Sleepwalkers at CES

which I don't normally play. I think so consciously. I was thinking about what Tristan Harris talked about in the first season Sleepwalkers. You know he was that former googler who told us that Instagram is actually supposed to feel a lot like slot machines. Well, that's right. Tristan studied at the Stanford Persuasion Lab and told us about how casino architecture has influenced the development of highly addictive type

products like Instagram. So it's interesting for me to actually see Vegas and the bright lights and the impossibility of escape firsthand, not to mention the replicas of the Empire State Building, the Canals of Venice, the Colisseum of Rome, and you know, I was lucky enough to see the Seattle Space Needle for the first time. I just didn't know that it was in Las Vegas. But that's not

why we were there. We were there for CEES, the Consumer Electronics Show, and this episode we're actually going to talk about some of the coolest things we saw there, but we're going to focus more on the innovations that are at the intersection of technology and humanity rather than

talk about you know, infamous toilet paper dispensers. One of the big reasons we went is because we were invited by WaveMaker, which is an agency part of w p P, to do an interview on stage, a live podcast, so to speak with Matt and on a hand, who is head of product at ARC Publishing and ARC Publishing is

part of the Washington Post. Yeah, and ARC is also an interesting case of AI and action because they're forward thinking in terms of increasing the visibility of content through personalization and optimizing everything from headlines to photo selection, all using machine learning, and those are things that really matter for journalists and readers. Yeah, and this use of AI stands out to me because it provides a solution to real problem. How do you get eyeballs on the right

content when there's just so much. That said, the issue of personalization does also raise questions about what happens when machines start to know us better than we know ourselves. Not to mention, what are the appropriate limits of how companies use AI and dature about us. Yeah, AI can definitely streamline processes by detecting patterns that you know, human beings cannot see, or it can allow you to do things at the scale like tag hundreds of thousands of

articles that again, human beings just cannot do. So greater efficiency is on one side of the spectrum and extremely attractive to people, but on the other side you have issues of taking humans out of the loop, like the black box problem and authenticity in a world of deep fakes. So a question for businesses and users of technology is sort of when does AI add to our experience and when does it maybe hold us back or take advantage

of us. You know, for example, from seeing news stories that we should see, but maybe the algorithm doesn't think we want to see it or that we won't click on it. Right in the old days when everyone received a print use of paper on their doorstep, everyone has

the same front page and the same headlines. Nowadays, when you log onto a news website or onto social media, everybody has a different version of the world, and that is obviously positive for driving engagement, but may not be so positive in terms of having conversations with the same facts about the same stories. Equally, we have to ask do we want articles where the headline has been written by an algorithm or do we prefer headlines written by

a person. And that's something we talked about with Matt because ARC actually tested the headline writing technology. Let's talk to Matt seriously. Let's cut to the chase. Are really came out of a collaboration trying to better understand what actual journalists needed. Can you talk a little bit more. Yeah, at the very beginning, you know, we were just trying to solve problems for ourselves seven or eight years ago.

You know, we knew and he had to make some pretty fundamental transformation to the post and to really prepare ourselves for the digital future. We didn't have the right tools to do it, and we couldn't really find the right tools on the market either. What we did was spent a lot of time with the journalists and the editors trying to figure out what it was that make their lives easier. It's trying to figure out how do you make journalists work better, how can they publish faster?

What are the little things you can do inside of a product to make it easier for them to write stories or publish From there? About four years ago is when we started evolving it into a commercial offering. Today we're running hundreds of websites around the world. We're in about twenty different countries. We're running companies like BP their

internal communications as well as some of their marketing. We're running large broadcasters and all their live video and body and of course we're still running a lot of newspapers and news publishers like The Post and many others around the world, Lucky and publishing. You know that AI and artificial intelligence are made in headlines, and there was a story in the Financial Times last year we said fort of AI startups use no AI whatsoever. So I bet

it's probably higher. But when we talk about using AI, or when you talk about using AI, what do we

actually mean? So it can span the range of technologies from something like machine learning, which is basically a way to use algorithms, to take large sets of data and either uncovered patterns in it or try to model a way to predict a certain outcome to technologies like computer vision, which you can use to look at images or video and extract information about them by recognizing patterns and trying to identify objects inside of them, and so a lot

of those technologies, then when you put them together, you can form some really interesting workflows that you know in the past you might have had to use humans to do that, you can actually do much more simple automatically. Was there a particular business challenge or challenge at the Washington Post that you couldn't have solved if you hadn't been using AI. Any story that we write on Washington Post, we're mapping to a set of two or three hundred topics.

