Hello, and welcome to the NVIDIA AI Podcast.
I'm your host, Noah Kravitz. I spent most of Nvidia's GTC 2018 conference locked in a glass booth.
Don't feel too bad for me though. I had about 15 of the best conversations anyone's ever had at a tech conference during those two days.
And lucky you, they've all been recorded to be published.
Some of them published already. to the NVIDIA AI podcast.
So as they say, subscribe wherever you get your podcasts.
All right, that shameless plug out of the way.
Let's talk about the day I got out of the podcast booth.
One of the talks that I made it to actually led to today's episode.
It was given by our guest today, Anamesh Garg and Marnell Vazquez of the Stanford Vision and Learning Lab down at Stanford University in Palo Alto.
Their talk was called Generalizable Autonomy for Robotic Mobility and Manipulation And essentially, we talked about, or I should say they talked about, and I listened, to teaching a robot to cook Teaching a robot to avoid those awkward, shall we dance moments that happen when two humans come face to face in a crowded room and can't figure out how to get around one another.
In more broader strokes, teaching robots to move from specific tasks to generalizing their learning to apply to new situations.
Pretty awesome stuff, but better than me tell you about it.
Welcome Animesh and Marnell. They're both postdoctoral researchers at the Stanford Vision Learning Lab, SVL, like I said.
And it is really a pleasure to have you on the podcast.
Thanks so much for joining. Thank you very much for inviting us, Noah.
Yeah, thank you. Let's go back. Let's start by talking about your GTC talk.
And there's actually a blog on the NVIDIA blog as well that folks can check out.
But we'll start with you, Marnell. Tell us about Jackrabbit and robots and crossing crowded rooms.
Yeah, so Jackrabbit is a robot that is central to a project on enabling a mobile agent to move around people in a way that is appropriate, in a way that doesn't bother nearby bystanders or users in, of course, human environments.
And so generally we speak about this as trying to address the problem of social robot navigation.
And Jackrabbit is a robot that we put together to basically try to get as much sensor data as we can from the environment. to then enable this mobile platform to move from one place to another nearby people.
And so before we get into the hows of how you did it, a little bit of history.
There's a picture on the blog and More information, there's a YouTube video that came from SVL, I believe, of a robot All dressed up in a hat with, I think, a cardinal colored sash on the hat or ribbon on the hat. and a tie roaming around the streets of campus.
So you actually or the lab actually deployed Jackrabbit back in 2015 initially?
Yes, that is correct. I think now the the actual robot has been on Stanford for at least three years.
And kind of the story behind The tie and the hat is that often moving around people is not just about the particular velocity say that you put in your robot, right?
There's a lot of social components that come to it.
The tie is just to represent, right, that we're trying to make this robot polite as it's navigating from one place to another.
Manners may be in jeopardy in human civilization, but amongst the robot society.
I liked it. Taking these social cues seriously.
And so what did you learn? What were some of the problems, the hurdles, the challenges that, you know, you, you got past in, in, teaching the robot to move around you know uh both the example of the crowded room but also um you know jack robin was making deliveries on the stanford campus and so learning to be safe and polite and appropriate out amongst the people.
Well, one of the first things that happened when you start talking about this project, right, because you put the robot out there and then people start asking, right, what are you doing?
And then you say, well, we're trying to have this robot navigate around people.
One of the first reactions that you get is that people get all excited about standing in front of the robot.
Just to kind of tease and see how it would react to them.
And of course, safety to us is very important.
And this raises one of the hardest questions that actually we have not solved yet, which is how can we have this robot try to influence the people around it so that they collaborate with it, right?
So that they actually help it to get through where it needs to go. rather than just, I don't know, exploring and kind of play with it.
Sure. And that I would say is kind of a long-term goal of, of this project that actually we don't have a good solution yet, though we have some ideas of what we might want to try.
Does that, without getting too far away from the vision and deep learning stuff we want to talk about, does that get more into social engineering of a situation?
Yeah, that definitely kind of falls within the area of human-robot interaction. and how you want this robot to basically become a social agent more than say, a computer with wheels in our environment.
From what I think of as more technical, but really what you're just describing is a huge technical issues or problem to solve as well.
But turning to the computer vision stuff and using deep learning, tracking the movement of not just one human, going across the space, but multiple humans who are traveling on different vectors and Jack Roberts got its own path that it's following.
Tell us a little bit about tracking the untrackable and kind of figuring out what I think you referred to as predicting human trajectories as part of navigation.
