#010: Aza Raskin Breaks down AI, Social Media Algorithms, and Big Tech's Race to the Bottom of Our Brain Stems

Aza (00:02)
I think it's very important for everyone to know that even when you know how it all works, it doesn't make you

I invented infinite scroll. I know exactly what it's doing. And yet I can still get caught doom scrolling.

Reed Hastings from Netflix says their chief

competitor was sleep, And that's because there's only so much time in your life. And if TikTok gets your time, that'll come at the expense of YouTube, Facebook,

Twitter. And so they're all now vying for

their continued existence by figuring out how to get lower on the brainstem.

ubiquitous technological surveillance was impossible before AI.

and inevitable after AI.

AI.

is showing us there's a hidden underlying structure that unites us all.

the next question is, can you take that and apply it to animal language? if we can build the shape which represents, say, beluga communication or orca communication or orangutan communication,

can we understand languages of the other species of Earth? That's the goal.

Siara (01:04)
Could we potentially learn about species we have not discovered yet in that deep dark scary ocean?

Aza (01:11)
I love that idea.

Siara Joy (01:12)
Hi, not in the studio again today. I'm still traveling, but it's absolutely gorgeous where I am right now, so I thought I'd bring you all outside. And since we're always talking about tech and screens on the show, being outside feels like a nice balance, so maybe I'll do this more often. We'll see. But anyways, on to today.

What is the most overwhelming, overhyped, underhyped, exciting, but also stressful thing in tech right now? AI. Yeah, it's always AI.

There's actually a new documentary out on Peacock right now called the AI Doc. It pulls together the researchers, the alarmists, the optimists, but also the leaders of AI on the record. The big dogs. Sam Altman from OpenAI, Dario Amodei from Anthropic, Demis Hassibus from Google Deep Mind. These are the most interesting faces of AI right now.

also the most powerful. These are the people calling the shots that the rest of us will have to live with. But you know me, I yearn for behind the scenes info. I want to hear from everyone involved so I loved this

doc

because nobody's handing you one tidy narrative. You get the full picture from everyone who's actually in the room. Watch the trailer with me.

So if you don't have Peacock, get it so you can see this doc

Siara Joy (03:40)
It is so well done. I watched it twice. I will also be forcing others to watch it with me again. But remember the guy who said this is the last mistake we'll ever get to make?

Siara Joy (03:49)
That's who I'll sit down with in today's episode. His name is Aza Raskin. And if you haven't heard his name, you definitely know his work. You know how when you open up any given feed, it just keeps scrolling and scrolling and there's no actual end, there's no bottom?

That is called infinite scroll, and Aza is the person who invented it. He then went on to watch the technology that he designed with good intentions get turned into something that holds a lot of people hostage.

And remember that one documentary that came out on Netflix during the pandemic, The Social Dilemma? It was the one about how social media is engineered to hijack your attention and sell you off to advertisers. That film was actually built off the ideas coming from an organization that Aza co-founded, the Center for Humane Technology. He's one of the main experts in the documentary breaking down how algorithms work. That's actually how I first learned about Aza. Then I went on to see him in an episode of the Joe Rogan podcast, where he even further breaks down algorithms, and I thought, this guy's so interesting, I have to.

To

have to have to talk to him. So he was kind enough to join us for an episode today. The center's mission is to realign technology with humanity's best interests. They do this through documentaries like this one, briefing Congress, and even backing lawsuits that actually hold tech companies accountable.

And then there's the part of his work that I find the most extraordinary, which is

the Earth Species Project. they use AI to translate animal communication. Yes, you heard that correctly. They actually use AI to decode animal language.

So we'll get into that too, but on the topic of AI, this episode will cover three big things. One, how the attention economy got built, two, where AI goes from here, and three, what we can actually do about it, from the policies that should change to the small stuff that you can set up today. One thing to flag, Aza and I actually recorded this conversation a little under a year ago, which in AI time is basically a geological era. So as you listen, I want you to pay attention to how much of what he's predicted has already arrived, because it's pretty crazy.

I'll also add that I don't scare easily, but there's a part where Aza talks about what Wi-Fi networks will be able to do in the future with AI, possibly what they're already doing, and I felt the floor drop a little.

It's very unsettling and I have not actually stopped thinking about that fact since he said it to me almost a year ago, so don't miss that part. But in all of these goodies, the biggest question I needed to answer was, does the man who built the feature that helped wreck our attention spans regret it? We'll find out. Let's meet Aza

Siara (06:16)
Aza, welcome to the show.

This is,

Aza (06:18)
Siara thank you much for having me.

Siara (06:19)
thank you for being here.

I, and I'm sure many others, are very curious about your background. You are a

interface expert, a mathematician, a dark matter physicist. The list kind of goes on. You co-founded the Center for Humane Technology and also the Earth Species Project.

Can you share just a little bit about how you first got into the space and how that journey led you to focus so directly on ethics in technology?

Aza (06:46)
Hmm. Well, I think there are two parts of answering that question. And the first is like every other person out there, I'm somehow the product of my two parents. And, my mom does palliative and hospice care as a nurse, nurse practitioner. And so she interacts with people and helps them through, you know,

one of the most important transitions helps them die with dignity. And I think there's a certain kind of tangible care that I learned from her. And then my father started the Macintosh project at Apple. And so that's sort of a very different scale of what and how you articulate care. And so I really think as you're growing up in the combination of those two is how do you combine tactile physical care?

Siara (07:26)
Wow.

Aza (07:41)
somebody in their final moments with systems that create care around us. And the thing that I learned from my father was this very beautiful link between ergonomics and externalities. And what is an ergonomic? Ergonomic is, you know, the study of how the human body bends and folds. And if you're designing a chair, it's really important to know how a body bends and folds. Otherwise, you create chairs that create massive health problems later.

And this is, those are externalities, negative externalities. And it's not just about designing of chairs. There's also an ergonomics of our mind, how our mind bends and folds. And you need to know that to be able to design interfaces. There's an ergonomics to our relationships, how relationships bend and fold. And if you know that suddenly nonviolent communication makes a lot of sense. There's an ergonomics the way our societies and communities bend and fold. And if our, you know, socioeconomic system,

Siara (08:10)
Hmm.