Maybe an example of one of those might be like congressional policy or narcotics crime. What you're trying to do is say, if I look at all this content, I'm not just pulling specific words out of it. I'm actually trying to figure out what is this content about, what is the fundamental concept of this content. So you pick a set of articles, let's say a hundred thousand news articles in the case of this example for the post, and at first you use humans it's called micro labor

to do this training set. And the goal is you're building an algorithm based on a set of real data. And so the humans are going there and saying this article, yeah, this is about congressional policy. Why because I know it is, I read it, that's what it's about. This one's about

narcotics time, and this one's about soccer. And so you train all these articles against that algorithm until finally the algorithm is basically sufficiently advanced to predict a new article that you put into it and determine an outcome with the same high probability of success that you were able to with humans training it. Now, every time a journalist saves or publishes a story, we're able to parse over

all the content inside that story. Then we can predict the strength at which it's likely to belong to that topic. How do you create a better user experience, in your case, news experience for an individual or consumer With that metadat you can do a lot of interesting things. We can figure out that hey, this is something that they're interested in reading, perhaps they'd like to read more, and it

actually serves the signal into our recommendation algorithms. From your perspective, where can businesses sort of harness the power of machine learning to really hone in on who their customer is and what that customer wants. We want to deliver more content to our readers who want to help them find more content that we've created. We have about nine journalists at the Washington Post. We write something like, you know, three or four hundred original stories a day, so there's

a lot of content there. To get readers to all that different content and have them continue moving through your content you spend a lot of money to produce is really challenging, and so that's a great use case for personalization. But where you can make it really come alive is by having more sophisticated metadata, more sophisticated information about that content. It's more likely to bring readers to it, and so that's where these machine learning models really come in handy.

I think part of what's fun about this conversation is there's a lot of cases out there where average users, you know, they imagine they see something like that. You see the boots on Instagram and you think, oh my god, like these companies must crazy, like you know, indiscernible for magic, right, like, there must be some crazy model out there doing this,

and perhaps there is. But in a lot of ways, you know, your users aren't necessarily as aware of the advertising ecosystem, the data ecosystem, and how these things tied together between platforms and sites, and I think, as like industry professionals, we always kind of underestimate that fact. And so the net effect is that users are completely surprised by this. They think you must be doing something completely unheard of to achieve it, when in fact, you know

it could be really simple data sharing. And so the reason I think that's important is then when you do build technologies that actually utilize some of these more sophisticated methods to build data sets, you have to be aware

that your users. You know, first of all, your users aren't gonna necessarily anticipate the outcomes that you can create, and if you don't do a good job on the product side of making sure that you really think through the use case and how you're leveraging technology to solve it,

you can generate unexpected outcomes. You know, there was the example of a retailer who produced advertising flyers that were able to predict folks who are pregnant, right, even if some of those folks didn't necessarily know that themselves yet or hadn't shared it with with their family or their spouses. And so that was a case really of both the company and the consumer being shocked by outcomes we're generating

exactly right. I mean, the you know, the algorithm doesn't do anything magic, but that's a case of you know, putting together in that case, like a marketing program, where you don't really think through what's the possible data that this could produce and what are my users? What do they already know about this data? You know, you need to think really hard about your users and what they want and what they're trying to achieve, and what the

dangers are and leveraging this technology. It's no different than in that way than any previous technology solutions you might have used to build a great product for people, and

it can be misused just as easily. Funny enough, the first episode of Sleepwalkers season one open with a story of Washington Post employee Gillian Brockell, who, to your point about pregnancy and data stuffered a miscarriage but continue to receive targeted ads for pregnancy goods after a miscarriage, and she wrote this openless is the technology companies saying please stop targeting me. But that raised a big question for us,

which is what happens when the algorithms go wrong? Yeah, I'd almost be more specific with the way that you say that, and like the algorithm didn't go wrong, right, but like the implementation of it and the product that they built around it did because it wasn't really correctly conceived. And we have to make sure that like what you're trying to do automatically fits really well with what your users are trying to accomplish, doesn't happen in a way

that's not expected. Is a well designed product, you know, So in that specific case, yeah, I mean it always starts and ends with kind of good product design. If you're not doing that, just like any other tool, you can misuse it. One of the other things we did on the show was we used a language generator to can't with pickup lines based on a data set of all none of them were actually I don't have kind of things like you are a thing and I love you, you know, which is now the name of the book

by the woman who. Yeah, woman Jell Shane her wrote a book about it, and then she also did these things like AI recipes, like one was for chocolate chocolate chocolate chicken cake. So there is funny things and Shakespeare on it, and I didn't revealed two things. One is when you turn these deep learning algorithms onto big data sets, they reveal passions you might not necessarily be aware of, like with a lot of chicken and quite a lot of chocolates. On the other hand, like these were clearly

