Article · AI translation
RSI Cannot Be Infinitely Fast
Tingyu reporting from Aofei Temple
QbitAI | WeChat public account QbitAI
Tingyu reporting from Aofei Temple
QbitAI | WeChat public account QbitAI
RSI is so popular, but can it really make the pace of AI progress infinitely fast?
Jeff Dean's judgment: that's not very realistic.
He gave an example: someone starts an experiment that finishes at 2 a.m., but has to wait until returning to their desk at 8 a.m. to check the results.
An automated system can eliminate that wait and promptly decide what to do in the next round. But training, computation, saving data, and evaluating results still take time.
Building and improving a foundation model still requires extensive training and experimentation. Those processes cannot happen at infinite speed.
Recently, in a conversation with Fei-Fei Li at Stanford HAI (the Institute for Human-Centered Artificial Intelligence), Jeff Dean discussed the limits of RSI and also revealed what his new company is doing.
The presentation still had the very minimalist slide style typical of Jeff Dean at Google...
According to him, the new company, Discovery Loop, already has about 20 automated improvement loops running and has built a collaborative framework that supports people and agents working together.
They want a small number of experts to conduct research with a large number of agents, aiming to make domain experts 20 to 50 times more productive.
To build this team, they conducted 400 interviews in cafés and other places during the first 3 weeks.
Discussing the technology behind these changes, Jeff Dean also admitted that he had misjudged a few things...
One of them was coding.
Coding was already one of the team's areas of focus on the Gemini project. But he did not fully foresee how profoundly this capability would affect a model's ability to solve all kinds of problems.
He had also expected neural networks to develop more unusual, irregular structures, but that path of evolution did not unfold entirely as he had imagined.
On AI doing mathematics, he recognizes the value of machine-generated proofs, but places greater importance on whether people can understand the results and extract methods they can continue to use.
What follows is an edited Chinese account of the talk and conversation.
Enjoy~
Jeff Dean's keynote: AI runs its own experiments, with the goal of making experts 20 to 50 times more productive
Automating the experimental loop
It's great to be with you today. I'd like to introduce Discovery Loop.
Our mission is to build AI systems that automatically solve important problems in machine learning, science, and engineering.
We believe this will ultimately lead to more discoveries, and that's what excites us.
Many disciplines share a common process that could be called the general loop of science and engineering: propose an experiment, carry it out and run it, observe the results, examine the measurements, and then decide what to do next. Often, the next step is another experiment.
Our idea is to automate these experimental loops. We believe that if we can shorten the time for each iteration, automatically propose and run many experiments, learn from each evaluation so later experiments draw on earlier results, and increase both the quality and quantity of experiments, we will achieve breakthroughs.
That's what we want to do.
Look at this list of Grand Challenges from the U.S. National Academy of Engineering. These are important problems to solve in the 21st century, and about a quarter of the century has already passed.
Think about the subproblems that must be tackled to solve them, especially in the areas marked in red in the image. We believe automated discovery loops can help.
Why we formed a public benefit corporation
One point matters greatly to us and is part of why we formed an independent company: if applied well, technology can have a very positive impact on the world, but it takes a collective effort to ensure that happens.
So we established the company as a public benefit corporation.
Our public mission is to develop technologies that address important societal problems, make their benefits available to more people, and improve people's lives.
The backgrounds of the 4 founders
Let me briefly introduce the 4 of us. My 3 cofounders are Sanjay Ghemawat, Quoc Le, and Oriol Vinyals.
Looking at each pair, we've worked together for 14 to 30 years. We enjoy tackling hard problems together and have gained experience in different areas.
We all come from Google, so all the work shown here is related to Google. We've worked on multiple generations of Search products, early advertising systems, contextual advertising, and products such as Gmail, News, and Translate.
The second column is infrastructure. The systems in the upper half help people organize and manage vast amounts of data, run large-scale computations, and extract valuable information from them.
The 4 items below are successive generations of large-scale machine learning infrastructure used to train increasingly capable models. We also worked on the TPU, specialized hardware for machine learning.