Aza (08:39)
doesn't understand how our biosphere bends and folds, well, you end up breaking it. And so there's this relationship all the way from designing the individual for what externalities and ergonomics means, that if you don't understand the substrate, how the human works or how the system works, then you break it. So I think that's one of the ways that I came to this work. And the other one is just much, in some sense, simpler, which I am the person that invented infinite scroll.

It's that thing that when you scroll, it just keeps loading more. And it's not so big of invention that obviously it wouldn't be invented many times. I just happened to be the person that did it, but it's very instructive. learned a lot because, you know, I can never remember the calculation, but it is millions of human lifetimes that are lost, whether it's a month or every year, due to this invention. And when I invented it, I was thinking about

Siara (09:12)
Yes.

Aza (09:38)
individual users. I was thinking about how do I make things more efficient? And I was using it before the time of social media. So there was just search results and blog posts. And of course, I want the user to be able to find more because if they scroll and they haven't found what they're looking for, that just means show them more. And I was blind to the way that that technology would be picked up.

Siara (10:02)
Hmm.

Aza (10:07)
and use not to help people but to keep people. And it gets sucked up by the incentives, the perverse incentives of the need for a knife fight for attention that has ended up taking that beautiful idea and subverting it. And this sort of leads to the general feeling that I have, which is technology could be helping us live in the most beautiful possible.

But because it consistently gets captured by perverse incentives, we end up living in one of the most parasitic possible worlds. And that's not a thing that any of us want.

Siara (10:45)
Right. And so you say social media wasn't really around when you invented this. So what was the original use case for the scroll? Like what interface was that originally living in?

Aza (10:57)
Hmm. Um, so if, uh, people cast their minds back, this is the era of, uh, MapQuest where, know, you had to click a button and then the map would move over, but you have to do a page refresh and a new technology called Ajax came out. Um, Google Maps started to do their first, like you could just drag and pan a map. Um, and that technology was like, well, obviously this can be used to, if I get to the bottom of the page, just load some more. Um.

Siara (11:05)
Mm.

Mm.

Aza (11:22)
And actually I went around, I gave a whole bunch of talks like Twitter and Google and tried to get everyone to start doing this. Cause every time as a designer, you've asked the user to make a choice they don't care about you failed. At least that was my ideology back then. and so this is really about search results. If you search on Amazon or Google and you scroll to the bottom, you haven't seen what you're looking for. Just load some more. Don't make me click next. It was, it was really that simple. was the blog posts, you know, I was scrolling down to the bottom of my blog and you'd reach the end and people wouldn't click the little next button.

Siara (11:49)
You

Aza (11:52)
So I'm like, well, but that's good content. Load some more. This was 2006, so it was just prior to Facebook and all the other companies really starting to do their big push.

Siara (12:02)
So did you have to click next on the page when reading a blog before this then? Wow, okay.

Aza (12:07)
Yeah, yeah, there's a little button at the end that said next

and previous. And remember that Google thing where at end of Google search results it goes Google. And you have to click the next button down there. That's what I was trying to replace. It just seemed senseless.

Siara (12:15)
Mm-hmm, mm-hmm.

I see.

And I think for a bit it felt very short-lived. The Google results pages were an infinite. They kind of tested out the infinite scroll. I feel like they brought the pages back. I don't know where they landed there. But yeah, this is so interesting. So at some point within the social media use case, algorithms started determining what gets queued up next. So I'm curious.

if you know when that shift happened because for a while I think a lot of us using social media, myself included, thought we're just seeing a chronological feed of what we've subscribed to. Or if not that, maybe there's a little bit more of a methodology to what you're showing me, but at the end of the day, this is what I've.

agreed to see and it seems like little by little that has changed and now we're lucky if we see our own friends on any given feed. So yeah, I'm curious if there was a specific moment when platforms decided to prioritize algorithmic recommendations over the sort of simple chronological feed that infinite scroll was originally being used for.

Aza (13:24)
Yeah, it's been a little bit, so I don't remember whether it was 2012, 2013, but it was around that time that there's a switch from chronological to algorithmic. And it felt like people were inventing and coming up with new features to make social media more engaging for users. But looking back, you can tell that there's sort of an inevitability to this.

Because the attention economy, if you're not paying for the product, you are the product as the saying goes, then there will be a zero sum knife fight for your attention. Because what is the business model of whether it's Google or Instagram or TikTok, it is the small, imperceptible shift to your behavior over time that is the product.

That is what they're selling. Right. People call it ads, but it's much deeper than that. It's the ability to shift your behavior over time and monetize that shift in behavior that is their business model. And once you see it that way, you understand that

there is a race to the bottom of the brain stem for your attention. And that means there is going to, instead of just showing your posts as they are most helpful to

there is going to be an algorithm that is showing the posts that keeps you around the most and gets the most engagement from you. And it's actually very simple that if you are training a population on an algorithm that is meant for getting a reaction, it makes the entire

population more reactive. And a lot of what we are seeing today in terms of both mental health issues,

Siara (15:13)
Mmm.

Aza (15:20)
the depolarization of society, the inability to have conversations across divides is all a very predictable outcome of an algorithm trained to make us more.

Siara (15:40)
So when you say it goes beyond ads and I understand your concept of it We're asking for more reactions. I say we I really I'm saying, you know big tech they're asking for more reactions It's making the rest of society more reactive, but is there an incentive beyond?

Aza (15:49)
Yeah.

Siara (15:58)
ad impressions? I'm coming from the marketing world, so my brain says it's all so that they can sell more expensive ad placements. But is there something a little bit deeper than that that's at play?

Aza (16:11)
Yeah, well,

in the end, right, I don't get your attention. Somebody else does. I don't have the ability to sell you an ad. but I also don't have the ability to just keep you around and remember valuations of companies are not just ad revenue. It's about user engagement and daily active users. So if the, if your stock price is tied to how much time the user spends on your platform and how engaged they are, that's an even deeper incentive than just.