not something human would ever make. So how do you think about the line between doing fun things in AI and doing stuff which is valuable for business and also not getting lost in the uncanny valley. So a good example of this for instances. We spent some time at the Post trying to build a headline generation algorithm we

could automatically create headlines for stories. And you know, the idea I think at the beginning wasn't necessarily that, you know, journalists are never read headlines again, but we'd be able to create some alternative headlines in different ways to think about a story. Our intention was, let's see if we can come up with something so that we can create several different variants of a headline. Part of our software

platform we include content testing framework. So one of the things that we can do is say, for a given story, let's have three different headlines for it, Let's run a test as soon as it publishest of the audience is going to get each variant, and then as people start to click one more than the other, we're gonna shift the burden of traffic to the most successful variant. And that,

I would them, by itself works really well. If you know, folks in the audience here were to look at the homepage of our site right now, there's probably two or three stories that are running those types of tests where different people would se different headlines or different images, or in fact maybe actually just complutely different stories, and those tests will resolve in like fifteen or twenty minutes. So

that works well enough by itself. But then we realized, well, we could probably create more of these tests if only we could automatically create headlines for them. We could just be running these tests all the time for every single story. But what we found was, you know, not exactly so if the idea was to save journalists time and doing that in the end, I mean, you'd have to come up with something that's fairly solid and ready to publish out.

We were able to create something that allowed you know, journalists basically have different formulations that they could play with and maybe gave them some ideas of what to create,

but it still require people to look at it. In the end, how can businesses work better with their engineers, with their tech teams to sort of create and not stay siloed in a way that like somebody who works in marketing feels like, well, you know, there's actually this need that I have, but I don't know who to talk to about it, and that don't really know what

to do. It's an awesome question to me, Like one of the best things that you can do as a business is to put those people together, sometimes even physically. So when we started this project, you know, we literally co located engineers, product people directly inside the news room to sit with the folks who are doing this work. Now, when it comes to a I M L, you remember, these are just tools. These are tools to make work easier. Their tools in a lot of cases for automation and efficiency.

There's some problems that can't be solved without it. In the end, though, you know, you're still trying to solve some business problem, and most of those involved some sort of users that you need to get to know. So you know, even at the post we had data sciencests who were on those teams embedded in the news room as well. You know, they weren't kind of seated somewhere else thinking of problems on their own. There's a time and a place for creating room for prototyping sometimes that

has to happen to especially with really advanced technologies. But beyond prototyping, putting those teams together is super crucial. So how do you make sure, speaking metaphorically, you write a good brief to your AI team, well, your engineering team. I still think, you know, start with the problem that you're trying to solve. Like, if you're going in thinking let's use AI to solve something, I think you're probably starting the problem the wrong way, and start by framing

up the problem in business terms. For people, you'd be surprised, I think how much you know engineers and product folks really actually prefer to get that first before they start diving into what's the technology that I'm getting used to solve this problem? With a buzz around AI, especially right now. You know, people tend to go into it. I think thinking this is kind of something that's pretty close to magic.

We just use AI. It's going to solve these problems that we haven't been able to solve some other way. And that's not really the right way to approach it. But a I will be transformative, I think for organizations that apply it the right way, with the product mindset, with a good knowledge of the problem that they're trying

to solve, with empathy for their users. When we're doing our research for this panel, as an article in Bloomberg News saying that Jeff Bezos is pestimally very invested in the product, and then you called him Jeff in conversation, I found very impress So we will unders people sending me an email already without obviously telling us the contented your meetings. How does his vision imbue what you do? So certainly for us it's boon to have him owning

the company. I think that's one of the greatest things is you know, obviously at the Washington Post were known for an amazing newsroom, but we've also spent a lot of time investing in our engineering team that started to some extent before you know, we were purchased by Jeff, but certainly after we purchased us. It opened a lot of new doors for us, and it gets people excited to come and work for us and some of the

problems that we're trying to solve. It really inspires people to be able to build like a platform like we built, you know, within a newspaper company. I think would have been hired to Fathom probably ten years ago, but I mean today we really can say that, you know, we're a content company and we're a technology company. And I think part of that starts with him in the leadership

that he provides. More Sleepwalkers after the break, So, Kara, that was our conversation on stage with Matt monahantan CS in early January. It was interesting because we hear so much about tech companies becoming publishers, whether it's Facebook, YouTube, or Twitter, but we hear less about publishers becoming tech companies. I guess that's where Jeff Basos as an owner is what we might call a differentiator. So I was personally

struck with Matt's experiments with the headline generator. You know, for the time being, it doesn't work well enough to be a commercial product, but I think it will soon. You know, look at autocomplete when you send a Gmail like Sincerely Comma. You know I get those all the time, and it works. You know, in an apocalyptic reading, that means that machines will take over our lives and there will be no work left for humans. We won't have

to come up with smart headlines. But I think in a more optimistic reading, using algorithms to generate writing suggestions could actually enable originality. Reminds me of that Chinese science fiction writer who you and I have talked about named Chen show Fund, who actually used an algorithm to create ideas for his own work, and he used it when he had writer's block. He wasn't using it to replace his creative skill. He was using it as an enhancement tool.

