Thoughts on web development, tech, and life.

Category: NLP (Page 1 of 2)

AI and the future of jobs

I was invited to give this year’s keynote address to the German International School of Silicon Valley, where both of my kids attend. The high school students have started taking internships and thinking about jobs, so the school wanted me to share some thoughts on how AI may change the future landscape for work and how the students should think about preparing. This is something I have a lot of thoughts about, although (as you’ll hear) I’m the first to admit that no one has a clue how the future will play out, especially these days.

Nevertheless, I was excited to pull together a number of strands from my own career and related topics I’ve read and thought a lot about like wealth inequality and even music, and deliver it in what I think turned out to be a pretty engaging and broadly consumable format. I was pleasantly surprised how many students (and also their parents) came up to me later (even days later) to say how inspiring and thought provoking they found it, so I thought others might like to hear it to.

My wife recorded the talk on her phone from the back of the room, so the quality isn’t the best, but I’ve extracted the audio, cleaned it up a bit with Audacity, and made a (lightly edited) transcript using Descript. I’ve included both below. Let me know if you have any feedback or if you see things differently! When I polled the students at the beginning of the talk, they were generally more worried than excited about how AI will impact their future job prospects. I hope my talk injected some cause for optimism. 🙂

Transcript

[00:00] Now that you’re in high school, it’s natural to start asking “what’s the world going to want from you all?” as you get out of school the next five years, or what’s it going to want for you in the next 10 years, 25 years, 50 years, there’s a lot of future to be had. Right now is when you’re making choices about how to prepare yourself for that. Trying to figure out what’s the optimal thing to learn and do. Honestly, and I say this as a parent, it is really hard to figure out right now. I think that you’re entering one of the most uncertain times in our future that I can certainly think of an analogy for, for a number of reasons that we’ll talk about.

So the short answer to my talk is I don’t really have a clue what to tell you to do. And anybody who thinks that they do and gives you a confident answer about what to study or how to learn or what’s going to happen, what jobs robots are gonna take or not is full of shit (excuse me). But it’s important to know that we’re entering an era of pretty big uncertainty.

I do think, however, there’s a lot that we can think deeply about that will help you, I hope, ground your own decisions [01:00] about what you want to spend your time doing. And so that’s what I thought we could talk about today. And there’s a number of sources of uncertainty, but obviously AI I think is one of the biggest ones that’s the elephant in the room because it is the thing that could be displacing a lot of traditional knowledge work.

So I’m curious first, just to make sure I know what your own experience is, raise your hand if in the last month or so (i.e. recently enough) you’ve used one of these frontier large language models like ChatGPT or Claude or Gemini, or one of these things. Looks like not everyone, but almost everyone.

That’s good. How many of you are using a paid version that either you or your parents are paying for that’s higher power than even the normal free version? Not many. I mention it because there is a difference, and I think one of the best ways you can start to think about how to change your own approach to things with AI is to just use the latest and greatest stuff. So tell your parents, from me, I think it’s a wise investment.

How many of you that raised your hands have done more than just played with [02:00] it, but have actually say made something or built something or learned something that you’re genuinely proud of and that you probably wouldn’t have happened if you hadn’t used AI? A few of you. I certainly feel that way. I think you have to taste it to realize what’s behind it.

Okay, last question we’ll do, this as an A/B test, but there’s no wrong answer. How many of you, when you think about your own future, think that the presence of AI in the future of technology generally makes you more excited than worried? And how many of you are more worried than excited? Excited first. Yeah, very few. How many more worried than excited? Oh, okay, the majority. Interesting. Interesting. Okay, maybe I’ll be able to change your mind a little bit, because I’m fundamentally an optimist about this stuff.

Just a very quick bit about me, so you know where I’m coming from. So, as you can hear, I’m American born and raised in America. My connection to the German school comes from the fact that I had the good sense to marry a nice Austrian girl 20 years ago. And, [03:00] so I have kids here in sixth grade and ninth grade. So that’s why this is a personal topic for me as well as a topic of professional interest. I came out to California in the late nineties to study AI at Stanford, because actually even 30 years ago, it was clear to me then, and, obviously not to as many people as it is now, but that AI was going to completely change the world and usher in a sort of second industrial revolution and have a really profound effect on society, and I really wanted to be a part of that.

So this is something that I’ve been working on and thinking about ever since. I did academic research for a while. I’ve been in a number of startups. I was at Google for over a decade, including working on Google Assistant, which was the predecessor to Gemini. So it’s something I’ve done a fair amount with personally. But I also have thought a lot and read a lot about the sort of economic and societal implications of technology change in AI in particular, because I’m very concerned about [04:00] wealth inequality and making sure that we can not only create the future, but all share in it. So that’s where I’m coming from on this, and that’s what I’ll talk about today.

So why do I say that we’re in such an unpredictable time? From my vantage point, and I’m curious if you agree with this or not, I think we could be on the brink of a kind of technological utopia in your lifetime. That is the stuff of only science fiction until fairly recently. If you think about it, we could have abundant clean energy and self-driving cars and flying cars and limitless intelligence helping us to solve long unsolved problems in science and climate change and curing diseases and brain computer interfaces and exploring space and all that. These are all things that have just seemed completely out of reach for pretty much all of humanity, right? But you can just go around here within a 10 mile radius and see people working on all of these things with really credible paths to making, [05:00] continued progress, right?