Yeah, so this is a project where the main person behind it is a PhD student in our group.
His name is Amir Sadigan. And the idea behind tracking the untrackable is thinking about tracking beyond mirror pixels.
And so normally in computer vision, when you think about the problem of knowing where people are and knowing how they're moving, you're basically associating pixels from one frame to the next that hopefully correspond to the same person.
And here in this project, the idea was to think about not only the appearance of people, which is what typically the pixels directly tell you, but also think about the way that people are moving, and also think about the way that they are interacting with other people in their surrounding.
And so the particular model that was proposed in this case is a model that has a structure of recurrent neural networks. that each of them takes into account these three kinds of features.
So appearance, velocity, and interaction with other people.
And then by means of optimizing for a particular score that allows you to associate whatever you're seeing at a given time with whoever you've seen before, then we optimize the network to basically mix these different kinds of information together so that you get a good tracking result. for sort of the lay person like myself at all.
When it makes sense, it stems from thinking about human behavior, which obviously is part of what you talked about. at the show and in your work.
And I'm thinking both about watching a room and thinking about watching sports.
The whole idea of studying your opponent and learning their tendencies and knowing, oh, in this situation, this player always tries to drive to the right.
And this player always goes left. And I'm imagining poor Jack Robert trying to cross a campus full of college kids who are heads buried in their phones, perhaps, who knows. uh and trying to to understand um and predictive behavior what kind of um success have you had?
Or even more interestingly, what are some of the things you've learned that maybe surprised you along the way?
I'd say in terms of success, like we have very good results in situations where normally there is very brief occlusions, but significant number of them. which tends to be challenging for this kind of problem, particularly tracking people.
But I would say we're still kind of having difficulty in terms of being able to connect people who disappear for a long time from the view of the robot.
And this, I think, is interesting from the perspective of Well, how much information should the robot remember and for how long, considering that At least in our project, we're trying to process everything real time on a mobile platform.
Right. And so just from a practical perspective, the challenges involving making sure that this can run on board. is often something that turns out to be difficult to estimate from the very beginning.
Was the project either from the start or was there a point after which you started designing things specifically for the power constraints of the platform?
And as I ask that, it seems kind of like an obvious question because if you're building a robot, it has to be mobile.
But I'm just kind of wondering, you know, sort of along the way where you started to realize, oh, we can work with five seconds of memory or 10 minutes of memory or what have you before running into those walls.
Yeah, I would say at least for our research, we don't think too much about that when we start as in terms of trying to push the state of the art. often requires you to kind of push the limits also in terms of computation.
But it is part of our goals in this project to transfer everything that we do that is related to the problem, to the platform.
And when this happens, then yes, of course, these problems come along and they're very, very important.
And so one of the things that I think is special about this project is that it is, at least to my knowledge, one of the first robots that has had a full onboard GPU to try to do as much as we can in terms of deep learning and computer vision mobile on the go.
And now that we're actually building a second robot, we're now doubling our capability by adding not only one, but two GPUs to try to continue that trend.
So let's turn to your work, Animesh, for a moment.
At GTC, you talked about, well, you talked about cooking and YouTube cooking videos, which got my attention immediately.
But once you had my attention, you really lured me in with talking about something called neural task programming.
So tell us about your work with neural task programming.
So that's actually one of the recent pieces of work.
Let me back up for a minute before I actually get into the details. sort of a lot of the work that we do here in the group is driven by this a challenge task of understanding videos and being able to go from YouTube style videos to have a robot actually attempt or have a policy to do the task.
And there are a variety of technical problems that need to be addressed on the way.
It is a very good challenge task. but we as a community are not yet ready for in the next sort of six months.
And one of the problems in this high and mighty goal, I would argue, is we need mechanisms or we need algorithms that can look at the video one time or for a new task, let's say a new recipe and be able to make up a plan for this particular recipe on the go.
And that is what new task programming was about.
The idea was that if, let's say, let's take an example of you trying to teach me to cook your favorite dish.
Let's say if you were in a kitchen and you were going to show me how this is done, you would probably assume that I know how to handle the spoon.
I know how to turn the stove on and so on and so forth.
The basics would be assumed. Yeah. I'm such a bad teacher.
I'd probably assume you knew how to do the whole thing and I'd sit down and let you cook.
Of course. But yes, no, I'm with you. We're assuming a baseline of some common knowledge.