In AI research, we've contributed to important technologies behind many of today's highly capable models, including model distillation, mixture-of-experts models, or MoE, and neural architecture search.
Neural architecture search can be seen as an early form of an automated experimental loop: finding a neural network architecture suited to solving a particular kind of problem.
We've also worked on applications of AI in different fields, including scientific research. Quoc and I worked on AlphaChip, which uses reinforcement learning for chip design, including placement and routing.
Oriol worked on AlphaFold and neural network weather forecasting. Our experience complements each other's well.
We've also managed very large teams. The Google Brain team I managed had as many as 600 people, while the Google Research and AI teams had about 4,400 at most.
Later, I realized I was spending too much time in meetings discussing second-quarter operating expenses. I hit send on an email and went from managing 4,400 people to having just 4 direct reports.
The following week, Geoff Hinton retired, leaving me with just 3 direct reports.
Later still, Oriol and I served together as technical leads on the Gemini project, while Quoc led post-training.
We're also very proud of people who worked on our teams and later achieved great things, including founders of companies such as Anthropic, OpenAI, and Physical Intelligence. Physical Intelligence should be fairly well known at Stanford.
Our experience combines AI research and distributed systems.
This image lists some highly cited researchers in AI and distributed systems based on Google Scholar citation counts. We're in the red boxes, and the people marked with asterisks have also worked on our teams.
Building the team and hiring
We also have a lot of fun together and have traveled to many places. The photos show London, Patagonia, Vietnam, and California.
Once, after dinner, Quoc and I took a motorcycle service similar to Uber back to where we were staying.
In the first phase after the company was founded, we put a lot of effort into assembling the founding team. People came from different places, as you can see from the logos in the image. Many have experience founding companies or leading important projects.
The team includes people with experience leading the Go language project, as well as major contributors to projects such as Gemini, DeepSeek, and Kimi K3. A coauthor of the Kimi K3 paper who is also an author of the Attention Residuals work recently joined us.
Our team includes both experienced people and rising stars early in their careers. About 10% of the company have received the ACM Prize in Computing, and 10% have won medals at the IOI, the International Olympiad in Informatics.
We're excited to work with these people, and we place great value on everyone's ability to work together.
Why we can do this now
Why now? Because frontier models have become very capable. This is the result of innovation at multiple levels of AI over the past 10 to 15 years.
Hardware performance has improved greatly, including TPUs, GPUs, and other more specialized hardware. Software abstractions have improved too, making it easier for people to implement what they want to do and map computation more efficiently onto hardware.
Model architectures have also evolved, including word2vec, sequence-to-sequence models, Transformers, and MoE. Training algorithms have made many advances, and techniques such as distillation allow smaller models to achieve very good results.
Inference algorithms have also advanced considerably. We can spend more computation at inference time to get better results, for example by using chain-of-thought reasoning or trying multiple ways to solve a problem and then checking the results.
Models need tools, creativity, and accumulated experience
A powerful foundation model alone is not enough to solve the problems we want to tackle. We need to start with a strong foundation model and combine it with agents so the system can use tools effectively and complete complex tasks.
We particularly want agents and models to be creative. Breakthroughs in science or engineering often require combining different insights, proposing something new, and then testing how well it works.
These systems should also learn from past experiments and experience. That includes learning while solving a current problem as well as accumulating knowledge over the long term.
Suppose a system has already solved 5,000 problems. When it faces problem 5,001, it can draw on the experiences most relevant to the new problem.
The challenges of large projects also lie in dividing and coordinating the work
Every stage of the traditional experimental loop requires substantial human effort.
With larger problems, things get more complicated because you usually have to divide them into many subproblems or modules, and that division is itself a difficult problem.
Within each subfield, many people run experiments and judge which approaches work and which do not.
This creates coordination overhead, resource conflicts, and issues at the level of the larger system. Breaking a problem apart is hard, and integrating the results of the various subproblems is hard too.
What's more troublesome is that once a large project is underway, it's hard to change how it was originally divided up. Whether you're building an airplane or training a frontier model, it's difficult to flexibly revise early decisions.
