We're sunsetting PodQuest on 2025-07-28. Thank you for your support!
Export Podcast Subscriptions
cover of episode Creators of "The Gradient" on its Origins and Purpose

Creators of "The Gradient" on its Origins and Purpose

2020/7/16
logo of podcast Last Week in AI

Last Week in AI

AI Deep Dive AI Chapters Transcript
People
A
Adi
A
Andrey Kurnikov
H
Hugh
N
Nancy
Topics
Andrey Kurnikov:The Gradient是一个在线社区驱动的杂志,旨在讨论人工智能和机器学习领域的研究和趋势,提供易于理解的技术性概述和观点。 Hugh:他从斯坦福大学经济学专业毕业,将加入Facebook的Kari实验室从事博弈论和机器学习方面的工作,并计划秋季申请机器学习研究生课程。他加入The Gradient是为了提升写作能力,并促进机器学习研究成果的传播。 Adi:他从斯坦福大学数学专业毕业,将开始攻读计算机科学硕士学位,主要专注于统计学习理论,并计划申请AI研究生课程。他创建The Gradient的初衷是为了建立一个AI领域的社区,并提升技术写作能力,填补媒体报道和学术论文之间的信息空白,提供更易于理解的AI研究信息。 Nancy:她从斯坦福大学数学和计算机科学专业毕业,曾在Illumina公司和一家专注于AI自动化的初创公司工作。The Gradient的创立初衷是为了弥合媒体对AI的宣传与AI研究和发展之间存在的差距,并关注AI对社会未来发展的影响。The Gradient为那些没有大型平台宣传其研究成果的研究人员提供了一个展示平台。 Andrey Kurnikov: The Gradient is an online community-driven magazine that aims to discuss research and trends in artificial intelligence and machine learning. It provides accessible and technically informed overviews and perspectives. Hugh: He graduated from Stanford University with a degree in economics and will join Facebook's Kari Lab to work on game theory and machine learning. He plans to apply to graduate school in machine learning in the fall. He joined The Gradient to improve his writing skills and promote the dissemination of machine learning research findings. Adi: He graduated from Stanford University with a degree in mathematics and will begin a master's program in computer science, focusing on statistical learning theory. He plans to apply to graduate school in AI. He created The Gradient to build a community in the AI field and improve his technical writing skills. It aims to bridge the gap between media coverage and academic papers, providing more easily understandable AI research information. Nancy: She graduated from Stanford University with degrees in mathematics and computer science and has worked at Illumina and a startup focused on AI automation. The Gradient was created to bridge the gap between media hype and AI research and development, and to focus on the impact of AI on the future development of society. The Gradient provides a platform for researchers who do not have large platforms to promote their research findings.

Deep Dive

Chapters
The Gradient, an online AI magazine, was born out of a desire to bridge the gap between media hype and actual AI research. The founders, all former Stanford classmates, reminisce about the early days of the project, their motivations, and the collaborative efforts that brought it to life.

Shownotes Transcript

Translations:
中文

Hello and welcome to Let's Talk AI, our podcast where you can hear from AI researchers about what's actually going on with AI and what is just clickbait headlines. I am Andrey Kurnikov, a third year PhD at the Stanford Vision and Learning Lab. And today we're going to have a slightly different type of episode from things we've done before. So before we've usually discussed weekly news or we've interviewed people in AI and

And this time we're going to try to have more of a chat among a few people I've known and worked with at Stanford.

And we're going to focus primarily on a project we collaborated on, which is The Gradient, which is an online sort of community driven magazine centered on AI and machine learning. I'll quickly read through a longer description. The Gradient is a digital magazine that aims to be a place for discussion about research and trends in artificial intelligence and machine learning.

We provide accessible and technically informed overviews of what's going on in AI, as well as a platform for perspectives on recent developments and long-term trends. In short, the gradient points in the direction of the field. And it's been live now for more than two years. There's been dozens of articles. It's sort of grown, I would say more than I expected.

So yeah, it might be interesting to reminisce on how it started and what was the motivation and how it went. So with me today are Hugh, Adi and Nancy, who were formerly classmates at Stanford, now all graduated and moving on to bigger and better things. So to start with, how about I let each of them introduce themselves.