And then what you do with, let's say, the ingredients in the room or in the kitchen is going to be something that will be new to me, assuming that this is a recipe that I have not seen before, or at least there is a twist to it that I do not know about.
And That is what Neural Task Programming is about, that it takes in a new demonstration and based on some prior set of baseline actions, which can be thought of as primitive actions.
It can compute a plan that will explain the new recipe with one demonstration.
Now that we understand the context, what I showed in the talk and what our current work can do is In this setup, to kind of be able to demonstrate the power of this approach, what we do is we create a domain which is fairly simple but abstract and rich.
So the block world domain is a toy domain in one sense, but it's very rich because it really enables us to explore the set of quote unquote, recipes, where the recipe can be any sort of structure you can build within this domain.
And there are very many to be able to show beforehand.
So you can come in and do whatever you like with this domain and the system can One, parse what you did, and two, now attempt to do it itself.
Because this is not just a parser, if you will, it can actually correct its own mistakes along the way.
So if you do it one particular way, and then the robot or the agent attempts it but fails, it does not need to sort of throw its hands up, but it can actually attempt to keep doing the task based on its knowledge of the domain.
Okay, let's unpack a couple of things there for a second.
It's one thing to feel like I understood your talk, but actually talking to you about it, there are so many questions I have.
So the first part of it where I can, the robot can watch, if we go back to our example, I can teach you how to make spaghetti and meatballs.
And you can then, or the robot, and it's standing up for you, could then break that down into those individual tasks, the boiling water and cooking pasta.
The making a marinara sauce, um, you know, making the meatballs, depending on if we go, how far back we want to go, you know? either just cooking them or going back with the... You get the idea.
Yes. And then you can apply those things to a new situation.
So you could... make pasta on its own. You could maybe, somebody else could teach you how to make a totally different kind of sauce, and then you could combine the two things together, that sort of thing.
So we've got that, but tell us a little bit more about the robots being able to learn from their mistakes.
To what extent can the robots in the system learn from their mistakes and sort of recover and reteach themselves?
So in this particular setup, this is sort of a longstanding goal to do reasoning at a very long-term plan.
So let's say in this cooking example, what happens if you completely mess up?
Can you redo the whole thing? That seems to be an open problem.
In this particular work, what you can do is, let's say if the task was boil water fry meatballs and then add sauce.
And if you accidentally spilled water, after boiling, then you can go back and boil the water again.
Basically, the setup is that You have to add you have to boil water, then add pasta to it.
And if any of those individual steps fail, then the system can go back and fix those. or at least that's the idea.
And how new is this idea of NTP neural task programming and leveraging hierarchy and modularizing these tasks.
How new of a thing is that? It sounds fairly radical.
I would like to claim it as my own, but I have to say that we are here standing on the shoulders of giants.
These ideas of modularity and compositionality have been around in AI and in robotics for about 30 to 40 years. the novelty or the revival really here is being able to combine these ideas of modularity, symbolic planning with deep learning.
This is where neural task programming is new in the sense that earlier we had to hand-design these state machines or the sequences of primitive actions, But now the system can do that automatically.
So the novelty or really what is the radical here is Being able to take in a video and convert that into a sequence of primitive actions and being able to take that representation and go execute that.
That is really novel. Our guests today are Animesh Garg and Marnell Vazquez.
They're postdoctoral researchers at the Stanford Vision and Learning Lab. at Stanford University.
And we're talking about their work broadly and specifically stemming from a talk they gave at NVIDIA's GTC 18 this past year covering, well, Robotics, but neural task programming, the dream of general purpose robots, tracking multiple targets and human behavior in crowd spaces, all kinds of really fascinating stuff.
It's research being applied to real life situations we can all relate to, which amongst other things makes a great podcast fodder.
But let's shift gears for a second and maybe talk a little bit about how you each got into the field and wound up at SVL.
Marnell, did you grow up dreaming about electric sheep, so to speak, and wanting to become a robot researcher?
What was your path to Stanford? I like technology, I would say, and I don't think I really had a very clear path all the way to here.
I was born in Venezuela in South America, and I did my undergrad all the way to there back at home.
And then when I was finishing undergrad, typically people would go and work in something related to IT.
So I did computer science there. And I had worked on robots during my undergrad, kind of the way people do it here in the US in high school. with Legos.
And that kind of got me very, very excited about the opportunity of mixing computer science with something physical.