We think a small number of domain experts can work with automated systems. Ideally, the system can automatically identify subproblems worth investigating and solve them.
Of course, people may also need to guide how the problem is divided and decide which subproblems matter most.
That would allow a small number of people with domain expertise to tackle large, consequential problems. We believe the right infrastructure can support this way of working.
You need to run tens of thousands of automated experiments, and you need the system to understand the value of different experiments and decide what to do next.
Reinforcement learning or evolutionary algorithms can be used for automated decision-making, making it easier to reassess earlier decisions. Every early experiment should inform later ones.
About 20 automated improvement loops are already running
What have we done so far?
During the first 3 weeks, we conducted 400 interviews, mostly at Cafe Venetia on University Avenue, looking for the kinds of people I mentioned earlier.
We also looked for office space and have now moved into the former Zibibbo restaurant, which has quite a distinctive space.
We built our own agent operating framework, or agent harness, for internal development and automation.
It has some useful collaboration features: you can talk with a model and ask it to solve a problem, and you can invite a colleague to join so that 2 people can interact with the model and agents together.
Other people can comment on work in progress and offer design input before a proposal is handed over to an agent.
From the outset, we've built our infrastructure for agents, with the aim of scaling to tens of thousands or even hundreds of thousands of agents carrying out tasks, rather than merely supporting hundreds of people at work. We think this is important.
About 20 automated improvement loops are running now, and we can see them operating and improving.
We're also evaluating how to make these experimental iterations more effective and faster, and how to use results and measurements from earlier experiments to help the system learn faster.
The goal is to make domain experts 20 to 50 times more productive
That's where we are now: building software loops and automating machine learning tasks, because we believe this can advance more autonomous scientific research.
Initially, we focused on building automated systems that could solve problems in machine learning research and machine learning engineering.
Our goal is to make domain experts 20 to 50 times more productive.
We want these domain experts, and people who genuinely want to solve a particular kind of problem, to be able to solve it with less effort. People still have a role in the process.
The machine learning breakthroughs of the past 15 years have produced powerful models and created the conditions for automating machine learning, science, and engineering, thereby accelerating discovery. Thank you, everyone.
Oh, and we're hiring. You can apply at discoveryloop.com.
Jeff Dean in conversation with Fei-Fei Li: RSI has physical limits too, and their distillation paper was once dismissed as lacking impact
Why leave Google to start a company?
Fei-Fei Li: Thank you, Jeff. That was a great talk. I think this is the last session, and we happen to be standing between the audience and refreshments.
Jeff Dean: That's never a good position to be in.
Fei-Fei Li: I had hoped this discussion could be more technical. But since it's a conversation, I'll be asking questions and taking part in the discussion.
Thank you for your talk. My first question was going to be: Why did you start Discovery Loop? Congratulations.
Jeff Dean: Thank you.
Fei-Fei Li: The 4 of you have all worked at Google for a long time, especially you, since you joined Google in 1999.
We've all seen the technological progress of the past few years, but I'm still curious: Was there a particular moment, perhaps in the past year or 2, that made you, Quoc, Sanjay, and Oriol feel you had to do something different, and that you needed to leave Google to do it?
Jeff Dean: We've long been excited to see how automated systems can increasingly help people develop software.
Quoc did a lot of work on neural architecture search 10 years ago. You could think of it as a kind of early recursive self-improvement system: having it search for model architectures suited to computer vision or language problems.
We knew these methods worked. At the time, though, you had to specify the search space with a lot of rather messy Python code.
Now you can use natural language to ask a capable model and agent system to, for example, explore model architectures that would help interpret MRI images, and then let it get to work.
This has changed what we imagine computers and AI systems can do. If we can make these systems run reliably, there is a great deal we can do. We think science and engineering are particularly well suited to producing a positive impact this way.
Fei-Fei Li: So it was these loops and self-improvement capabilities that gave you the idea to start a company.
Jeff Dean: Some might say we're a little “loopy.”