Sounds great. Thanks, Andre. I'm Hugh. I just graduated with a degree in economics from Stanford. And next year, I'll be working at Facebook's Kari Lab on game theory and machine learning. Everything's still off the media at this point, of course, but I'm planning to apply to grad school in machine learning this fall and do more research from that point on.

Cool. And then Adi, how about you do a quick self-intro? Sounds good. I'm Adi. I just graduated with a degree in math from Stanford. I'll be starting a master's program in computer science, focused largely on statistical learning theory. And yeah, I think also probably will apply to grad school in AI. Cool. And last but of course not least, Nancy.

Very sweet, Andre. Hi there, Nancy. Like Audie and Hugh, we met through Stanford. I've been out for about a year now, and I was an undergrad at Stanford studying math and computer science, and recently left and worked for a while with the CEO of a genome sequencing company called Illumina, and then now at a startup that focuses on AI for automation.

Great. Yeah. So it's going to be interesting to see how this conversation among the four of us will go, if it'll make any sense at all to the listeners, but we're going to try and make it work.

And to start with, maybe I can introduce my involvement and why I thought of doing this episode. So I have been helping edit and sort of invite people to write for the gradient for a while from pretty much early on in life before it was even live on the Internet.

And actually the way I got involved was that I was invited by Nancy to write an article. So at that point I had already been blogging on my own site and writing kind of little things. And somehow I had met Nancy beforehand and she suggested I consider writing for this new online magazine that she and some people at Stanford were creating.

So, yeah, from that point on, I had an idea and I actually got started on it. And that's how I got involved. And somehow since then, I have joined the team to some extent and helped other people write blogs and edited many things. And it's been really interesting to see how this project took off.

So given that, maybe I'll ask Nancy, can you tell us how you got started with Gradient and what is sort of your view on the origin story from the early days?

Absolutely. It's funny, Andre, how you remember that story. And I actually, it's really thanks to folks like Andre, who were some of our first writers for The Gradient, that we were able to get it kicked off the ground. We were just talking about this earlier, and I think one of the

first memories I have of getting the gradient started was originally it was just Hugh, Adi, myself, and actually we had one more person, Eric, who's unfortunately not on the call today, sitting together in a coffee shop at Stanford called Coho. And at the time, Adi, who was sort of the glue between all of us, had brought us together and we'd all known each other peripherally and in different communities at Stanford. And I was like, why don't we start an AI magazine?

And this was before all of these AI podcasts and large sort of AI publications had started. And really at the time, there was basically just this bill and maybe some academic blogs. And we fundamentally felt that there was this gap between what the media portrayed AI as being, which oftentimes was this catastrophic force that would take jobs away from people and overrule humanity. And really what was truly happening was

at these academic institutions and also in industry as well. And really the sort of human elements of the, and oftentimes exponential progress that was happening in AI. So going back to the coffee shop, four of us were sitting there trying to figure out how to get this thing started. We looked at each other. Neither of us had ever started an online magazine in AI before. Like frankly, we never, none of us had ever started an online magazine before.

And we decided that the first thing we were going to do was basically sit down together and write a letter of what we thought would be what the values that we cared about for the organization. And that sort of started and became our first ever post on The Gradient, which was an open letter from the editors. And a big part of that initial letter, which...

I think we wrote in January of 2018 or early 2018 was that we wanted to push access in the space of AI and make it more transparent and more accessible for people and sort of bring in more representation. I'm going to take a pause here and let Hugh and Adi also share their thoughts on this as well.

Yeah, for sure. For sure. Thanks. That's quite a vivid picture of, you know, I assume you were all, I think, undergrads at that point. So you were sitting around coffee shop and sort of thinking of a new project, which is always exciting when you're young in your career. So, yeah, maybe next we can ask Adi. It sounds like you were sort of a connective glue guy.