So I think if we can stick around and if you can be part of ushering that in, it could be an absolutely wonderful thing there. Humans were subsistence hunter gatherers and farmers for almost all of human existence, right? And it’s really only since technology started accelerating that we’ve seen this sort of massive ability to bring everybody’s standard of living up and there’s no reason why we can’t go significantly further and we better hope we can.

Because we still have a lot of problems in the world. Despite all that optimism, we also, I don’t know about you, but if you read the news recently, it’s hard not to escape the fact that we might be on the brink of civil war and pandemics that we’re totally unprepared for, and runaway climate change and cyber attacks, or AI running amok or mad men running around with nuclear weapons. It’s pretty bleak out there, right? Let’s be honest. And that’s happening at the same time, right?

I think the more you study history, the more you come to appreciate the fact that the future is not foreordained. There’s a lot of fairly random things that can cause it to go down [06:00] very different paths. You are all going to have to navigate these, sort of real high highs and real low lows that we’re both staring at. A lot of possibilities. And on top of that, I think even if we are able to create a lot of technological abundance, there’s no guarantee that it’s going to be shared widely enough to maintain social cohesion. It could very well be that it continues the path that we’re on right now of concentrating wealth and power in a few hands and everybody else not sure how to participate. That’s something that you’ll have to deal with as well. We’ll talk a little bit about that

The last thing I would say is that it really was true, maybe not for your parents, but certainly for your grandparents, maybe even your parents, that you could think “I’m going to school for 10 or 20 years, I’m going to learn some skills, and then I’m gonna practice those skills for the next 40 or 50 years, maybe even at the same company the entire time, and that’s sort of it”. I just think those days are over for sure. I think the rate of change is accelerating itself so quickly [07:00] that you’re going to have to be lifelong learners. You’re going to have to change your identity, you’re going to have to do multiple things. You may live a lot longer, if we cure a lot of diseases. But you may also just have to reinvent yourself and think of school as more of a launching pad for more fundamental meta-skills like learning and curiosity and so forth. So I just think that makes it that much harder to predict, because it’s not just “oh, I’m gonna learn X and that’s gonna gimme a good job”. What the AI can do, and what you can do is going to change and the set of possibilities are going to change.

So anyway, that’s the backdrop I would think about. Hopefully I haven’t depressed you too much! I do think there’s good news. What I try to do when I think about my own kids’ future, when I think about the world more generally, is try to think more deeply about what is the way that you can add value to the world that is fairly universal? What is it that causes a job to be valuable in the first place in a more fundamental way? And also what are human universals that are unlikely to change even in the world of rapid [08:00] technological change? So that’s what I wanted to talk through with you, because that’s how I at least maintain some grounding and some optimism despite all this uncertainty.

So I don’t know if you’ve thought much about it, like why do some people get paid more than others when they’re both working hard and the same number of hours? It’s a huge range, living in Silicon Valley, there are people working super hard, making no money and there are people working super hard making unfathomable amounts of money and everything in between. Anybody have any ideas about what sort of fundamentally drives that? You can shout out if you want.

Economists will differ on this. I’m not sure there’s a well defined agreement on this. But the way I think about it, for what it’s worth, is I think there’s sort of two factors that are interrelated. One is how much supply versus how much demand exists for the skills and knowledge that you have and the connections you have and all in what you can offer. The other is how much [09:00] leverage do you get in the world based on the application of that effort.

So as an example, if you’re a Uber driver or you work on building construction or there’s lots of jobs like that where it can be hard work and you can spend a lot of hours doing it and you still don’t make a lot of money. The reason is lots of other people could also do that job without a ton of training. And you’re only driving one car at a time or fixing one house at a time or whatever. Even traditionally high paying jobs like doctors and lawyers and that sort of thing, they top out because you’re very well trained, so there are not that many people who can do what you can do, but you’re only fixing one person at a time or writing one contract at a time. Whereas the sort of amount of unbelievable wealth that’s been generated by Silicon Valley, or Hollywood for that matter, comes from the fact that not only do you have a lot of very specialized knowledge and a lot of very specialized connections and work and experience and so forth, but you’re changing, you all have cell phones in your pockets and you’re all watching Netflix and all these things, right? The amount of impact you can have from the leverage of your impact.

[10:00] It always made me a little humble at Google. I would write some code and push a change and it was like, “okay, a hundred million people are gonna wake up tomorrow and see that”, it was like, “I’d better not fuck it up”. It’s huge leverage. So I would just say in general, there are lots of different problems in the world that need solving that you can go out and decide you wanna get excited about, but I would ask you to keep in mind: where can I build up a set of, like I said, not just, book smarts, but like skills, practical, real world experience, people, like the whole set of what it is you amass? Where can you get on the right side of that supply and demand curve, and how can you find ever more leverage to use from that? Can you, instead of fixing something in one place, can you build a system? Can you teach other people to do it? It doesn’t have to just be traditional technology, but other ways to find leverage.

The problem of course, is when you’re trying to think through supply and demand, what AI is going to fundamentally do is add a lot more supply of a lot of different skills that previously didn’t exist, right? So like in my own [11:00] field of software engineering, AI is getting really good at writing code. So does that mean it’s no longer going to be valuable to write code because the supply and demand is going to get totally messed up? You can ask that question both in terms of what’s in demand and also where is the AI going to provide the supply. Of course, it’s hard to predict because the history of AI, which dates back now well over 50 years, is constantly a history of people saying: no computer can ever play chess or translate human languages or talk in a natural voice or compose new music or, like all these things people have said very confidently. And of course computers have steamrolled over all of those expectations. And they’re showing no signs of slowing.