[Note: “Loopy” is a pun on “loop,” jokingly suggesting “a little crazy” or “a little eccentric.”]
Jeff Dean: There's another important point. We've been through the process of starting with a very small project many times, and we really enjoy that feeling: at first there are 5 people, then a 6th, a 7th, and a 10th join. It's really fun.
You can have that experience in a large company too. But we wanted to make this a public benefit company because that reflects our values.
Starting a company also means bigger challenges. You have to deal with office space and the other things involved in setting up a company, and that's fun too.
Fei-Fei Li: I had a similar technological moment when I started World Labs.
We're focused on a different area from you. What prompted us was the convergence of computer vision and generative models, which showed us the possibility of creating generative models that go beyond text.
Jeff Dean: The generative models you've built, and what they can do now, are truly impressive.
How to distinguish self-improvement from recursive self-improvement
Fei-Fei Li: Let's talk about self-improvement and recursive self-improvement in a little more depth. We studied machine learning at around the same time.
I was just thinking about how many names this concept has had: incremental learning, continual learning, meta-learning, lifelong learning, neural architecture search, and now self-improvement and RSI.
Could you give the audience, especially those with a technical background, a more rigorous definition? What is self-improvement, and what is recursive self-improvement? I want to distinguish imagination from fact.
Jeff Dean: The loop I described earlier can be applied to many problems. My understanding of a self-improving system is that you give it a goal, usually a measurable metric such as the accuracy of a machine learning model. The system tries different model architectures to improve that metric.
Fei-Fei Li: Let me stop you there. A lot of machine learning algorithms can do that.
Jeff Dean: Right.
Fei-Fei Li: That alone isn't recursive, right?
Jeff Dean: For example, we give a system a set of MRI data and corresponding outcomes and ask it to develop a model to help interpret MRI images. The system can develop models automatically, try different approaches, and hopefully improve accuracy as the experiments progress.
But that doesn't improve the underlying system responsible for proposing the next experiment. Recursive self-improvement is about applying the same self-improvement method to the underlying system that proposes improvements.
Why can't the self-improvement loop run infinitely fast?
Fei-Fei Li: I understand. But this concept is easy for many people to confuse. Especially in Silicon Valley, people use the term loosely, and as a scientist that makes me somewhat uncomfortable. I'd like us to define RSI clearly here.
Let's also talk about the physical limits of RSI. It's now described as though it were an endless system that could keep accelerating to infinity. What's hype, and what's speculation?
Jeff Dean: Think about how humans have improved model capabilities over the past 10 to 15 years: researchers and machine learning engineers run many small-scale experiments together, work out which ideas succeed and which fail, and then apply the successful approaches to the next generation of models.
That's how Gemini 1 was followed by Gemini 2.
That's the normal research and engineering process. Automation really can speed up some parts of it. For example, someone starts an experiment, it finishes at 2 a.m., and they wait until they return to their desk at 8 a.m. to look at the results.
An automated system can inspect the results promptly and decide what to do next. It might also make better decisions about the next step across an entire large project than individuals deciding separately.
But building and improving a foundation model still requires extensive training and experimentation, and those processes cannot happen infinitely fast.
There are real physical processes involved: computers have to perform computations, results have to be checked, data has to be saved, and another model has to analyze the results before deciding what comes next.
Some people think this loop can run infinitely fast. I don't think that's very realistic. We need to stay grounded in reality. Automated systems can certainly increase the speed at which these systems improve, but they won't make that speed infinite.
Are humans also recursively self-improving?
Fei-Fei Li: I think that's an important point. I haven't actually applied for a job at Discovery Loop, but I have had coffee with Jeff at Cafe Venetia.
Jeff Dean: We have coffee for many reasons, because we like coffee.
Fei-Fei Li: We've been talking about the hype and the facts around RSI all along. It seems we were already preparing for this conversation then. I thought of an analogy you haven't heard yet. You can agree or disagree.
Are humans self-improving systems or recursively self-improving systems? I think humans definitely improve themselves.