What was your inspiration to think and get excited about this idea and work on it in those early days? Yeah, absolutely. I think there were several different motivations, but I remember early on, probably when I was a freshman or sophomore, I just sort of observed that many of my friends...

that I respected very highly were all working on some AI-related project, whether that was in a research lab or on their own as a side project. And

At least from my perspective, there wasn't really a group of people or a club or a place to talk about these things in some shared context. And so I think originally for me,

I was sort of wanting a bit more community to share some of my ideas and talk with other friends thinking about ideas in AI. And I think part of it at the time was there were probably various forms for graduate students, but as an undergrad, at least from my vantage point, didn't seem like

there was much of an established group. So originally, I think the motivation was community. And I think secondly, I...

I always wanted to... One of my concrete focuses in undergrad was I wanted to get better at technical writing. So I started a blog somewhere around freshman year where I've written about math and computer science topics of various kinds. And I think doing so...

helped me get better at communicating technical topics in a way that was accessible to a broader audience. And I think this general skill is very useful.

And somehow, I think after lots and lots of discussions, the gradient kind of emerged as this way to do both, to create some kind of community within AI locally at Stanford at first, and also as a way to practice and hone this essential skill of technical writing. And yeah, of course, now it's grown beyond Stanford, but I think originally...

Um, it was, yeah, largely about those two things for me. Yeah, that makes total sense. And I think we share this quality of, uh, I also started a blog in undergrad and started sort of discovering that I enjoyed writing things for its own sake and clarifying thoughts. And in some cases, uh, writing about more technical subjects, um,

So that was certainly part of what I think brought us together was kind of the idea of it being enjoyable to write and help other people write. So on that note, I wonder, Hugh, what was your outlook on this early on? What was the appeal and how have you enjoyed working on it?

Yeah, Andre, you bring up a really good point about, like, technical writing and learning the skills of writing. I actually think that one of the big things that brought me to The Gradient was I wanted to improve my own writing and get, like, feel, find a better way to express my thoughts because, you know,

in a lot of machine learning things, the papers are hard to read. And even though they contain really good ideas, if the rest of the research community can't find and discover those ideas, a lot of it doesn't get built up on. And so, Jen?

Just for personal development, I actually thought it was extremely helpful for my own writing and thinking to be able to write about machine learning research, edit other people's machine learning research, and develop all that while also making sure the community also learned more about all the stuff we were writing about.

Right. Yeah, it makes a lot of sense. I've also worked some on the Stanford AI Lab blog and I've generally helped edit people's writing in various contexts, sometimes just personal writing, someone just a personal blog. And it's always interesting how much you can learn about writing and communicating thoughts by helping other people write. Right. And I think that's part of the appeal for us early on.

So, yeah, we sort of covered the early days and what drew you to a project. And it sounds like you sort of started thinking about it around January 2018. And if I remember correctly, I got involved around maybe March or so.

And then we actually had quite a bit of meetings and trying to get initial articles together before it went live on May of 2018. So it's now a little bit over two years old as far as public appearance. And then a couple of months after that, we had the expansion of the scope to include perspectives.

So since we were talking about the early days, I kind of I'm curious to hear how you remember that sort of lead up to launch the launch of a site and how your kind of expectations when you started to do a project versus the actual effort to do it and make it happen, how that went. And yeah, maybe let's start now and see what's your memories on that.

So my earliest memories back when The Gradient first started was, you know, initially it was the four of us. We really, it came out of an idea, it was a complete blank slate, and we sort of decided we were just going to do whatever it took to make it happen and sort of get articles for folks to be able to read The Gradient. And a lot of the initial authors for The Gradient ended up

being from people in our own personal network. So, Audie, Hugh, myself, and Eric as well all spent a lot of time reaching out to folks in our network who had done AI research and who we felt were also excellent communicators who would be willing to write articles for us. And we were very lucky to have a number of people who were very eager and excited, Andre like yourself,

to varying degrees of follow-on collaboration with the gradient of writing articles and then later forming that initial repertoire of articles that we would publish. So what we decided in the beginning was

We didn't want to... We wanted to make sure we had a nice stream of articles when we first launched. So we actually had, I think, maybe a handful of publications or articles that we had stashed in the background before we even published our initial sort of note from the editors. That was our first official...