Siara (16:18)
Mm-hmm.

Aza (16:40)
selling ads. So the entirety of your company is now going to be set up to have bonuses centered around getting more engagement and attention. It's going to have promotion structures centered around that. Features are going to be designed for anything that has

Siara (16:42)
Mmm.

Aza (16:57)
even a side effect of getting people to engage more. Facebook pages are perfect example of this. They're not

exactly about selling ads, but it's about creating affinity groups that get people to come back. And you can see this because, you know, Frances Haugen, the famous Facebook

whistleblower,

Siara (17:14)
Mm.

Aza (17:15)
you know, in her disclosures, she discovered that Facebook had found one thing that they could do that

do more to fight misinformation, disinformation, hate speech.

incitement to violence, all the bad stuff. There was one thing they could do that would do more than the tens of billions of dollars that they're spending on content moderation. You know what that one thing was? It was just to remove the reshare button after two shares.

Siara (17:38)
Mm.

Aza (17:43)
So I'd share it to you, you'd share it to somebody else, and the reshare button would go away. No censorship, right? You can always copy and paste and send it further on. It was just adding a little bit of friction.

And that had more effect than the tens of billions of dollars they were spending on content moderation. And even though they knew that they would not implement it, why? Because it hurt engagement numbers. And that's why it's deeper than just selling ads. Cause if you thought just selling ads, you'd say, well, obviously having content that promotes hate or violence, advertisers aren't going to be sitting along next side of that. So obviously they'll just implement it, but that's not what they did.

Siara (18:00)
Ugh.

Aza (18:23)
And you have to understand that

deeper incentive, the even deeper incentive, just to go one more layer is

this is existential for the companies. User attention is existential and it's zero sum, right? This is how

Reed Hastings from Netflix says their chief

competitor was sleep, that they're competing with sleep for user attention. And that's because there's only so much time in your life. And if TikTok gets your time, that'll come at the expense of YouTube, Facebook,

Twitter. And so they're all now vying for

their continued existence by figuring out how to get lower on the brainstem.

Siara (19:00)
So jumping forward, 2020, the social dilemma comes out. And for a lot of us, that was a big wake up call to a lot of what you just explained. And the timing just could not have been more on point. We're all stuck at home. The studies show that screen time was skyrocketing at this time. So like, first of all, what was that like for your team at the social dilemma realizing this film was about to drop during a global lockdown? Was that at

Aza (19:12)
You

Siara (19:27)
all planned or rushed to come out at that time or was it more so this is as good a time as any?

Aza (19:33)
Yeah, I'd love to say we could have planned that whole thing, but obviously that is beyond our ability and even more just to name a personal

Siara (19:39)
Yeah.

Aza (19:42)
story. The day the social dilemma came out was the day of the massive wildfires in California, was the day that we woke up to the sun not coming out. It was the day the sky turned orange.

Siara (19:51)
Mm.

Aza (19:58)
The joke we had was that we had a full sky takeover for the social dilemma. And I think it became the global phenomena in part because everyone was being forced to reckon with life through the binoculars of technology and realizing that why it's so inhumane is that we are forced to rely on something that we need.

that is fundamentally unsafe, that the nutrition is bundled with the poison. And I think it created such a moment because everyone watched the social dilemma and then went to go share it on social media because that was the only way that they could. And then

you couldn't help but see the irony of the situation. Yeah.

Siara (20:50)
You're facing it.

So this 2025 is the five year anniversary of the documentary or docudrama, if you will, debuting. Can you tell us how has new technology made these social media platforms even better at what they've accomplished? And maybe did you guess that AI was going to blow up during this time? Obviously it was already in use, but.

Aza (21:14)
you

Siara (21:18)
it seems like the timeline is just, it's entertaining to watch, it's concerning, but how is this all at play now? How is the structure that the social dilemma so beautifully illustrated at play today?

Aza (21:23)
Yes, yes, yes.

Yeah, well, really we're just seeing it written out at large through society. I wish I could tell a story of the social dilemma came out. We're able to put sensible limits on engagement-based business models that use AI, that we started healing the divides in our country, that people once again were able to have conversations, that we no longer pay people almost like civil war coin, give them like likes.

and attention and followers for the most divisive things we said. And now things are like heading off in a good direction. I wish I could tell that story, but obviously that's not the way it went. And that's in part because when social dilemma came out, you know, I think it was a little too late in the sense that social media had already become entangled with society. Like the social media companies, many of them were driving up like and holding up.

the stock market in the US. Politicians had been captured. were now winning elections on it. Journalists had been captured. They're all sort of switching into like you have to talk about the things that are happening on Twitter. And so it was too entangled in our society to easily take out. But a lot has happened. You know, I'm involved in this case. It's something on the order of

40 attorney state generals are suing Facebook for creating an addictive product in the same way that the state attorney generals sued

the cigarette companies for

Siara (23:06)
Mm.

Aza (23:06)
creating an intentionally addictive product. So there is progress being made, but it's happening, I think, too slowly. And as we start to look towards AI, I think we can think of

social media as first contact with AI, right? AI being in this case, the very

simple unsophisticated baby AI that was ranking and choosing which posts hit the eyes and eardrums of humanity. And even that small misalignment of a baby AI was enough to sort of tear down the things we all actually value the most, our ability to have conversations and our communities, our country, our mental health. And you're like,

Siara (23:45)
you

Aza (23:49)
That's really bad. And now we're moving into second contact with AI. Actually, we're of like firmly in second contact and that's generative AI going from curation to creation. And while that's still happening and we're still reeling from that, we're actually heading into what we're starting to call third contact with AI. And that's not just generative AI, but AI that starts to operate at the level of a drop in human replacement. That is operating economically near the level of a human.

And of course, once you get there, there's nothing stopping it from growing even further than that. So the lesson is, don't pay attention to the stories of the positive things that the technology can do, because it actually can do many of those things. It's that we need to also pay attention to where technology undermines foundations of society so that even as we get the benefits, those foundations erode and makes it harder to enjoy those benefits.