Even in the real world, robotics are still lagging behind AI just in a digital format, but I think there’s a lot of progress going on right now with robotics and actually embodied cognition is one of the things that really unlocks, like having a good brain from AI makes a lot of robotics tasks possible in the real world that [12:00] work. I think the most salient example, maybe you don’t think about it this way, but if you’ve seen Waymo’s driving around in the streets here, those are robots, right? They’re machines that are using perception to navigate the real world and not bump into things and decide where they want to go. They’re robots. They just have wheels instead of legs. But it really does work and it really is a total game changer. And it is potentially displacing a huge amount of human labor, right? Actually, driving a car is one of the top professions in most places, right? And, totally unclear if that will still be a thing 10, 20 years from now.

So you have to be careful about what AI can and can’t do. But if I look at what I’ve gone through as a software engineer, just give you some personal experience, which is, on the one hand, it is really amazing how much good code AI can write and how much more it knows about esoteric aspects of different things that I don’t know about. But it’s also been amazing for me to see, as someone who uses it every day now for years, how much value I still have to provide on top of that. I don’t say that to brag. It really is just [13:00] fascinating to me that I wouldn’t have been able to tease apart the parts of my job that are replaceable by AI from the parts that aren’t. But there’s actually a ton of both.

For example, if you tell the AI: write a program that does X, Y, and Z, it might do a decent job of that, but it is not keeping up with what’s going on in the industry and what was just happening in the meetings and what’s the overall roadmap of the team and what’s our unique advantage and, oh, this thing you used, we were planning to tear that down next year anyway, so please don’t use it. There’s so much additional context that you as people bring to the puzzle. There’s also, for lack of a better word, the wisdom I’ve accrued as a multi-decade software engineer, which is there’s lots of ways to build things that are technically correct today but are unlikely to be as good in the future. They’re too brittle or they make too many of the wrong assumptions, or they’re making things unnecessarily complicated. One of the things you learn in any profession as you get older is those areas where you’ve done it a bunch of times and you know the sort of attractive pitfalls, right? [14:00] If you ever work with more junior people, that’s really the difference: they can be really smart, but they make a lot of “rookie mistakes” because they just haven’t been in it enough to really understand those second and third order effects. And you really see that too.

Maybe the coding systems will continue to get better at that, but unless they’re going to be sitting in all the meetings and chiming in on the slack threads and going to industry events and reading the news and so on? You could imagine a world in which my AI text editor is doing that, but it’s a pretty far stretch from the way things are today. Then putting even on top of that: I have to still wake up every day and decide I want to work on this and what I want to work on, and I have to motivate the team around me to get excited to work on stuff, and I have to get people outside to be excited about it or to get people to want to give me money to build my thing. All of that human stuff: building trust and getting people excited or having people feel like they need to work harder because they don’t want disappoint me, and vice versa, right? All that stuff is very human and it’s just not the same when [15:00] there’s blinking cursor on the screen, right?

So even as much of a optimist as I am about the continued progress of AI. I’m very aware of all the stuff that it’s not really on a path to do anytime soon. And that’s where I would suggest you anchor your own thoughts about how can I take an area. So in medicine, if you’re just being trained to read a radiology image or something like that, that narrow task might be very well doable by AI. But if you think about what it takes to be an effective doctor, let alone a medical researcher, let alone someone who helps change health policy in the country, etc., there’s so many layers above that that really need that human touch and I think will for a long time.

Now, like I said, I think you want to figure out how to be part of creating a lot of value, not just because you wanna make a lot of money, but also because you want to use your precious time as effectively as possible, right? I still think there is this big question, and I’ll be honest, I don’t know how we’re gonna fix it, but I encourage you to think about not just what can I do to [16:00] create a lot of value, but also how can I be part of shaping our future so that value gets shared broadly enough that the whole system is sustainable? Because, I think the history is pretty clear here that when wealth and power get too concentrated. a the economy slows down because there’s just not enough money circulating, there’s not enough people able to buy stuff. We already see today there’s a huge affordability crisis that most people have. Then at some point, political, societal stability breaks down too. And ultimately you get the French Revolution or things like that, right? And so we could absolutely be down that path where you’ve got a few trillionaires in their bunkers and we’re all using their stuff and we’re serfs to the AIs and we don’t want that world. They have solar panel arrays everywhere and so forth.

So it’s just interesting for you all to think about. It doesn’t mean you have to go into politics necessarily, but I do think that positive sum thinking is the secret to Silicon Valley. It’s basically “we can grow the pie”; you don’t have to lose for me to win. We can just make things work better. We can find new ways to distribute value. But I really encourage you to think about [17:00] that as well. what, any field you go into, what’s, where’s that opportunity to create that kind of positive sum game?

Alright, let me just wrap up here. I don’t wanna talk too long. The last thing I would encourage you to think about when you’re trying to figure out what you can do uniquely in the world is think more about what makes humans special. I talk a lot about this with my kids, but you really have to remember that as advanced as we are, we are monkeys, we’re social primates, right? That is not going to change any time soon. We are hardwired for social, understanding what each other thinks, building status in groups. That is a lot of what humans want and a lot of what drives us. And that is something that we really don’t want from machines. I think you can probably see that around you already.

I see a lot of FuĂźball Trikots here and there’s a lot of passion around that. There are robots that are pretty good at playing soccer already, but I just don’t think anybody is going to [18:00] be as passionate about a robot soccer player or a robot soccer team anytime soon as they are about the actual players that they care about. Why is that? It’s not because robots can’t be good at playing soccer. It’s because you don’t just care about the execution of soccer; you care about the human drama. Watching a player go from a rookie up or the team that you’ve stayed with through the highs and the lows or all the fans that are around you. It’s actually the human aspect.