I'm a terrible cook. The first time I make an omelet, I might burn everything. The second time, I'll learn from the experience and do a little better. Maybe after 10 tries, I can make an omelet I can just about eat, though my children still refuse to.
That's self-improvement.
But recursive self-improvement involves the brain. As I learn to make omelets, something must change in my brain. Neural connections need to adjust so that I have a better system for doing it.
If we're talking about hardware improvements, maybe I could even grow a 6th finger.
Jeff Dean: Maybe you were generated by AI.
Fei-Fei Li: Exactly. I keep making omelets, and my brain improves further. If a loop like that happens, humans are RSI systems. Does that analogy work?
Jeff Dean: I think both happen. People learn many things over a lifetime, such as making omelets and solving calculus problems. That's self-improvement. People also learn how to learn.
Fei-Fei Li: And learning how to learn enables people to improve themselves recursively. Once you understand how you take in knowledge and information most effectively, you can learn more things faster.
Jeff Dean: Right.
Fei-Fei Li: Recursive self-improvement might also be a kind of evolutionary algorithm, right? Let's not go further into that now. I think it's valuable to be able to discuss this publicly, because there's so much hype around it at the moment.
Jeff Dean: There are many claims that it will become infinitely fast and therefore frightening. If you understand it as an engineering system in operation, it's less frightening and easier to understand.
Starting with research in digital environments; healthcare is a future direction
Fei-Fei Li: You've said a lot about Discovery Loop, but you're still somewhat vague about exactly which scientific problems it will tackle.
I'd like to press you on this, to the extent you're willing to share: Of the many scientific problems that could benefit society, which one do you most want to study right now?
Jeff Dean: I've long been interested in applications of AI in healthcare. We'll work on that in the future, but not now. It's an important field and fits well with our mission as a public benefit company.
For now, we're focused on scientific and engineering problems where the entire experimental loop can be completed in a digital environment. These problems are better suited to speeding up the whole loop.
If physical experiments or other real-world processes are involved, speeding things up is harder. Of course, many important disciplines require physical processes, and we also have some hybrid approaches that can help accelerate research in those fields.
Initially, we're focusing on 3 areas: machine learning research, machine learning engineering, and hardware design. It makes sense to focus on a few areas first. We also have experience in them, and this work can help enable other research.
For example, if a system can automatically solve machine learning engineering tasks, it can be used in healthcare, materials science, and other fields to develop models suited to particular tasks. That's self-improvement, not recursive self-improvement.
We're focused on these 3 areas now and hope to expand into more areas later.
Can automated experimental systems conduct theoretical research?
Fei-Fei Li: I studied physics as an undergraduate. In physics, there's a distinction between experimental and theoretical research. The loop you drew begins with “propose an experiment.” In physics, we would usually say “form a hypothesis,” then conduct an experiment and evaluate the results.
Whether or not they're recursive, these powerful AI agents and self-improvement loops place a lot of emphasis on forming hypotheses and running experiments in parallel. Does that point more toward experimental research? The process of theoretical discovery seems harder to describe.
Theorists like Newton and Einstein didn't personally conduct that many experiments. They read experimental results, but they worked differently. Is your approach more inclined toward experimental research?
Jeff Dean: Not necessarily, though it might lean slightly in that direction. But I do think this kind of system can also do excellent theoretical research.
You can ask it whether it can come up with a theory to explain the data. The system can try repeatedly to find a theory that explains the experimental data while being as simple as possible.
Many physicists also follow developments in experimental fields and try to explain phenomena that people have observed but do not yet understand. I think that's possible, and we shouldn't completely separate these disciplines.
Imagine a system that has run 100 million experiments across 100 scientific fields and solved 1 million problems. It has seen all kinds of experimental results and knows which approaches have worked and which haven't in fields relevant to future problems. It could well gain a lot of insight from that.
That experience may not have the strict disciplinary boundaries of human academic departments. One thing I like about HAI is that it can cross organizational boundaries and bring people from different fields together to talk.