And at that point, we had made a Twitter account and made all these other accounts with the hopes that a couple people might just follow us. And since then, I think it's exceeded a lot of our wildest expectations in terms of how popular the grading has gotten. And I think right now, we've hit 1 million page views since we first started, and we

more and more monthly viewers and it's been an incredibly exciting and satisfying journey and along the way

There's also been, like you said, Andre, the core team has also changed and expanded over time. So we've brought on a lot of other folks who've been able to help us get more authors, help us manage sort of the website, think about recruiting more editors. So, for example, when we first started, there was sort of the initial core group of editors was with us. And we would solicit articles from other people because we didn't want...

We really wanted outside opinions, not our own opinions to be published in the first sort of couple articles. And we were editing them. And then after a while, we realized that, wow, this editing process actually takes a lot of work. And at that point, that's when we realized, like, wow, we should really get some more people to help us edit. And we started to bring in people from Stanford.

AI lab and then from other research institutions across the U.S. and then eventually outside of the U.S. as well. To summarize, I think the first three months and four months was a lot of us getting folks to help us write articles and really trying to push the vision within our own communities so that other people would be excited to help.

I see. Yeah. And I guess it was good that you reached out to many people, including me. And then I sort of found what you were doing interesting and started sort of giving more suggestions. And maybe I'll point out like this whole conversation between us obviously is not to make any claims as to the importance or significance of the gradient of the larger community.

But at the same time, it's a project we've enjoyed doing and we've had some articles that have seen some leadership and it's been interesting to see kind of evolve and to do as one of the activities we have done to various extents. So that's why I thought it would be fun to discuss.

And on that note, yeah, maybe I'll also ask the other two the same question of like thinking back to your initial idea versus the work leading up to launch and getting it off the ground and basically making it a real thing. How did...

how did your initial imagination of it versus the actual process compare and how do you remember it now? And I'll let Ali speak to that first. Sure, yeah. I think Nancy summarized many of the beginning conversations and discussions well. I think one of the

important moments early on was when, I think as Andre mentioned, we decided to launch Perspectives. So looking back at some of the earlier articles, we sort of just asked our friends to write about a topic that they were interested in. And most of them were

fairly, um, fairly factual in the sense of being, um, very, very detailed and, and, and, uh, and interesting, but not really making an, uh, an opinion. So, uh,

I guess the word is the original articles were largely expository. We had one that Hugh wrote on speech recognition systems being vulnerable to adversarial attacks. We had an article on machine learning on graphs. We had an article on semantic segmentation. But I think Andre, his first article was something like,

why reinforcement learning is fundamentally flawed. I think we renamed it to Reinforcements Learning Foundational Flaw. And I think we needed a slightly different platform to share opinions. And I guess it was about a month after we had published our first articles that we launched Perspectives. And I think these are...

in my view, I think some of the coolest pieces we've put out where the authors have been able to share some really strong takes on various topics in AI. So I think that was a pivotal moment. For sure. Yeah, I agree. And I think it is interesting to point out how

You know, once you start on a project, once you started a side project, you never really know what it's going to become. Right. It sort of evolves in a sense. And so this idea of perspectives and sort of what types of articles would be on the site kind of evolved and wasn't necessarily something you knew from the start. But then it sort of happened.

And yeah, I think lately, so the team has been changing and evolving. And I think lately our official lead of the team and of the site and the whole project has been Hugh. And he has been around and trying to keep us in a consistent pace as we have started getting more articles and more collaborations and things.

So, yeah, I think same question, Hugh. How have you how do you look back on the early days getting it launched and then kind of sticking around and doing this side project alongside your undergrad, alongside research, along other things? Like, how do you think about it looking back?

I obviously underestimated the work that the gradient would be in the same way that I underestimate the work that any project that I undertake. But I think it was especially true for the gradient. Some other things that I got wrong is I actually thought that overview would be the most important thing that we did. But I think it's actually around 50-50 between overview and perspective with possibly perspective being slightly more influential at least later on.

There were lots of little things about how to write the tweet, how to post on Reddit, what time to post on Reddit, stuff like that, which we basically learned through trial and error. A lot of it was just testing out things, seeing if it worked. If nothing really drove, we would just keep doing it. I personally had a lot of fun trying to learn quite a bit about how an online magazine works.