This really hit home for me when, last summer we were lucky enough to be in Munich, when Taylor Swift played there in the Olympiastadion, and I don’t know how many of you have been to that place, there’s this whole Olympia Park around it and it was just completely full, probably 50,000 thousand people all dressed up with bracelets and singing songs and posters and shirts. Half them didn’t even have tickets for the show. They were just hanging out. And it was such an amazing energy and I was just thinking it’s so cool that people do this, but would anybody do this for an [19:00] AI-composed piece of music, no matter how good it is? Even if it was super catchy? No, of course not. It’s not just that it is Taylor up on stage, it’s the whole ethos of the culture that is around that and that’s what we care about, right? None of that is going to change even though AI will get better at writing music and maybe even really good music, but there’s still, people want to still see a human, right?

I don’t about you, but like I love live music, and I always think it’s funny, you can go see somebody play live even in like a coffee shop or whatever, not a big stage, and it’s still way more enjoyable than listening to recorded music through the speakers, even if the person who recorded that music is more talented than the person who’s playing it. It’s weird, right? Because if the goal was just, I want the best music possible to enter my eardrums, like the sort of very narrow idea of the work, then you wouldn’t ever have the need for live musicians, right? Because we already recorded it once and we can spread it everywhere. And yet I think if anything, the opposite is happening, which is we really enjoy that [20:00] live communal experience. We really enjoy seeing the person play. I love playing music myself too. Even though I’m not very good, it’s incredibly fulfilling.

So I think that’s the human universal stuff that I think is not going to change. And that’s to say nothing of, if you’re actually caring physically or emotionally for a loved one, a, child, a parent. Again, I think robots will have a huge impact on healthcare, but I think there’s nothing that’s going to replace the human touch; the feeling that a teacher really believed in you and made you feel like you wanted to be more because they inspired you or that sort of stuff. Again, we’re monkeys. So embrace your inner, monkey! Don’t, don’t run away from it.

The last thing I would say about that is that, I don’t know if you’re familiar with the psychology literature on human drive? You can read Daniel Pink or one of these people, but basically the thing that motivates humans is a combination of desire for autonomy, mastery, and purpose. Have any of you heard that framework before? So autonomy is you want to be able to [21:00] be your own boss or march to the beat of your own drum; mastery is you want to feel like you’re getting really good at something; and purpose is you want to feel like the work you’re doing actually matters to someone. Pretty much everything humans do is trying to get one of those things. Actually, making a lot of money is not one of them. You have to make enough money that it’s not a problem, but after that, more money isn’t nearly as motivating as these other factors.

You see that in Silicon Valley, right? You see a lot of people who have more money than they ever spend, but are working really, hard. Why is that? Why aren’t you just on a beach? But laying on a beach doesn’t give you a lot of mastery and purpose, right? Whereas getting back into building something new, or even back to the street musician example, why do people become starving artists and make no money, but play music or make art or whatever? It’s because it really does fulfill a lot of those drives, and they’re willing to do it despite the fact that it doesn’t really pay well. So if, you strive for something that actually lights your candle, gets you excited, makes you want to go push and do things to make some change in the world you want to see, I think you can start to optimize for that.

And, maybe one of the good sides [22:00] is maybe it’ll even become less important that it creates a lot of economic value. Because if we are successful at creating overall abundance and sharing it widely, then it probably won’t matter as much that what you’re doing personally is creating a lot of value. A lot of people who make a lot of money really hate their jobs, right? They really don’t love it, in fact, a lot of them have a “side hustle”, or they’re like a 12th level Orc on World Warcraft or whatever, because that’s the thing that actually fulfills them, not the work. So, as much as possible, try to find work you can do that you can throw yourself into for a long time. It’s just such a hedge against what will and won’t be the “hot job” at the time. Does that make sense?

To sum it up, if you take away one thing from this, it’s that I think you should all know that the future is very uncertain, but that there’s a lot of promise and if you can just try to, maybe it sounds cliche, but cultivate a sense of agency, have a growth mindset, remember [23:00] that there’s lots of things in the world that should and could be different, and if you can just be curious about them and then have the courage to learn things and try things and fail and iterate and just get out there and try to make a difference. That set of skills will always be valuable no matter what the set of “building blocks” that are out there are. Remember, everybody else is going to have to same building blocks as you. The reason why having ChatGPT do your homework is not a good idea is not just because this is the way school is set up. It’s because everybody else can press that same “yes” button that you could have, right? So you’re not differentiating yourself in any way if you’re just the like passive conduit. What you want to do is figure out how to use that technology to do something you couldn’t do before, or to do something better or faster, or to have or be more ambitious because you’re like, “I have no idea how to do this”, but I can learn it on YouTube and I can try it with this. That’s, to me, the positive way of thinking about it. That’s where you’re still providing that unique set of supply and demand skills.

So, be curious. Have a growth mindset. Care about other people, right? Don’t be a robot yourself. Humans care about [24:00] humans. The more you understand, not just technology (I do think it’s very important to stay up on the forefront of technology), but I also think it’s really important to understand what drives art and culture and fashion and empathy and all these things.

And schnall dich an, because it’s gonna be a wild ride, but I think it could potentially be a really great one. I hope that’s helpful. Thanks.