The value of AI mathematical proofs and the role of elegant solutions
Fei-Fei Li: Then we have to talk about mathematics. While preparing for this conversation, I also invited Terence Tao and called him, but he couldn't make it.
I have mixed feelings about these recent developments. First there was work related to Navier–Stokes, then more than 720 mathematical manuscripts, and yesterday OpenAI released some results on AI solving math problems.
Of course I welcome these discoveries. Mathematics, and problems that have existed for hundreds or even thousands of years, are important expressions of human intelligence. Mathematicians see both solutions and beauty.
Now we have an entirely different mechanism for tackling these problems.
I've talked with Terence Tao quite a bit. He has also spoken recently about “Mathematics 2.0,” discussing which problems AI can help solve.
As he puts it, some problems remain beyond AI's capabilities, especially those that have no known solution and lack data.
What do you make of these changes in mathematics over just the past few weeks?
Jeff Dean: These models, and the agent systems that allow them to work in a highly parallel way, have already shown an ability to solve very difficult open problems in mathematics. That shows us what is possible now.
I think people are sometimes dissatisfied, or even uneasy, because some proofs consist of things like 500 pages of Lean code.
A machine can check them, but people don't yet quite understand how to arrive at the result in an elegant, concise way.
Knowing that a proposition has been proved or disproved is certainly valuable. But when human mathematicians study a problem, they also develop elegant methods. Those methods often become new mathematical tools that can be used on other problems.
Fei-Fei Li: Do you mean that new mathematics can only come from elegant solutions?
Jeff Dean: No. I mean that an elegant solution often yields a good method that can be applied to other problems.
In some very long machine-generated proofs, people may not yet see such a clear, elegant method.
But simply knowing that something can be done is important too. If people know a problem is solvable, they may try harder and perhaps find another, more elegant solution.
Having so many problems solved at once may make people uneasy, but it has now become possible. We need to find ways to use these capabilities to do more mathematical research.
A Pop Quiz About the Transformer Paper
Fei-Fei Li: There's much more we could talk about. But I promised to give Jeff a surprise quiz onstage, and I didn't tell him the questions in advance. It's fine if he can't answer, because I don't remember either.
Jeff Dean: Okay, no pressure at all.
Fei-Fei Li: Do you remember what happened on June 12, 2017? No cheating with AI.
Jeff Dean: Was that the day the Transformer paper was posted to arXiv?
Fei-Fei Li: Wonderful, A+! I'd have to look it up myself to know the date. Did you read the paper on June 12?
Jeff Dean: No, maybe June 10.
Fei-Fei Li: Oh, right, you may even have had to approve it.
Important Technical Advances and Architectural Expectations That Went Unfulfilled
Fei-Fei Li: Looking back on your career, the four of you did a great deal of work at Google and across the field. Let's take that day as a historical milestone, in a sense the beginning of the era of generative models.
Over the nearly 9 years since then, which technical moments have mattered most to you? And please be candid about which predictions were right and which were wrong.
Jeff Dean: A few things stand out. Much of the work we did led to important improvements, like the ones listed on that earlier slide.
The Transformer is a more efficient model architecture for handling these problems. The idea of retaining previous states and then accessing them through attention is valuable.
Earlier approaches instead updated a smaller vector at each step and forgot previous states. Compared with earlier models such as LSTMs, Transformers can learn more with less computation.
Then there was the distillation work that Oriol, Geoff Hinton, and I did. NeurIPS rejected it in 2016, with a review saying it was unlikely to have much impact.
[Note: This refers to “Distilling the Knowledge in a Neural Network,” coauthored by Geoffrey Hinton, Oriol Vinyals, and Jeff Dean.]
Fei-Fei Li: We should make a website collecting all those rejected papers.
Jeff Dean: This one would fit right in. We put it on arXiv, and it now has about 40,000 citations.
Distillation is important for building capable models that are efficient at inference. You can have a trained, large, expensive model serve as a teacher to teach a lightweight student model. It's a fitting analogy to use at a university.