Yeah, yeah, exactly. So that's another reason I think I wanted to have this chat is it's kind of fun to just reflect on how projects go and how we sort of like projects that take teamwork and that go on for a while, how they

You do learn a lot as you go. And basically, I think it's interesting that the three of you as undergraduates just decided to do this, right? Having not done anything and not been directed to do it necessarily. And then, yeah, you figured out how to do it on the fly. And then we are still figuring out how to do it on the fly. I mean, it's still a scrappy kind of side project, but

But I think it's very, very fulfilling to have something like that. And I would say that I want to partially share this with the gradient listeners to let them know that this is kind of how things go. Is you have an idea that seems exciting and you start working on it and then you realize all the little things that are required and how much work it takes to

But as long as you still enjoy it and find it exciting, somehow, maybe eventually it works out and you figure something out and it actually becomes a real thing. Moving on from that, maybe on a slightly broader topic, I wonder sort of...

Yeah. How much do you read about AI? What are you thinking about AI as a topic? What drew you to it and drew you to making a project related to it and its discussion early on? And again, let's do our little cycle. So Nancy, yeah, what is your current, let's say, large view on AI and an interest in this sort of writing?

Wow, that's a big question, Andre. So I'm going to narrow it down and just try to answer sort of what exactly about AI we wanted to address with the gradient and what the sort of initial impetus for the publication was, and then sort of maybe some thought reviews that we're interested in.

And I personally am very excited about today. So I think the answer to both is actually kind of similar. So when we first started The Gradient, there was, like I mentioned before, we saw this gap between what was portrayed as AI in the media and what was actually happening in terms of AI research and AI progress and sort of its implications for society and for work and for humanity. And

I remember when we first started the gradient, one of the words we used a lot was hype. Because a lot of what we wanted to target was how do we distill the hype or how do we counteract the hype that the media was portraying where you would look online and then suddenly in CBS News or something, we'd be talking about AI taking over the world. And what's really...

for me from like an AI perspective right now is thinking about how AI is going to change the way society is going to look in the next decade or several decades. And I think if you look at, for example, in the early 1800s, I think about 85% of jobs were in agriculture. And every decade since then, so the 1810s and the 1820s, there were 5% fewer people who worked in agriculture. So by the 1900s,

of Americans no longer live in agriculture. And I think this is something that we're going to see in the next couple of decades with AI, where I think a lot of people are scared that like AI is going to change the way that people work, um, and, and the jobs that we have. And, um,

And I think that's rightfully true, though I think it is oftentimes exaggerated in the media. And what will really happen is, I think the transition is inevitable and it's technological progress in the same way that steam engines or steamboats or even like fundamentally electricity or farming technologies has changed the progress of society. And really the underlying question is,

bringing people on board to understand where AI is going, what it looks like today, and helping them understand that it isn't necessarily going to mean that jobs will be lost. It's more of a transition of jobs being lost and new jobs being created, and sort of those new jobs being potentially more creative or sort of more topological in nature. And I think if you walk into a room today now and you ask people,

10 people, if they want to go back to a farming society where they're farming weed all day long, 9 out of 10 of those people probably won't say yes. And it's interesting because I think when people think about AI now, it's very much the same conversation where you ask them, well, they view AI as this sort of demonized...

that in two years is going to make them lose their jobs. And I think that's very sad because it's actually, there's so much potential for AI to do good in this world and helping communicate that and sort of

helping more people get on board with that larger picture of where the field is going so that it isn't some scary research that only very few people and labs are privy to is very exciting. Yeah, definitely. That's a great reply to my overly vague, very big question. I think I agree with a lot of what you said. In particular, I think

What I like still being an editor for Gradient is this idea that it's a platform where researchers can talk directly to non-experts instead of having to go through journalists or through media experts.

that is a little more closed access, the gradient you're trying to make it so anyone can collaborate with us. And as we get more readership, we can allow researchers to share their perspectives directly. And that's also partially the idea of this podcast, of course, is to sort of let people outside of AI who think it's interesting

get to perspectives of researchers and listen in directly on those thoughts. I guess I'll pose a similar question, maybe a little more specific to Adi. You've said that in undergrad you saw a lot of people around you, a lot of cool people working on AI. So I'm curious, kind of, were you also drawn to AI and its technical concepts and potential?