Leaders in Tech podcast appearance (part 2)

After my initial conversation on the Leaders in Tech podcast, the host asked me to come back and follow-up with more of my thoughts on AI and the future of work and what we humans will or won’t still want or need to do in the future. We discussed what we can learn from the study of the human brain, and in particular how the pattern-matching cerebral cortex is distinct from the goal-oriented “old brain”, the latter of which is still largely missing from the AI models we’re building. While a lot of knowledge work will undoubtedly be augmented if not replaced by AI over time, we reflected on how much of being an effective leader in tech (or in most professions) still comes down to innately human characteristics of passion, empathy, group coordination, and so on, as well as how we will continue to be driven by work that affords us autonomy, mastery, and purpose, even if it becomes disconnected from how we provide for our basic needs.

Throughout the interview, you will hear why I am still fundamentally optimistic about “team human” and our potential to thrive in a world of technological abundance, which AI can help us usher in (if we don’t mess things up in the meantime, of course!).

Leaders in Tech podcast appearance

I was recently a guest on the Leaders in Tech podcast. We covered a lot of ground, from my childhood and how I got interested in tech and AI, to many lessons I learned working in startups (first Plaxo, now Triller) and inside big companies (Google+, Google Photos, Google Assistant). In particular, the conversation focused on the advantages startups have when it comes to driving innovation, and why, despite their advantages in terms of resources and distribution, it’s hard to get the same results inside larger organizations. We finished with a discussion of how AI is likely to impact the careers of software engineers (my bet is it will remain more of an amplifier than a replacement for years to come).

I think this is one of the best summaries of my experience and thoughts on Silicon Valley and entrepreneurship that I’ve managed to capture. I hope you’ll find it useful and would love to hear your feedback!

Starting a new conversation…

Google AssistantI’m excited to share that I’ve recently joined the Google Assistant team! Like a lot of people (including our CEO), I’ve become convinced that natural language conversation represents the future of how we’ll primarily interact with computers and the myriad smart devices that will soon proliferate around us. This new “UI” will be personalized based on both knowledge of you and your history of interactions. It will also be proactive (reaching out to you with pertinent questions and updates) as well as responsive. And it will execute tasks across multiple, interconnected services on your behalf.

Which is to say: it’ll be a lot different than how we work with computers today. And it promises to be a lot better, too — if we can get it right.

I’ve been fascinated by interacting with my Google Home (which I’ve had early access to for a while). It highlights both the challenges and opportunities of this new conversational modality, and it surprises in equal measure with how far we’ve come and how far we still have to go. For instance, my 5 year old daughter walked into our living room the other day and proclaimed, “Ok Google, play some Lady Gaga”, then started dancing to the music that immediately began playing. Think about that: She would never have been able to accomplish that task with a traditional desktop/mobile app, nor would I have been able to help her as quickly as she was able to help herself. She didn’t have to be unnaturally precise (e.g. select a particular track or album), and it was an enormously empowering interaction with technology overall. It feels like “how things should work”.

I’ve had countless similar “wow moments” asking Google questions about the world, my own upcoming flights and schedule, or streamlining tasks like playing music or showing some recent photos I took on our TV to the grandparents. But for all the magic Google can deliver already, this is still very clearly early days. The dialogs are often too fragile and require too much custom crafting by too many Google engineers in a way that clearly won’t scale to the ambitions of the medium (and the team). There’s not yet much deep learning or true language understanding under the hood. And only recently has there even been a focused effort to build The Assistant, instead of just “voice enabling” individual products here and there. The industry as a whole is still only starting to figure out how “chatbots” and “conversational commerce” and so on should work.

Which is why it seems like an ideal time to get involved–we can tell “there’s a there there”, but we also still need many foundational innovations to realize that potential.

On a personal level, this change also represents an opportunity to get my career “back on track” after a wonderful decade+ diversion into the emerging world of social networking. I actually came to Stanford in the late nineties to study Natural Language Processing with the aim of empowering ordinary users with the superpower of Artificial Intelligence, and even published several papers and helped build Stanford’s first Java NLP libraries while earning my BS and MS. I originally joined Plaxo, an early pioneer of social networking, to build an NLP engine that automatically parsed contact info out of emails (e.g. “Oh, that’s my old address. I’ve since moved to 123 Fake St.”), but eight years later, I was still there, serving by then as its CTO and trying to open up the social web that had sprung up around us as the new way to stay connected. That in turn led me to join Google in 2010 as part of the founding team that created Google+, and I’ve been at Google ever since. But now I’ll actually be working on NLP again, and I have a feeling my years advocating for user-centric identity and data portability across services will come in handy too!

I’m uncomfortably excited to be starting this new chapter of my career. If you think about all the exhilarating potential surrounding AI these days, realize that a surface like Google Assistant is where you’re most likely to see it show up. One of the senior-most engineers on the team remarked to me that, “I’m sure a full Turing Test will be part of our standard testing procedure before long,” and I think he was only half joking. If you’re building similar things in this space, or have ideas about what we should prioritize building in the Google assistant or how you’d like it to integrate with your service, please let me know. I’m ready to learn!

The paper that would not die

Sources of Success for Boosted Wrapper Induction
Journal of Machine Learning Research, Volume 5
Written October 2001, published December 2004

Download PDF (29 pages)

Download PPT (900KB; presentation at Stanford’s Seminar for Computational Learning and Adaptation)

I co-wrote this paper during the first summer I started doing NLP research, but it didn’t see the light of day until a year after I’d finished my Master’s degree. Yeesh!