There has been progress in architecture too. I originally thought we would see more unusual neural network architectures. MoE splits into a group of experts, brings them back together, then splits and combines them again, while learning which experts are good at which tasks.
This is useful: a model can have a lot of capacity while activating only a small part of it for each token or task. But it is still quite regular. In most architectures, the experts are the same size and are brought back together quickly.
I had expected to see more irregular, more “organic”-looking networks: some neural activity forming column-like structures, and a few large-scale cross-connections in some places linking distant parts of the network. We haven't achieved that yet.
Fei-Fei Li: We thought about these things when doing neural architecture search too, but we didn't manage to realize them, right?
Jeff Dean: The models produced by neural architecture search are indeed more irregular than manually designed architectures. Manual designs tend to stack one layer on another, occasionally adding some skip connections. Searched architectures can have very unusual connections.
Fei-Fei Li: Like connecting this layer to the 17th layer above it and doing some strange things.
Jeff Dean: Right.
Fei-Fei Li: So that's one thing you predicted incorrectly?
Jeff Dean: I've been wrong about many things. This one stands out; I had thought architectures would develop more in that direction.
Underestimating the Importance of Coding Ability
Jeff Dean: Coding ability turned out to be very important. I didn't fully foresee that it would become a key capability for models solving many different problems.
Fei-Fei Li: I agree. We knew to some extent that it mattered, but we didn't foresee how profound its impact would be.
Jeff Dean: In the Gemini project, coding was one of the areas we focused on, but at the time it wasn't our primary focus.
How to Address AI's Social Impact
Fei-Fei Li: Since this is an HAI conference, we have to talk about the H—that is, humans.
Which of AI's social effects—including changes in employment, psychological anxiety, and discussions about human extinction and alignment—did you foresee? What worries you most now?
Jeff Dean: AI is a very powerful, very general technology that affects nearly every part of society. It can play a role in mathematics, coding, image generation, and self-driving cars alike.
Some past technological revolutions had a deep impact mainly on five or ten fields. AI affects more fields at the same time, so it brings excitement and concern, as well as fear.
Some coauthors and I wrote a paper called “Shaping AI’s Impact on Billions of People,” discussing how AI affects the lives of billions of people.
It's the only paper of mine with its own website, because one of the coauthors made shapingai.com.
[Note: shapingai.com is the website for Jeff Dean and others' research project on AI's social impact. It explores how AI can better serve fields such as education, healthcare, and science.]
We studied 7 areas, including employment and the economy, AI research in science, and misinformation and media. We also spoke with experts in each area to understand the effects AI has already had and the changes it might bring. It was a valuable experience.
My main concern is how to realize as many of AI's benefits as possible across different fields while reducing the harm it might cause. That requires the participation of technologists, regulators, policymakers, and the public.
For example, when people's jobs are affected, they need ways to adapt to new demands, continue to find paid work, and live prosperous lives. This is a multifaceted problem that many people will need to solve together.
Advice for Young People
Fei-Fei Li: One last question. Many people don't know how much you care about the younger generation. We don't often appear onstage together; I remember the last time was at the Grace Hopper conference.
Jeff Dean: You went for several years in a row.
Fei-Fei Li: We're on a university campus now, and this generation of students is anxious about jobs and their economic prospects. Stanford students are no different.
As an AI pioneer, and someone who can see technological changes relatively early, what would you say to these students who are both anxious and hopeful?
Jeff Dean: I can understand that anxiety. The world is changing quickly, and it's genuinely hard to plan for a rapidly changing future.
My advice is to learn as much as you can about new technologies that are emerging or will emerge soon, and learn to apply them to the things you're most passionate about.
You might want to become a doctor, write literature, or develop software more efficiently. AI can help with all of those.
People still need to decide what we want to do together. You should also think carefully about what you want to do and strive to do it as well as you can, including making good use of the tools that have emerged.
Fei-Fei Li: So, in the age of AI, human passion still won't become obsolete. Thank you, Jeff.
Jeff Dean: Thank you.
Interview link:
https://www.youtube.com/watch?v=ooGKNwJWdb8&t=1s