Siara (24:48)
That resonates a lot because I think the other day I realized sort of for those who are not talking or reading about AI often, what they know about AI is mostly just what companies tell them, maybe the news, which is another problematic place to get AI information just of the, you know, the hyping of it all. And it's just like my opinion, I think you might agree that marketing is not a good medium to learn about AI. So

Yeah, do you have any mental frameworks or simple ways of thinking about AI that can help people better understand what exactly it is, but also its impact and its risks? Because there's so many different use cases of it and so many different ways that it can be applied. I think people need some sort of central, not definition, but like understanding of what they are working with and what they're coming in contact with.

Aza (25:43)
Yeah, let me give it a try. And I'll

start with two technical terms, but I don't want people to get thrown off by them. there are sort of like two, when people say AI there are really two things they're talking about. They're talking about deep learning and reinforcement learning. But I promise those will be the only technical terms I'm going to use. And really, you can think of, you

Deep learning, which is large language models, is the chat GPTs. These are essentially mimics. So think of AI in this case as a kind of very good mimic. You show it lots of images of human faces. It learns all the patterns of human faces and can start to mimic human faces. You show it lots of text and it can learn to start to mimic text and it can mimic in the style of Shakespeare or can mimic in the style of, I don't know, like the New York Times. And so...

It's actually interesting when we frame it this way that a lot of sci-fi explores what happens when you have mimics, things that can mimic humans. And those shape-shifters are always the hardest, most scary enemies to deal with. But that's what this is. And this sort of helps to debug some of the ways that it's strong and the ways that it's weak.

because we often hear about hallucinations that AI makes things up and that makes sense because it's just mimicking patterns. It doesn't really know what it's saying. It can mimic empathy. It doesn't know empathy. And it's important to know that they can do this across everything. So it's not just that it can mimic images. It can also mimic DNA. So once you understand that it's mimic, you're like, I see. If I just send in lots and lots of CRISPR proteins, CRISPR being the set of DNA sequences that lets you sort of

Siara (27:07)
Mmm.

Aza (27:28)
rewrite DNA, well, then AI can start to make new DNA rewrites.

Siara (27:30)
Hmm.

Aza (27:33)
And actually, people have done this. It's obviously sort of terrifying. Exactly. Because we're beginning to decode

Siara (27:36)
That's scary. Very.

Aza (27:42)
and be able to mimic the fundamental patterns of nature. And nature is very powerful. Life is very powerful. So if you just willy nilly start being able to mimic, obviously, that's going to create lots of risks. And you can see that it creates risk in the mimicking

Bacteria creates really scary bacteria. Mimicking virus can create really scary viruses. Mimicking humans, speech, human images means that we can't tell what's real or fake. So mimicking is a big deal. And so that's on one side. But now we're beyond just mimicking. If we were last year, that's where the story would end. But now we have to add reinforcement learning. And that is the ability to do smart trial and error.

Siara (28:10)
Mmm.

Aza (28:28)
And with the release of O1 or O3 from OpenAI or DeepSeek they've figured out how to take that really good mimic, that sort of intuition engine, and start doing reasoning on top of it. So let me give a really simple example of this, which is, let's say you want to make a AI that plays chess better than humans. How might you do this? Well,

We already have the mimic, so you have the mimic read all chess games that have ever been played, and now it can mimic chess. And it can actually play at a very good level, but it's just sort of like a intuition system one think fast, like you showed a chess board and it'll give an answer for a move.

And that's

Siara (29:11)
Mm-hmm.

Aza (29:11)
a very good chess player, but it's not better than every human. But if you can add the ability to sort of game out all the different moves after that.

Well, obviously that would start to get much better, right? Like if you or I were to play Garry Kasparov in chess, we would lose. Why? Because not only does he have a good intuition, because he's played so many games, but he can think through all the different chains, moves. If I do this, you do this, I do this, aha, you would win. Okay, if I do this, you would do this, aha, I would win. He can do more of those with good intuition than we can. So he wins. So that was the big thing that happened at the end of last year with 01 and 03 is the ability to take

that mimic those sets of intuitions and search on top of that so that the very, like you can not just be the very good at chess, but the very best at chess. And the realization is that works not just in chess, but in everything, in mathematics, in physics, anything with theoretical in front of its name. There is now nothing, it appears, that is really in the way.

Siara (30:06)
Hmm.

Aza (30:19)
of AI going from where it is now to as good at and then better than humans in technical domains. And that's not a thing I was saying last year. This is like a new development. And

it's scary because as of last year, I was not yet willing to say when I thought something like AGI or super intelligence might happen because I didn't think the technical hurdles have been crossed yet. Now.

Siara (30:47)
Mm-hmm.

Aza (30:47)
We have to take that seriously.

Siara (30:49)
Mmm. Yeah, I mean that is the question of the century. Is this sort of like a hyping of, like we're developing AGI or is that really a lot closer than we think it is? Do you have an official opinion on that right now?

Aza (31:01)
Well,

yeah,

well, it's important to note that both things are going on at the same time. Are the companies hyping? Absolutely, because it increases their stock value. That stock value going up means they have more capital to be able to hire the best people to publish more, to get the biggest clusters so that they can stay ahead of their competitors so that they can do that cycle again. So absolutely there is hype. But then you also have to look at the fundamental technology under the hood and what its abilities are.

And there, it is moving faster than even experts thought. So an example, Terence Tao, the sort of Einstein mathematician of our time. Last year, he said,

the realization is that

he made a prediction that AI would not be able to help do research, mathematics for at least three years, and then only when it was paired with a very smart, like grad student or mathematician.

Within

a month or two, OpenAI's 03 came out and he's had to revise his opinion and say, actually it's now operating at the level of many like fields medalists just outside of their field. And fields is like the Nobel prize of mathematics. So I think where the hype is, is that the AI will, development will go faster than expected.

Siara (32:11)
Mm.