It all started when I decided to spend the Summer of 2001 (between my junior and senior years at Stanford) doing research at UC San Diego with Charles Elkan. I’d met Charles through my dad on an earlier visit to UCSD, and his research exhibited exactly the mix of drive for deep understanding and desire to solve real-world problems that I was looking for. I was also working at the time for a startup called MedExpert that was using AI to help provide personalized medical advice. Since one of the major challenges was digesting the staggering volume of medical literature, MedExpert agreed to fund my summer research in information extraction. So I joined the UCSD AI lab for the summer and started working on tools for extracting information from text, a field that I would end up devoting most of my subsequent research to in one form or another.

As it happened, one of Charles’s PhD students, Dave Kauchak, was also working on information extraction, and he had recently gotten interested in a technique called Boosted Wrapper Induction. So Dave, Charles, and I ended up writing a lengthy paper that analyzed how BWI worked and how to improve it, including some specific work on medical literature using data from Mark Craven. By the end of the summer we had some interesting results, a well-written paper (or so I thought), and I was looking forward to being a published author.

Then the fun began. We submitted the paper for publication in an AI journal (it was too long to be a conference paper) and it got rejected, but with a request to re-submit it once we had made a list of changes. Many of the changes seemed to be more about style than substance, but we decided to make them anyway, and in the process we ran some additional experiments to shore up any perceived weaknesses (by this time I was back at Stanford and Dave was TAing classes, so re-running old research was not at the top of our wish list). Finally we submitted our revised paper to a new set of reviewers, who came back with a different set of issues they felt we had to fix first.

To make a long story short, we kept fiddling with it until finally, long after I had stopped personally working on this paper (and NLP altogether, for that matter), I got an e-mail from Dave saying the paper had finally been accepted, and would be published in the highly respected Journal of Machine Learning Research. It was hard to believe, but sure enough at the end of 2004–more than three years since we first wrote the paper–it finally saw the light of day. It was the paper that would not die.

Charles had long since published an earlier version of the paper as a technical report, so at least our work was able to have a bit more timely of an impact while it was “in the machine”. I’m glad it finally did get published, and I know that academic journals are notoriously slow, but given how fast the fronteir of computer science and NLP are moving, waiting 3+ years to release a paper is almost comical. I can’t wait until this fall to get the new issue and find out what Dave did the following summer. :p

A nifty NLP paper that never made it

Conditional Estimation of HMMs for Information Extraction
Submitted to ACL 2003
Sapporo, Japan
July 2003

Download PDF (8 pages)

Download PPT (500KB; presentation to NLP group, including work discussed in this paper)

A conditionally-trained HMM in a toy domainThis is another paper I wrote that didn’t get accepted for publication. Like my character-level paper, it was interesting and useful but not well targeted to the mindset and appetite of the academic NLP community. Also like my other paper, the work here ended up helping us build our CoNLL named-entity recognition model, which performed quite well and became a well-cited paper. If for no other reason, this paper is worth looking at because it contains a number of neat diagrams and graphs (as well as some fancy math that I can barely comprehend any more, heh).

One reason why I think this paper failed to find acceptance is that it wasn’t trying to get a high-score in extraction accuracy. Rather it was trying to use smaller models and simpler data to gain a deeper understanding of what’s working well and what’s not. When you build a giant HMM and run it on 1000 pages of text, it does so-so and there’s not a lot you can learn about what went wrong. It’s way too complex and detailed to look at and grok what it did and didn’t learn. Our approach was to start with a highly restricted toy domain and minimal model so we could see exactly what was going on and test various hypotheses. We then scaled the models up slightly to show that the results held in the real world, but we never tried to beat the state-of-the-art numbers. Sadly, it’s a lot harder to get a paper published when your final numbers aren’t competitive, even if the paper contributes some useful knowledge in the process.

It seems both odd and unfortunate to me that academic NLP, which is supposedly doing basic scientific research for the long-term interest, is culturally focused more on engineering and tweaking systems that can boost the numbers by a few percent than by really trying to understand what’s going on under the covers. After all, most of these systems aren’t close to human-level performance, and the current generation of technology is unlikely to get us there, so just doing a little better is a bit like climbing a tree to get to the moon (to quote Hubert Dreyfus, who famously said as much about the field of AI in general).

If companies are trying to use AI in the real-world, their interest is performance first, understanding second (make it work). But in academia, it should be just the opposite–careful study of techniqus and investigation of hypotheses with the aim of making breakthroughs in understanding today that will lead to high-performance systems in the future. But I guess the reality is that it’s much easier (in any discipline) to pick a metric and compete for the high score. (The race for a 3.6GHz processor to out-do the 3.5GHz competition in consumer desktop computers comes to mind, when both computers are severely bottlenecked on disk-IO and memory size and rarely stress the CPU in either case. Ok, that was either a lucid metaphor or complete jibberish, depending on you are. :))

In any event, I enjoyed doing this research, and I’m proud of the paper we wrote.

Information Extraction for the Semantic Web

Finding Educational Resources on the Web: Exploiting Automatic Extraction of Metadata
Workshop on Adaptive Text Extraction and Mining
Cavtat-Dubrovnik, Croatia
Sempetmber 22, 2003

Download PDF (4 pages)

The Semantic Web is a great idea: expose all of the information on the web in a machine-readable format, and intelligent agents will the be able to read it and act on your behalf (“Computer: When can I fly to San Diego? Where can I stay that has a hot tub? Ok, book it and add it to my calendar”). There’s just one problem: the humans writing web pages are writing them for other humans, and no one is labeling them for computers. (A notable exception are blogs, like this one, whose authoring tools also generate machine-readable versions in RSS or Atom that can be consumed by sites like Bloglines. In a way, Bloglines is one of the few sites making good on the vision of the Semantic Web.)