And so when I was finishing, I just applied to grad school programs just to see what would happen.
I was very lucky to get into Carnegie Mellon at the time to do a funded master's.
And this was with Aaron Steinfeld, who was in the Robotics Institute.
And so going there to me was kind of a eyeopening experience.
I got to learn about human robot interaction, which is what my main area of research is now.
And I got to see everything that was happening there from mobile robots all the way to manipulation to... even crossovers with other areas like NLP or human computer interaction.
And so after that, I did a Ph.D., in Carnegie Mellon, and when I finished, I applied for jobs again, and I was very lucky to get a tenure position at Yale University.
So I'm going to be starting there this summer.
Oh, congratulations. You know, I want to learn a little bit more in terms of like managing students and doing all these things that often you can do as a PhD student.
Right. And when I was looking for what to do for a year, then I found this project at SVL, the Jackrabbit project.
That was actually a very nice continuation from my work at Carnegie Mellon during my PhD, where I was looking into the way that people stand with respect to one another during social conversations to try to infer things like who's talking with whom.
Do you have someone actually be part of a conversation or not based on their body posture?
And this, of course, in the context of human-robot interaction.
And I must thank also Disney Research for funding all of my PhD during that time at Carnegie Mellon.
And so yeah, I would say I've been lucky.
This was not something I planned, but somehow I got to here.
Well, for what it's worth, I would use the word fortunate because I think your own work had a little bit to do with your your good fortune so far.
Marna, how long? So this is a one year engagement at Stanford for you.
Yes, correct. And Animesh, how long have you been at Stanford and SVL?
I've been here for about two years now. I started in fall of 2016.
And before that, what were you doing? What was your journey been like?
I grew up in India and as most kids, it was fun to sort of start out career in engineering, I was actually as far away from computer science or at least as I wanted to be because I perceived the computer science job to be office job and I wanted some more action so i decided to study mechatronics or robotics in undergrad itself undergrad was fun.
It allowed us to build robots, not just sort of program robots.
And I really got into robotics during undergrad, where this is back in 2008, 2009, when autonomous cars were not yet a thing, but people were still working on them in research.
And as a naive undergrad, I decided that I will just build an autonomous car on my own with $250 budget.
As one does. Yeah. But as mistakes teach you, it was a one year long artist journey.
We failed, but it was a lot of fun. It wasn't a full autonomous car, but it did run autonomously somewhat.
And that really led me to that belief that I liked it enough that I wanted to go to grad school.
At the time, I wasn't sure what I wanted to do for grad school.
And then when I came to the US, I first started grad school at Georgia Tech, studying operations research and statistics.
And I realized a lot of the machine learning algorithms and algorithms in general have very close relationship to ideas in operations research and supply chain management.
So I started studying How can we bring algorithms to robotics?
And from there on, I started at Berkeley within about a year.
West Coast was much better than Atlanta.
And I got started into the grad school studying healthcare, algorithms for healthcare.
Oh, interesting. Okay. In principle, we were looking at how we can help surgeons by bringing algorithms and robots into the clinic. to enable surgeons to do better.
It was never about replacing people as much as it was about making people do more.
And along the way, we worked on cancer treatment.
We started working on robot surgery. And it turned out that a lot of problems boiled down to this fundamental problem of what you may think of robot manipulation, robots moving things around on a tabletop.
And it turns out that that led me to a much more basic problem of robots working with rigid bodies or simpler objects than surgical machines. setups.
And that led me to Stanford. When I finished PhD, I was still working on surgery.
And then I wanted to study slightly more fundamental basic questions of perception and planning and learning in robot manipulation, which led me to Stanford, which was a powerhouse of deep learning at the time.
Right. Since then, we have been working on mechanisms for robot learning, for doing cooking or for doing manufacturing or a variety of these tasks where a person robot can be an assistant to a human in its everyday tasks.
So to what you just said, and you characterized it really well, and naively, to me, it seems like it's a fundamental thing. thing of research and applied research.
But, you know, you're examining these building block questions and problems.
And this is a question for either of you.
Do you see either in the recent past or now, or perhaps looking forward, that we might be approaching some kind of an inflection point where there are specific basic robotics problems in robotics or problems in AI and vision that Once a workable solution is found, it's really going to unlock all of these other things where we're going to see kind of an accelerated rate of growth in certain areas?
Are there these kind of big golden rings on the horizon?
I would like to think so. I really think robotics is the next sort of personal computers.
