is exactly the difference between the university environment and industry.

Peter Stone: Fei-Fei, you just mentioned that human intelligence is still very special and also very important.

One research direction I often give is: how to use fewer resources and less energy while maintaining the same performance.

I think this is a very important research direction.

Of course, this may conflict to some extent with the business goal of selling more chips.

I would like to hear how the two of you view this issue. Do you think that reducing energy consumption and improving efficiency should become an important research direction for contemporary students?

Lisa Su: Of course. This is absolutely a very important direction.

This brings us back to the question we were just discussing: there is no one-size-fits-all approach to computing.

The physical world does not have just one fixed scale either; rather, many different scales exist.

I think this is also why we are still at a very early stage.

You are right that we now have large data centers and various scaling laws.

But ultimately, we need to bring computing power everywhere.

Therefore, lower power consumption, higher efficiency, longer service life, and lower latency—all of these metrics need to be optimized.

In addition, I think we need to be careful about one thing: we should not simply conflate the 20 watts consumed by the human brain with the enormous energy required to train AI models.

Fei-Fei Li: Evolution itself is also very expensive.

Yes, the human brain is indeed very efficient, but it is the result of billions of years of evolution. I do not yet know how to calculate the cumulative energy consumed by the entire animal kingdom over the long course of evolution.

Therefore, we should not conflate these two factors.

Furthermore, I have always believed that brain science itself is an independent science. Whether it is understanding the brain, recording and replicating the brain, or helping to solve brain diseases, these are all questions that brain science needs to study.

Because the computing problems we are trying to solve have gone beyond certain limits of human capability—for example, writing code, generating simulated environments, and providing computing power for robots.

Peter Stone: There is another question on Slido. Given that you are currently in the honeymoon period after announcing your collaboration, someone asks: Do you have any disagreements? Is there anything where you disagree with the other's view?

Fei-Fei Li: Seriously, do we have any disagreements?

I don't know. As you can see, we haven't found one yet. We haven't had such a situation yet.

If you invite us back again in the future, you can ask again.

6. In a changing era, seek unchanging values

Peter Stone: From the questions everyone just raised, it is clear that students today feel both anxious and excited about AI.

If we combine your own experiences, suppose you were now in the position of the students here, how would you view this era?

Finally, what advice would you like to give to everyone here? For many students, this is indeed a turbulent period, because the whole era is changing.

Fei-Fei Li: Your own lives are undergoing change, because you are students, and your field is also changing.

From the perspective of human civilization as a whole, we are also at a very special moment, because we are experiencing a technology of civilizational significance.

I actually want to invite all students, in an era of constant change, to seek out those eternal, unchanging values.

As Lisa just said, life is long. Many things will change. Research questions will change, challenges will change, and opportunities will change.

But how to live the life you truly want depends largely on those things that do not change with change—that is, those eternal values.

For me personally, this experience has taught me that eternal values include curiosity, integrity, and the courage to pursue truth and explore truth without fear, while also daring to explore boldly in thought and remaining honest.

Lisa Su: I think I would encourage everyone to maintain lofty ambitions and to be truly optimistic about the future.

Because you are in the best of times. You have excellent tools, and there are many opportunities before you.

Many people seek advice from me about career development, such as what they should do and what direction they should choose.

I usually tell them a piece of advice I received when I was young.

At the time, I was a young engineer at IBM, having just worked for about two years.

One day, someone asked me: "Lisa, what do you want to be when you grow up?"

At the time, IBM had a position called IBM Fellow, which was focused on deep research; there was also the position of IBM vice president.

What I actually thought at the time was that I might not have enough intelligence to become an IBM Fellow, but perhaps I had a chance to become an IBM vice president.

So I decided to work toward that direction.

At the time, someone gave me a piece of advice: always move toward the hardest problems.

At that time, I did not really understand the meaning of that sentence. But this advice has stayed with me for many years. Now, I also tell this to others.

Because ultimately, as engineers, scientists, and researchers, our job is to solve problems.

We need to solve very interesting problems, very important problems, and exciting problems.

We can gain joy from solving these problems, achieve success through them, and drive the world forward through this work.

Therefore, if you are about to devote yourself to a job, go solve the hardest problems, and believe that you can have a huge impact.

I think this is also why this era is so interesting.

You now have tools that can help solve the most interesting problems, some of which have not been solved for decades.

What an interesting thing this is!

So, I want to say that now is a remarkable time for students starting their careers. I also hope everyone will consider joining AMD in the future.

Fei-Fei Li: By the way, she will be our hiring manager.

This article comes from the WeChat public account "Robot Frontier"; author: Liu Lishan; editor: Xu Lisi; published with authorization from 36Kr.

Source

36Kr · AI