What do people do when they’re looking for a piece of information, say a list of syllabi for NLP classes? There’s no database that lists that type of information in a structured and curated form. Rather, there are a set of web pages that describe these classes, and they’re all a bit different. But most of them contain similar information–the title of the class, the course number, the professor, and so on. So, in a way, these pages do constitute a database of information, it just takes more work to access it.

That’s where NLP comes in. One of the ways we were using information extraction in the Stanford NLP group was to automatically extract structured information from web pages and represent it in a semantic web format like DAML+OIL and RDF. The idea is that you send your intelligent agent out to the web (“find me a bunch of NLP classes”) and when it comes across a page that looks promising, it first looks for semantic web markup. If it can’t find any (which will usually be the case for the forseeable future), it tries running the information extraction engine on the site to pull out the relevant data anyway. If the site allows it, it could then write that data back in a machine-readable format so the web becomes semantically richer the more agents go looking for information.

Specifically, we built a plug-in to the protege tool developed by the Stanford Medical Informatics group. Protege is a Java-based tool for creating and managing ontologies (a form of knowledge representation used by the semantic web). Our plug-in let you load a web page, run our information extraction tools on it, and add the extracted semantic information to your ontology. You could build up a collection of general-purpose information extracton tools (either hand-built or trained from data) and then use them as you found web pages you wanted to annotate.

Cynthia Thompson, a visiting professor for the year, used this system to find and extract information about educational materials on the web as part of the Edutella project. It ended up working well, and this paper was accepted to the Workshop on Adaptive Text Extraction and Mining as part of the annual European Conference on Machine Learning (ECML). I declined the offer to go to Croatia for the conference (though I’m sure it would have been a memorable experience), but I’m glad that my work contributed to this project.

My most famous NLP paper (CoNLL-03)

Named Entity Recognition with Character-Level Models
HLT-NAACL CoNLL-03 Shared Task
Edmonton, Canada
June 1, 2003

Download PDF (4 pages)

Download PPT (3.8MB; presentation at CoNLL-03)

Every year that Conference on Computational Natural Language Learning (CoNLL) has a “shared task” where they define a specific problem to solve, provide a standard data set to train your models on, and then host a competition for researchers to see who can get the best score. In 2003 the shared task was named-entity recognition (labeling person, place, and organization names in free text) with the twist that they were going to run the final models on a foreign language that wouldn’t be disclosed until the day of the competition. This meant that your model had to be flexible enough to learn from training data in a language it had never seen before (and thus you couldn’t hard-code English rules like “CEO of X” –> “X is an organization”).

Even though my first paper on character-level models got rejected, we kept working on it in the Stanford NLP group because we knew we were on to something. Since one of the major strengths of the model was its ability to distinguish different types of proper names based on their composition (i.e. it recognized that people’s names and company names usually look different), this seemed like an ideal task in which it could shine (see my master’s thesis for more on this work). By this time, I’d started working with Dan Klein, and he was able to take my model to the next level by combining it with a discriminatively trained maximum-entropy sequence model that allowed us to try lots of different character-level features without worrying about violating independence assumptions (a common problem with generative models like my original version). Dan’s also just brilliant and relentless when it comes to analyzing the errors a model is making and then iteratively refining it to perform better and better. The final piece of the puzzle came from my HMM work with Huy Nguyen, which let us combine segmentation (finding the boundaries of proper names in text) and classification (figuring out which type of proper name it is) into a single model.

Our paper was accepted (yay!) and Dan and I flew to Canada to present our work. This was my first NLP conference and it was awesome to meet all these famous researchers whom I’d previously read and learned from. Luckily for me, Dan was just about to finish his PhD, and he was actively being courted by the top NLP programs, so by sticking with him I quickly met most of the important people in the field. Statistical NLP attracts a fascinating mix of people with strong math backgrounds, interest in language, and a passion for empirical (data-driven) research, so this was an amazing group of people to interact with.

On the last day of the conference (CoNLL was held inside HLT-NAACL, which were two larger NLP conferences that had also merged), the big day had come at last. My first presentation as an NLP researcher (Dan let me give the talk on behalf of our team), and the announcement of the competition results. There were 16 entries in the competition. In English (the language we had been given ahead of time), our model got the 3rd highest score; in German (the secret language), our model came in 2nd, though the difference between our model and the one in 1st place was not statistically significant. In other words, had the test data been slightly different, we might easily have had the highest score.

Doing so well was certainly gratifying, but what made us even happier was the fact that our model was far simpler and purer than most in the competition. For instance, the model that got first place in both languages was itself a combination of four separate classifiers, and in addition to the training data provided by the conference, it also used a large external list of known person, place, and organizaton names (called a gazetteer). While piling so much on certainly helped eek out a slightly higher score, it also makes it harder to learn any general results about what pieces contributed and how that might be applied in the future.

In contrast, our model was almost exclusively a demonstration of the valuable information contained in character-level features. Despite leaving out many of the bells-and-whistles used by other systems, our model performed well because we gave it good features and let it combine them well. As a wise man once said, “let the data do the talking”. Perhaps because of the simplicity of our model and its novel use of character features, our paper has been widely cited, and is certainly the most recognized piece of research I did while at Stanford. It makes me smile because the core of the work never got accepted for publication, but it managed to live on and make an impact regardless.