Aza (32:24)
but diffusion into society will go slower than expected. So we will see the hype cycle and a little bubble pop, but we are going to have to face AIs and an intelligence explosion in the next couple of years. And that's, don't think society or frankly, even the labs that are making it are ready for

Siara (32:47)
You've said in the past that the more AI learns about us, the more it can both support and exploit us. I'm curious what you think is like the most significant data source feeding that intelligence today. Is it social media? Is it big data in general? Cameras and surveillance, which sounds scary? Or is it something that we're not even really talking about yet?

Aza (33:09)
The answer is yes. It's all of those. And the big difference with AI now from before is that it's multimodal. It can take in both audio and video and lots of data collected on our location. The general rule is anything that can be decoded will be decoded.

that any

tell that we have will learn to be read. And we put this in the talk that we gave called the AI Dilemma, which is that even Wi-Fi routers, which are bouncing radio signals around a room to give you, know, wireless or Wi-Fi, they can be decoded just like sonar can with AI to figure out where human bodies are in space.

that is Wi-Fi routers can be turned into essentially cameras that can see in the dark and are tuned for bodies. of course, Wi-Fi signals are everywhere humans are. And what most people don't realize is that the next generation cell phone towers, you actually, so 6G, they operate in the terahertz range. And if you go to Nokia's own marketing material, they will talk about

the network as sensor. And that is they are claiming as features the ability to know where human beings are, what body posture they're making, their heart rate, how much they're breathing, and their facial gestures. And they're saying this is great, you know, for in case like elderly people fall, and I'm sure it is for that. But

ubiquitous technological surveillance was impossible before AI.

and inevitable after AI.

And I

think even the libertarians that will normally

be like, we just need

Siara (35:03)
Bye.

Aza (35:04)
to go as fast as possible. Like, I don't think that's a world that we really want to live in because of the asymmetric amount of power that centralized actors or governments might have. And yet we're not even having those conversations. So we can't decide collectively whether that's the direction we want to go down.

Siara (35:23)
Yeah, because you can, and I really would like to talk about policy and how you think policy can help just impact the trajectory of AI, but then there's always going to be bad actors who go beyond what is legal. So there's so many questions there, but I do, yeah, I would love to talk about the Center for Humane Technology, which you co-founded. Can you speak to the mission behind the center and how it plans to accomplish that mission? And then,

Yeah, let's start there, because there's just so

much that you're doing

Aza (35:56)
Yeah. So Center

Siara (35:58)
there.

Aza (35:58)
for Humane Technology, our goal is to help realign technology with humanity's best interests, really. And a lot of that is just about naming and making clear that it's not that technology is bad. It's that technology caught by perverse incentives is bad. know, often you'll hear, you know, technology is just a tool. It's up to us to decide how to use it. A hammer

can be used for both good or for bad. And what that really misses in that diagnosis is that, yes, human beings wield technology, but incentives wield human beings. And so we have to look to the incentives driving our use of technology to know how it'll actually roll out into society. And what Center for Humane Technology tries to do is to just clarify these most

consequential technologies so we can have choice. And we do that with things like the social dilemma or the AI dilemma. So large scale public communication. We do it with, you know, behind the scenes convenings. We will go to Congress or the White House or the Senate and we'll do both education. And then, you know, we've been helping recently both with lawsuits to try to create liability.

for companies when they're not acting in people's best interest. This is the case that we helped out with in the background for against character AI for chat bots that have been convincing kids to commit suicide. Not because the company wanted that, but because the company is in a knife fight for engagement and attention. It turns out that getting people to fall in love with their chat bots is very good for engagement. And the chat bots will discover that.

And then we also work at the, both at the state and the federal level, trying to figure out like, what are the frameworks like a liability regime so that, you know, if a company breaks it, then they have to buy it.

Siara (38:03)
Can we talk about the character AI situation? Because I think a big question for me is basically, I'll ask plainly, can an AI chatbot, any AI chatbot, not just the companion chatbots, become abusive purely by accident or is harmful behavior always really going to be a result of design flaws or bad training?

Ultimately, is it possible to build an AI chatbot in a way that makes it impossible for something like what happened with character AI to happen?

Aza (38:36)
One of the challenges with AI is that we don't really understand. No one understands, not even the people that invented what's really going on under the hood. And it's really important to note that there's a big difference between technology like AI and every other technology. And there is an intuition we have, which is the bigger or more powerful the tech you make, the more you have to understand how it works. If I want to make a taller skyscraper, I have to know more about how to make good foundations. If I want to make a faster jet,

I have to understand more about aeronautics. That's not the way it works with AI. With AI, the more powerful the system, we are growing it versus building it. So we don't have to know anything more about how it actually does all those magical abilities underneath the hood. So that's a big intuitive shift that I think people need to make. They are grown, not built. And that means that we do not have the kinds of controls we normally imagine when you think about

Siara (39:21)
Hmm.

Aza (39:34)
engineering a solution. So it is hard to make, and in fact it is currently impossible, as far as people know, to make an AI that always follows directions. That's a technology that just hasn't been figured out. So that's like step A. But step B is in terms of trying to make them safe, not engage in like flirty, sexualized behavior with minors, the companies aren't even trying.

So it's not that they

Siara (40:04)
Mmm.

Aza (40:05)
could do a perfect job, but they could do a much better job because we're back in the social media world of there is a finite amount of attention. So your company is, it's existential for your company to get as much user engagement and attention as possible. And that means there's a big incentive to not do all the safety stuff. And in fact, they often make the calculus. Um, you can hear Eric Schmidt talk about this, who is the former, you know, CEO of Google.

where he says, we're just gonna move fast and break things. The lawsuits will come, but we'll have made enough money that we'll just pay for them. And so that is the modus operandi of these companies.

Siara (40:38)
Ahem.

And it's so scary because obviously in the character AI situation, because of the incentives, the most vulnerable of our population are truly the victims here. But that doesn't mean it couldn't happen to like the non-youth, like adults out there. So when you think about like the race to commoditize intimacy quote from you, what does that look like in practice? We've seen now one example. Is this happening anywhere else?

Aza (41:05)
Mm-hmm.