And was that interest part of the reason why you got inspired to work on something like the gradient? Yeah, absolutely. I mean, I think my interest in AI,

to be honest, probably just came from the fact that there are some cool applications. I did a lot of robotics in high school and thought that seemed cool. And I just kind of got more into AI once I got to Stanford. I think when it comes to the gradient in particular, I...

I felt that a lot of the times at Stanford, there is always just like a lot happening within CS, but in particular AI. I know that in the past decade, CS majors as a fraction have gone up at Stanford by like five times. And I believe the most popular concentration within the CS major is the AI track. And

Yeah, it always seemed like there was a lot happening, but it was hard to know what was important amongst the thousands of research papers published in a year, which ones were actually going to be important.

And I think the gradient in that sense was largely about solving a problem I had, just trying to find the signal from this vast array of things happening within CS and AI in particular. And I guess also just echoing something you said, Andre, I think that previously...

And I guess today as well, my main two sources for learning and reading about AI have been hacker news and maybe sources like MIT Tech Review. And then on the other end of the spectrum, reading archive papers or perhaps using the popular open source package Archive Sanity, which I'm a huge fan of.

But it always seemed like there wasn't quite anything in between, something that was a bit more researchy than MIT Tech Review, but also readable and something that doesn't take several weeks to understand. And somehow I think the gradient was a way to...

fill this middle ground that seemed to be lacking, particularly when I first started at Stanford. I see. Yeah, it makes total sense. And I also view it a little bit like that where...

I mean, at the time, there were already kind of a tradition of blog posting started by people like Andrej Karpofy. And I think all of us read some of these seminal blog posts and were very inspired. And part of what I like about the gradient is that if you haven't already started a blog or you don't already have a lot of following, but you like the sort of writing and you would like to try your hand at it to benefit the community,

Now there is a place where you can work with editors and work on something without necessarily having to start a blog and do that all on your own. So on that note, maybe I'll ask you, is that also your outlook? Where do you see the gradient slotting into a broader conversation? Is it kind of what you thought of when you first started or has it changed?

So I agree with everything that Adi and Nancy have said. Well, I think one additional thing that the gradient does is it exposes the research of people who don't already have big platforms to announce their stuff on. So people with large Twitter followings or, you know, widespread name recognition can already

uh, spread their research really strongly. But a lot of people who don't have that can also write an article on the gradient. And oftentimes the well-written, if it's well-written and popular, it will, uh, like be read by a lot more people than would be read by normally. And I feel like that's one of the, the,

the most important thing that the gradient does. Because otherwise the discussion is just dominated by whatever is trending at the moment. And in this way we can get a lot of unheard but still very important voices out there for everyone to hear. Definitely. That's a great point. Yeah, I think on that note we've covered a lot of the origin story and the motivations, the process, the current outlook. I think it's been interesting to reflect and reminisce and also fun

To cap things off, I'll just ask any last thoughts, anything else you'd like to add with respect to the gradient or maybe AI or anything you think might be interesting to mention to listeners?

So I will say we're always looking for writers and excited folks who want to be a part of the community and helping edit and really sort of push the mission of The Gradient forward. And we welcome any feedback from folks who are listening to this podcast. And also, Andrea, thank you so much for taking the time to sort of reminisce with us a bit about the early days. And it was very fun. We haven't had a chance to talk about it

the gradient since it happened two years ago. Okay, yeah, that's a great note to add on. That's a great point. Let me point out that the thing we're talking about is the gradient. You can go to the gradient.pub. That's the URL to see it. We have a link to contribute on the navigation bar. So if you want to write to us, you can take a look there.

Thank you, Hugh, Adi and Nancy for joining us for this chat. I hope people do enjoy it. And of course, if you've enjoyed this episode of Let's Talk AI, please rate us on any platforms you're using and share it with your friends and tune in to our future episodes.