My first NLP research paper

Classifying Unknown Proper Noun Phrases Without Context
Technical Report dbpubs/2002-46
Stanford University
April 9, 2002

Download PDF (9 pages)

Download PPT (1.3MB; presentation of the paper to the NLP group)

As I describe in my post about my master’s thesis, I started doing research in Natural Language Processing after Chris Manning, the professor that taught my NLP class at Stanford, asked me to further develop the work I did for my class project. He helped me clean up my model, suggested some improvements, and taught me the official way to write and style a professional academic paper (I narrowly avoided having to write it in LaTeX!). I was proud of the final paper, but it wasn’t accepted (I believe we submitted it to EMNLP 02).

This was the start of a series of lessons I learned at Stanford about the difference between what I personally found interesting (and how I wanted to explain it) and what the academic establishment (that decides what papers are published by peer review) thought the rules and conventions had to be for “serious academic work”. While I got better at “playing the game” during my time at Stanford–and to be fair, some of it was actually good and helpful in terms of how to be precise, avoid overstating results, and so on–I still feel that the academic community has lost sight of their original aspirations in some important ways.

At its best, academic research embarks on grand challenges that will take many years to accomplish but whose results will change society in profound ways. It’s a long-term investment for a long-term gain. NLP has no shortage of these lofty goals, including the ability to carry on a natural conversation with your computer, high quality machine-translation of text in foreign languages, the ability to automatically summarize large quantities of text, and so on. But in practice I have found that in most of these areas, the sub-community that is ostensibly working one of these problems has actually constructed its own version of the problem, along with its own notions of what’s important and what isn’t, that doesn’t always ground out in the real world at the end of the day. This limits progress when work that could contribute to the original goal is not seen as important in the current academic formulation. And since, in most cases, the final challenge is not yet solvable, it’s often difficult to offer empirical counter-evidence to the opinions of the establishment as to whether a piece of work will or will not end up making an important difference.

I found this particularly vexing because my intuition is driven strongly by playing with a system, noting its current shortcomings, and then devising clever ways to overcome them. Some of the shortcomings I perceived were not considered shortcomings in the academic version of these challenges, and thus my interest in improving those aspects fell largely on deaf ears.

For instance, I did a fair amount of work in information extraction, which is about extracting structured information from free text (e.g. finding the title, author, and price of a book on an amazon web page or determining which company bought which other one and for how much in a Reuters news article). The academic formulation of this problem is to run your system fully autonomously over a collection of pages, and your score is based on how many mistakes you make. There are two kinds of mistakes–extracting the wrong piece of information, or not extracting anything when you should have–and both are usually counted as equally bad (the main score used in papers is F1, which is the harmonic average of precision and recall, which measure those two types of errors respectively). If your paper doesn’t show a competitive F1, it’s difficult to convince the community that you’re advancing the state-of-the-art, and thus it’s difficult to get it published.

However, in many real-world applications, the computer is not being run completely autonomously, and mistakes and omissions are not equally costly. In fact, if you’re trying to construct a high-quality database of information starting from free text, I’d say the general rule is that people are ultimately responsible for creating the output (the computer program is a means to that end), and that the real challenge is to see how much text you can automatically extract given that what you do extract has to be extremely high quality. In most cases, returning garbage data is much worse than not being able to cover every piece of information possible, and if humans can clean up the computer’s output, they will definitely want to do so. Thus the real-world challenges are maximizing recall at a fixed high-level of precision (not maximizing F1) and accurately estimating confidence scores for each piece of information extracted (so the human can focus on just cleaning up the tricky parts), neither of which fit cleanly into the academic conception of the problem. And this is to say nothing about how quickly or robustly the systems can process the information they’re extracting, which would clearly also be of utmost importance in a functioning system.

I witnessed firsthand this difference between the problem academics are trying to solved and the solution that real applications need when I started working for Plaxo. A core component of the original system was the ability to let people e-mail you their current contact info (either in free text, like “hey, i got a new cell phone…” or in the signature blocks at the bottom of messages) and automatically extract that information and stick it in your address book. This would clearly be very useful if it worked well (the status quo is you have to copy-and-paste it all manually, and as a result, most people just leave that information sitting in e-mail), and it clearly fits the real-world description above (sticking garbage in your address book is unaccepatble, whereas failing to extract 100% of the info is still strictly better than not doing anything). None of the academic systems being worked on had a chance of doing a good job at this problem, and so I had to write a custom solution involving a lot of complicated regular expressions and other pattern-matching code. My system ended up working very well–and very quickly (it could process a typical message in under 50 msec, whereas most academic systems are a “start it running and then go for coffee” kind of affair)–and developing it required a lot of clever ideas, but it was certainly nothing I could get an academic paper published about.

The irony cuts both ways–when I tried to solve the real problem, I couldn’t get published, but the work that was published didn’t help. And yet the academic community could surely do a much better job of solving the real problem if only they hadn’t decided it wasn’t the problem they were interested in. I only bring this up because I am a big believer in the power and potential of academic research, and I still optimistically hope that its impact could be that much greater if its goals were more closely aligned with the ultimate problems they’re trying to solve. By bridging the gap between academia and companies, both should be able to benefit tremendously.

If you’ve read this far in the hope of knowing more about the contents of my first NLP paper, I’m sorry to say it has nothing to do with information extraction, and certainly nothing to do with the academic/real-world divide. But it’s a neat paper (and probably shorter than this blog post!) and despite its not being published, the work it describes ended up influencing other work that I and people at the Stanford NLP group did, some of which did end up gaining a fair bit of notoriety in academic circles.

« Older posts

© 2025 Joseph Smarr

Theme by Anders NorenUp ↑