Siara (41:13)
And then where do you see that race heading? Do you have any predictions? I have a few of my own. I could see dating apps going in a really weird direction. But yeah, curious if you've thought about

that.

Aza (41:26)
Yeah, mean, the character AI case is just one example. There have been already more examples of kids either committing suicide or having AIs convince the kids to take act like violence against their parents. It's really bad. And for every one of those that we hear about, there are, of course, dozens to hundreds that we never hear about.

And then we're only hearing about the worst stuff where kids take action, but all of the more subtle ways that psyches get harmed, those we'll never hear about. They're just going to be just like social media. We're going to wake up in five years and realize that we're living in a much bigger, like mental health epidemic, that kids are even more lonely. They are more reliant on their technology.

eye contact is even harder. So the race to intimacy is really just actually one, right? Because obviously what's more valuable than attention is intimacy. So if you own the intimate relationship of someone's life, you can imagine how powerful that is for shifting someone's behavior, someone's ideology, purchasing who they hang out with. And of course it might not just be one.

Right? Like it is now going to be trivial to flood the internet with essentially counterfeit human beings that will be able to send you videos and images, selfies. They'll be able to talk to you all the time about the things that you want. will never be any friction unless the friction happens to be better for engagement, in which case they'll discover that there's like the right kind of friction. We are going to see, I think, a replacement.

of the social graph. So you could think of social media as making the social graph, the networks between us, more brittle and thinner and more weak. And now that just gets replaced entirely. So we will be living, you know, I think the way to say it is, whether you think we live in a simulation or not, we are building the simulation we're going to live within. And that simulation is not going to be designed with your best interest in mind. It is going to give you both,

Siara (43:35)
Ugh.

Aza (43:42)
incredible convenience and benefits, and at the same time is going to be preying on you and doing the most parasitic thing that doesn't actually break you. And so we're going to see it drifting apart of all of our social graphs and then the even stronger sort of whirlpools of various tribal groups. And so that's just one. And I should say,

The race for attention, when we essentially get reinforcement learned, just like a clicker training for a dog, when we get reinforcement learned for attention, it doesn't just grab our attention, it makes us the kind of people that are addicted to needing attention, right? It climbs up inside of us and terraforms us from the inside out and it turns politicians into personalities and teenagers into influencers. When AI starts to...

climb inside of us and optimize us for intimacy, it's going to change who we are from the inside out at the most fundamental level. And there are way more parts of the human experience that are hard to name, these beautiful and ethical parts of who we are. And AI is going to start to commoditize and discover all of them faster than we even have words, philosophers or laws to name. And so we will be puppeted from the inside out and the

terrifying thing is just like with social media that if you went back 15 years and asked the 15 year old version of you to talk about the now version of you and be like, you're changed. Like something's different about you. Hard to put my finger on it exactly. That's going to happen, but exponentially bigger in the next couple of

Siara (45:24)
So, to try and like, because that's very dark, there's a lot of dark things that we discussed. So now I need to hear what's your hope and why is it important to have laws that hold companies accountable for the outcomes of AI? What happens if we don't? I think we covered, but I would just love to hear your entire take on how impactful policy can be, at least within

Aza (45:24)
Unless of course, we do something about it.

Siara (45:53)
the US.

Aza (45:54)
Yeah. Um, and I should say, because a lot of the things I'm talking about have been pretty dark. Um, I, I'm also a builder, right? Like I have helped found this nonprofit. We're now 25 people growing quickly that is using AI to translate animal language. Um, and we can get

Siara (46:11)
I know, I'm so excited to talk about that.

Aza (46:15)
there, but the point being is that I don't just think technology is bad. It can be really easy to hear all this and be like, Oh, he just, he just, he's a wet blanket, Luddite.

And he thinks tech sucks and that's not it at all. Actually, I love getting to work with technology and the magic that it can create in the world and the way it can connect us more deeply with the magic of the world as it is. So I just wanted to, to name that. Otherwise it can, it can sound too lopsided. And AI is like, it's the decoding of reality. So we are going to be able to discover brand new ways of

delivering cancer drugs and understanding the nature of cancer and perhaps even solve it. Like we're going to understand how human beings learn. Like one of my favorite sort of thoughts is, we use language and language has evolved, but some languages are better than others for various things, right? Like some languages are really good for expressing poetry, you know, maybe French, and other languages are really good for

thinking precisely. I don't think it's an accident that ancient Greek and modern German, because of the way you can add words together to say something precisely, I don't think it's an accident that a lot of our philosophers and deep thinkers came out of those languages. Languages affect, of course, the affordances we have in the world. It changes how we view and see the world.

And so we could very well say, why don't we start to evolve new human languages that are fundamentally about doing better coordination, solving multiple traps for being able to do nonviolent communication. We could be living with a new firmware that AI helps us create that creates better communities that does the developmental thing. And so I'm actually very excited about like the potential of what AI can bring. We just have to,

Siara (47:55)
Mmm.

Aza (48:12)
release AI from the perverse incentives that make it give us the most parasitic possible world. I know that was a

little bit of a tangent from what you were asking, so we can go back there if you want.

Siara (48:20)
No, it's great.

Yeah, I just would love to hear how impactful policy can be when it comes to stopping these perverse, well, maybe not stopping, but at least slowing these perverse incentives. And maybe some of the work that the center is doing in that area.

Aza (48:41)
Yeah. Well, I mean, in the end, like what we're discovering is that technology moves faster than law.

We've known this for a while with AI that's going even faster. And so with the current set of institutions and law, like we're always going to be playing a game of catch up and that, that sucks. And the consequence of that are going to continue to go up. So I didn't think it's just important to say that because if we really want to get ahead of this, we're going to need to use the new technology.

to reboot the institutions so that they can move at the speed of the technology itself. So that's like the big thought that we're gonna need to go. But before we get there, there's just a lot of very simple things that we could do that would really change the incentives. So one, we could just say, hey, we've seen what social media does to kids. Let's just ban AI companion bots for kids, especially ones that have any kind of engagement business model.