If you remember in the late 70s, early 80s, people had a similar fear of computers bringing automation to jobs as we have about robots today.
But we clearly know now that that has not happened.
Right. at least I see robots going down the same path that computers or personal computers started where I believe recent advances in compute, a lower cost of hardware of robotics, together have now brought about a car.
And when When we can reduce the cost of a robot to somewhere around $10,000, that's when you can put robots out in the hands of people who are not experts. and have them use these things first as luxury products, but then as everyday products.
I would argue people outside who are not experts will surprise us with the kinds of cool things that they can do with these machines and tools that we are not building them for right now.
Much in the same way we use computers to do digital design, publishing, writing, so on and so forth, so much just moved using computers rather than without.
And that kind of changed the way a lot of industries worked.
Are you familiar with any of the current, what I'm seeing marketed as wave of social robots? and or without getting into reviewing specific products on the podcast, Google Assistant, Siri, Amazon Alexa products, those kinds of things.
Do you characterize those as robots in some way or is to you given, and both of you given your backgrounds, Is a robot inherently a physical thing that can move mechanically?
Like we could go on and on for days, I think, on trying to describe exactly what a robot is.
And I personally don't think that's a very useful term. you know, discussion to have.
I rather focus on the question of what can we do with this that will have a positive impact in people and what are the characteristics of that particular thing that we have that is an agent, right?
That is interacting with people. How does that change the particular problems that you're solving or the way that you're interacting with people.
In my personal opinion, I think all of these different devices that communicate with you verbally, but maybe are not manipulating the world as as some of the robots that we have here would, I think they still can classify as a robot in the sense that it's a situated agent that can influence its world, it can influence people around it, and it can definitely change, connect now IoT, right?
It's happening everywhere. It can connect to other things to basically change the state of the world. in a particular way, hopefully a good way for the people that are around.
Now, I want to hear what Animesh has to say.
I would love to be controversial, but I kind of have to agree.
Not to interrupt, but it was an eloquent answer.
Hard to disagree on this end as well. No, no.
Yeah, I think I'm with both of you. It's hard to disagree and I counted that argument, although I do have to add to this that while the algorithms that are being built that enable these agents, whether a robot in physical world or a robot which is, let's say, a Google Voice or Alexa agent. are very similar.
There are in particular some challenges that are much harder when a robot has to physically change or interact with the environment. versus where the system is not actually a physical agent.
Think of it like this, If Amazon Alexa makes a mistake and misidentifies what you said, you can just correct it on the way.
But now think of this in terms of your autonomous car or a robot that is handing you a glass of boiling hot coffee.
So the mistakes can be much costlier. And at the same time, there are slightly domain-specific challenges that are much harder when you're interacting with these things.
But that said, that is, I believe, going to drive both of our research careers.
This is a problem that we like to joke about is an open problem.
And it is a very... a right problem to enter the field because this is going to be a fairly challenging problem for a while.
So given that, where do you see things headed in your field or if there's another area of robotics that you want to talk to in the next five years out, 10 years out?
And even outside of the little bubble of the Bay Area, Peninsula area, are most of our deliveries going to be handled by robots?
Is the dream of a general purpose robot something that you see taking shape in the next 5, 10, 15 years?
Where do you think we're headed in the short term?
I would say the general purpose robot that we all want to have is a bit farther away than five years.
I don't want to say a number because that's very risky.
However, I'm going to speak from the side of of human robot interaction and from a technological point of view, what I think will happen within that timeframe.
So one of the things that is going to change the way that we interact with embodied machines is how much of the subtle things that people do will be able to measure. with the sensors that we have available, not at a crazy price, but rather at a commercial level.
And just to remind you of something that happened that, I don't know now how many years, but when the Kinect came out.
About nine years ago. Nine years? Yeah. Time flies.
The Xbox Kinect is nine years old already.
Yeah. So when the Kinect came out, the revolution in terms of gestural interfaces was huge.
And it was because that availability of body pose information was right there for anybody to use.
Right. And similarly, the same is going to happen when we can actually understand very tiny gestures that people do with their face, but that we humans are very good at feeling and at capturing. but that today robots completely miss out of the picture.
And today, for instance, when I was driving in the morning, I had this car in front of me.
And the car was barely moving. But I could very quickly sense the car wants to change lane to the left.
And I could not see the wheel of the car, but I could see the person's head, like half of the head turning towards the left and going back and forth.