And I think that's really interesting. Yeah, And in season one of sleep Walkers, we spoke to a filmmaker called Oscar Shop who actually shot a whole film written by Ai called Sunspring. Oscar and Chen turned the technology into a tool that actually serves their purposes. You know, you can develop all kinds of technology in a vacuum. About the technology that really serves people and fills a need.

Is the technology that sticks around, you know, speaking of technology that really sticks around, and then some technology that might not stick around. There's so much stuff on the floor of CS you and I have never been to see us before. I think we were very overwhelmed by what we saw and excited. It was kind of inspected gudgets paradise, and obviously, as someone who's obsessed with technology and consumer technology, I would have bought every single thing

I thought you tried to buy. I did try to buy that. I mean that keyboard with the mouse burnt into the keyboard. How much did this cost? I almost bought a laser cool laser patch for my back, which placebo or not made me look very hot. But no, in all seriousness, you know, there are things that were on the floor that are kind of amazing when you think about it, Like from this company called Pillow Health.

They've developed this device called Priya that looks like a little face, a cute little face, as they always do, and it's basically a pill dispenser that is voice and face activated, so anybody could have one. I could have one, You could have one. But I think they've developed it mostly for elderly people who have many pills that they have to take throughout the day, and who's children or health aids want to be able to control when their

medicine is dispensed. And I think for someone who might have memory impairment, physical impairment, the idea that someone who isn't in the room with that person could control when they're getting you know, vital medicine is really amazing. And you know, you say what you will about privacy. I think being able to do something like take care of your elderly parent with a device is you know, the perfect intersection of technology and humanity, right. I spoke to

the founder about exactly that. And you know, we have a lot of concerns about facial recognition that we've discussed at length on Sleepwalkers and will continue to discuss in season two. But in a narrow use case like this, in a voluntary use case where it can help somebody out to remember something very important, like what pills are taken, when it may well be that that's a sacrifice which

is very much worth taking. There was another startup on the floor that really caught my eye, which was called in New Pathy, and according to the card which have in front of me. It's the first device in the world which is equipped with technology to visualize your dog's

status from his or her heart rate information. And this is basically a harness that you put on your dog and it recalls your dog's heart rate and in particular the variability in your dog's heart rate to tell you if your dog is happy or sad, or anxious or excited or curious. And you know, people struggle to know what that dogs are thinking. And if you can use data from historical doggie feelings to model what a current dog is feeling and use that to have better interaction

with your dog, more power to you. I think it's cool. There was this other piece of technology from a company called we Labs. They were Japanese, right, and it kind of blew my mind in the same way that like thinking about automation of drive through blew my mind. You know. It was this like would beam that looked like a beam in a house, and it had a computer that

was inside of it. And the woman who was showcasing it basically asked ours to stay end up against it like you would when you're charting at child's height, and she took a pen or stylus and marked Oz's height, and then immediately that marking was uploaded into the cloud and displayed on a device next to this would beam And it just made me think, like this thing that millions of families do as their children are growing up is now being digitized, and again like going back to

the intersection of technology and human behavior, like imagine if someone moves from the house, those height markings that were such a part of your child's growing up can be taken with you in the cloud. It's just I mean, that's the kind of stuff where I'm like, do I need it? Does someone need it? Who cares? But the idea that it's like replicating this very very personal feeling and you know activity that we do in our childhood is I don't know, it kind of blew my mind.

All three of the things we ended up talking about, you know, pillow Health, the doggy heart rate monitor, and this Japanese WOULD device. You know, they go back to the most human things are our parents, okay, is our dog okay? Our children growing up? What does it make us feel as they grow up? And so technology that addresses those questions in a sensitive and humanistic way will always be interesting to us because it really allows us

to think about and tell stories about ourselves. The oldest stories we tell, the stories that are parts of novels and films and all other kinds of art. So that's to me where technology is most interesting and the types

of stories that will continue to tell on Sleepwalkers. So everything we just talked about is consumer focus and very interesting, but a I can also help address problems at scale, you know, issues ranging from climate change to pain management, and those are all things that we're going to talk about in our very exciting season two. Thank you for listening, and we're looking forward to seeing you for Season two of Sleepwalkers very soon. Und the Roots under the Roots for the f