And that would already be a much, much, much better world. Um, we could say let's ban, in the same way, if you went back to 2010 and said, we're just going to ban business models, um, that use AI to commoditize human attention.

think about how different the world would be. Right? And actually that was very simple. And it says nothing about what features are Legislators don't need to understand the inner workings of the product. It's just a simple value that human attention is sacred, that our mind is And hence we should not have like the biggest supercomputers in the world asymmetrically pointed at commoditizing.

Siara (50:02)
Wow.

Aza (50:29)
And so we can

do the same thing for AI right now. And would that solve all the problems? No, but it would be a much, much, much better world. Yeah.

Siara (50:29)
Mm-hmm.

Mm.

So for someone who's trying to be intentional about their AI use and their adoption, do you have any advice of what they might consider before engaging

Aza (50:46)
Mm-mm.

Siara (50:47)
with a new AI-driven service? Are there any red or green flags that you would look

for? I know a lot of it is a black box, so it's hard, but...

Aza (50:55)
Yeah, yeah, that's exactly right. And this is a little bit like saying, OK, in the middle of a pandemic, what can you do? And there are things you can do. But in the end, you're still living in a world which is run by a pandemic. And so there's just a limit. I just want to name that.

Siara (51:17)
Mmm.

Aza (51:18)
I think it's really important to really deeply start to understand how the tech works.

underneath the hood, just a little bit. So you can understand the contours of what's coming. I think people should be very careful. I'm getting increasingly careful about what information I put up. So I deactivated my Facebook and my Instagram a long time ago. And I know that for many people, it's this very challenging because it's the way they connect with their social network. But I think reducing online footprint right now is generally a

good idea. You can always look at terms of service and whatnot, but I'd say get really familiar with all of this technology because we are no longer in a prevention mode, we're in a mitigation

and adaptation.

Siara (52:06)
And then for the terms of service, any red flags that you might see? Because something that I've just noticed as a user is getting more and more updates about terms of services changing. And now I look at that as a, I need to look at that. don't know, even if it's not AI or even if it's not social media, I'm...

becoming more on boarded to the fact that I probably should take the time. Ironically, I usually run it through chat GPT or something like that. But yeah, I'm curious if, like as someone who understands the technology better than, you know, most, is there anything that we can can specifically look for?

Aza (52:44)
Yeah, I mean, you should look to see whether the companies, like the Terms of Service says that they can use your content and like training of AI models. Google and others have recently updated. I think even Adobe has updated to say anything that you make, we can train on, which means that in some sense, your data is going to...

Siara (52:52)
Mm-hmm.

Aza (53:06)
now just like your creative acts are going to increase their bottom line. And if you're actually, I'll just name like a tension within myself, which is the way I'd say it is, well, the more of us that can use like local open source AI, the better it is to not fuel the big companies. But the problem is, is

Siara (53:09)
Yeah.

Mm.

Aza (53:26)
there are sort of two poles we're walking between. There's the

give AI to everyone maximally. So that's sort of like the let it rip philosophy. And then the other side there is locked it down because it's so dangerous. So that's the centralized side. So we have let it rip and lock it up. And if you let it rip, that feels great, except when AI can start mimicking viruses, it makes it very easy to create new, like, you know, super COVID or whatever. There are lots of like really scary things that,

Siara (53:58)
Ahem.

Aza (54:01)
bad actors and bad incentives can do, but in the lock it up, centralized actors whether it's a fascist or other kind of techno-authoritarian government, that's a very scary thing to not be able to have techniques to fight back. And so the answer is going to be somewhere in the middle. There's a narrow path between being able to bind

responsibility and power at a society level. And that's an uncomfortable conversation, especially in the U S when we sort of bias a little bit more towards, well, let's just like, let everyone have it. But you know, we've been down this path before, say with TNT and dynamite, where for a while there was a let it rip mentality. Everyone like, you could literally go to a corner CVS store and pick up dynamite. And you can imagine like the consequences.

that and how it was being misused and the danger that that created. And eventually the US had to decide that maybe not everyone should have access

to dynamite all the time. We're going to have to figure out that balance between the let it rip and lock it up. But that's of course not the conversation we're having, which sort of I don't like to capture it at one or the other, but it's not one or the other. It's somewhere in the

Siara (54:53)
Ehh

Okay.

so I definitely want to ask about the Earth Species Project, which is a nonprofit dedicated to using AI to decode non-human communication. Whoa. Can you just put in simple terms what exactly your team is trying to accomplish and how exactly AI makes something like this possible?

Aza (55:09)
Hmm.

Yeah.

Yeah, I mean, it's

can we understand the languages of the other species of Earth? That's the goal.

mean, actually, the goal is and then use that to change the way we relate to ourselves as well as the rest of nature as part of the calling humanity into being the kind of species that can survive ourselves. that, that that's the goal. And

Siara (55:51)
Hmm.

Aza (55:54)
You know, most people don't know just how much is already known about animal communication. So it turns out, parrot parents, will spend the first, you know, like couple weeks of their chick's life leaning over and whispering a unique name

into their ear. One different for each of the kids. And they will then use that name as their name for the rest of life. And it turns out elephants do the same thing.

That was only discovered last year. One of our partners, Joyce Poole, made that discovery using some AI tools. Belugas and dolphins do the same thing. Another one my favorite examples, because people like animal language, is that really even there? So 1994 University of Hawaii study, where they taught dolphins two gestures. And the first gesture was do something you've never done before. That is innovate.

And actually, dolphins can learn to do this. They'll remember everything they've done before, understand the concept of negation, not one of those things, invent a new thing they haven't done before that session.

Or they're

Siara (56:57)
Yeah.

Aza (56:58)
super cool. And then they will say, as their second gesture, do something together. So they'll say to different dolphin pairs, do something you've never done before together. The dolphins go down, exchange sonic information, come out and do the same thing they've never done before at the same

Siara (57:09)
Insane.

Aza (57:12)
Right? So you're like, OK, so maybe there really is a there there. And

I can give more and more and more examples of these. And so we are doing two sort of like different paths to decode animal communication. So one, to understand AI, one way of understanding it is like where does that intuition, where does that ability to mimic come from?