And then like, I don't know, maybe five seconds later or something, the person turned on the light to turn.
The signal, right. And this is something that I thought for a second, whoa, if we could actually perceive this very subtle things are happening all the way around us that we're very good at sensing by means of us being there and realizing that they're happening all the time.
This will completely change the game for human.
And is that more of a hardware capability problem, a software algorithm problem or still both?
I would claim it's both. Yeah, that makes sense.
Along those lines, in addition to being able to do this, then the question is, how would you do that, right?
And me, of course, being at Stanford, being in SVL, I'm a strong believer that data is going to be able to help us unlock a lot of these problems.
And so one of the things that is very hard to do today in human-robot interaction is collect big amounts of data, of big data sets, to train algorithms that say for object recognition are working wonderfully.
The question is then how can we scale from a human-robot interaction perspective to get as much data as we're seeing in other fields kind of have such a positive impact.
And at the same time, how can we make algorithms that are more data efficient?
Is the problem that there just aren't enough touch points to gather the data, or...?
Well, it's so expensive, right? Like if you want the robot to interact with a person doing X or Y, for how long would that person be willing to be there In a way that he or she would not start, you know, just like, oh, doing the task in a very bad way or completely changing their behavior.
Yeah. There's something kind of manufactured about the interaction.
And so those two things, I think we're going to see big changes in the next couple of years.
Animesh, what do you see down the road? I think the idea of general purpose autonomy is like a carrot hanging in front of us.
And there will always be newer problems.
There will always be something that humans can do, robots cannot do, and we would want them to do that.
So this problem is going to be ever changing and ever evolving.
That may sound frustrating for you, but I think for a lot of people, it's an optimistic point of view.
We'll always be one step ahead of the machines. thing, I guess.
Not saying that machines will overcome us is really, according to me, undermining and understating The imaginative power of humans, we have persisted for a long time.
And I believe that's going to hold true for a long time to come.
That said, I think in the last five years or so, the big advances that we have changed. had in machine learning in a lot of compute hardware and also a lot of hardware for robotics has really enabled us to capture problems that were very structured earlier.
So robots were there making our cars all along.
We just did not know about them. But then now we have been able to take away some of those structures and take those robots out of the cages to maybe put them alongside people in their houses, maybe put them alongside people in small manufacturing shops where we can build systems or products that are more customized rather than mass produced.
And I think in the next five years, what we are going to see is with the one advent of e-commerce and this idea of delivery on demand, We are going to see more and more robots powering the Amazons and the Alibabas of the world. in both front end and back end to do these kind of deliveries. autonomous cars will be available in at least the developed part of the world.
I see that getting similar technologies across the world will take a few decades.
It's very much again like cars. The cars were up and running in the U.S. in the 20s, but really they became a thing everywhere else much more after the World And the technology arc here is going to take pretty much the same trajectory where we are going to have early adopters in the developed world or in US and in Europe for autonomous cars, for delivery robots, for warehouse robots, hopefully even for robots in retail stores like Walmart and the H&Ms and Macy's of the world.
While I believe the cooking robot is probably a few years out and so is the surgeon or the robo-doc.
But at the same time, I think this is such an exciting time that a lot of these new technologies are breaking ground.
We are seeing robots that are helping humans extend their ability really in a variety of things. enabling new tasks that were just not possible earlier, like deep sea mining, deep sea exploration, mining of metals in very harsh conditions, space exploration, all of these are places that will be totally new that were just not there maybe 20 years ago.
I'm now just seeing the carrot on the string.
And you know what? As long as it's a little ahead of us, we have more to chase after.
And that's a good thing. Animesh and Marnell, if folks would like to learn more about the work that you're doing, your colleagues at SVL are doing, I know there's some YouTube videos folks can watch, papers that can be downloaded and read.
Where online should they go? They should go to svl.stanford.edu.
That's the website for the group. And we are starting out a new blog.
For the general audience to keep up to date to our research and our plans. at Stanford, and I would suggest that you come and subscribe to that blog.
Excellent. And Marnell, as you mentioned before, you will be headed east to start a new position at Yale this summer?
Yes, correct. Best of luck to you with that.
They're lucky to have you. Best of luck to both of you and all your colleagues for more Just really interesting and obviously groundbreaking work that you're doing down at SVL.
And thank you so much for making the time to join the podcast.
No, thank you. Thank you very much, Noah, for giving us the opportunity to talk to you.
It's been wonderful.