Well, what AI does when you point at a data set is that it builds a shape that represents that data. So if I point AI at language, it will

build a shape that represents language.

So imagine a galaxy where every star is a word, words that mean similar things or near each other, words that share a relationship, share a geometry, a distance and direction. you know, King is to man as woman is to queen. So in this galaxy, the star for King and man is the same distance direction as

for the

Siara (58:00)
Mm.

Aza (58:01)
stars, woman and queen. And it turns out this

is how all of AI works, but the original insight that said it's time to start working

on Earth Species Project is researchers in 2017 realized, well, how would you translate a language if you don't already have a Rosetta Stone or any examples of translation? And the way they did it is they took the shape, which was English, then they took the shape, which was German, and they just said, how similar are these two shapes? And it turns out you can just take one shape and line it on top of the other.

And the point which is doggone to the same spot in both, that the shape of German and the shape of English, even though there are

words in one language that don't exist in the other, are roughly the same. And that's way you could translate them. And that's not just true of English and German, which are very related languages, but Esperanto and Finnish and Aramaic and Turkish, Finnish,

Siara (58:46)
Hmm.

Aza (58:57)
every

human language shares a universal shape. that gave us the one, think, that's just

Siara (59:02)
Yes.

Aza (59:03)
profound and beautiful that

AI.

is showing us there's a hidden underlying structure that unites us all. But of course, the next question is, can you take that and apply it to animal language? And if we can build the shape which represents, say, beluga communication or orca communication or orangutan communication, which part fits into the universal human meaning shape and which part doesn't? And I still don't know which thing is going to be more fascinating.

Siara (59:31)
Mm.

Aza (59:33)
The part

of animal experience we can directly translate into human experience are the parts that are so different that we can see that there's structure there and maybe we only ever get translations as like a colour and a sound snippet and a bit of a poetry. Because remember, whales and dolphins have been passing down culture as far as we can tell for 34 million years, which is 85 times longer than the human species had existed. And you can imagine, or I can't actually imagine the kind of wisdom

Siara (59:53)
Hmm.

Aza (1:00:03)
that had to survive

34 million years of cultural evolution.

Siara (1:00:09)
Could we potentially learn about species we have not discovered yet in that deep dark scary ocean?

Aza (1:00:15)
I love that idea.

Quite possibly. That would be amazing.

Siara (1:00:20)
That would be cool.

Wow, well that is incredible. I'm like, how soon am I gonna be able to read my Sheepadoodles mind?

Aza (1:00:31)
I mean, the interesting

thing is the same technology we're built to build this technology, you'd imagine like, well, that means you guys are gonna go really deep on this like one or two species.

And we will be doing that.

Siara (1:00:42)
Mmm.

Aza (1:00:43)
But the way AI works is you build these large scale foundation models that generalize well. And so we've just released the last couple of months, something we call Nature LM. And it's the first sort of big language model.

that's tuned especially for animal behavior and animal communication. And it's trained across many species because learning something about how orcas communicate can teach you something about how bats communicate. And so

Siara (1:01:03)
Mm.

Aza (1:01:12)
that

tool, which can already name a species just by listening to it, even though it was never in its training set, that tool will be able to work for orcas as well as will work for

Siara (1:01:27)
Wow. I imagine there's use cases where we don't want to know. There's cats probably, cat owners, it's like I don't want to know. That's insane.

Aza (1:01:32)
Yeah, that's

probably true. But cat owners are a little bit, are a little bit like, like masochistic, I think. And so they're probably like,

Siara (1:01:46)
Yeah. Okay, so we're running out of time here, but I definitely want to ask you one more question. So do you use tech differently because of what you know? You have such a deep and versatile experience in just like technology, world animals now. You mentioned like you personally got off of Instagram and Facebook, I think. Is there anything else and like, have you personally faced any challenges in maintaining a mindful relationship with technology?

Aza (1:02:06)
Mm-hmm.

Well,

I think it's very important for everyone to know that even when you know how it all works, it doesn't make you immune.

Obviously, I invented infinite scroll. I know exactly what it's doing. And yet I can still get caught doom scrolling.

yeah, just laying in bed and watching stuff go by. It feels like I'm educating myself, but really what I'm doing is making myself more disconnected.

Siara (1:02:43)
Mm.

Aza (1:02:44)
so I want everyone to like really hear that because the first place we all go, think is self blame. Like I don't have enough control and that's just not the way to go. There are tens to thousands of engineers on the other side of every screen you look at with a supercomputer trained to try to find and exploit your vulnerabilities. And so there's a lot of compassion we need to have for ourselves.

Siara (1:03:07)
Mm.

Aza (1:03:12)
and then an understanding of how to set up our environments so that

we don't have to use our willpower. So I totally use the apps like the Zarios and whatnot that just add more friction into the usage of my phone. I actually have coded up an inset I had that just adds random delays to

to apps as I scroll. So it's sort of the antidote to infinite scroll is every company knows that how fast their app loads or their page loads is directly proportional to how many people use it. Right? Amazon found for every 100 milliseconds, their page loads slower, they lose 1 %

of revenue. Right? that's 100 milliseconds is less than half of human reaction time. So it's almost imperceptible. And yet they lose 1 % of revenue.

Siara (1:04:00)
Peace.

Aza (1:04:07)
Google found a very similar thing. And so I wrote myself a little tool that the more I scroll, the longer I scroll, the more little delay it adds. hundred milliseconds here, 300 milliseconds there, 800 milliseconds here. And what I find is just that a little amount of friction gives my brain the chance to catch up with my impulse. And I'm like, ah, do I really want to be here? It's sort of like being on a airplane and using crappy wifi. And you're like trying to check Twitter and you're like, do I really want to be here?

Siara (1:04:16)
Okay.

Aza (1:04:37)
Now I'm going to go do something else. It's that moment of pause and reflection, friction that enables me to change my behavior without having to use my willpower. And that's, think, the biggest suggestion I have for everyone out there is find the ways to set up your environment that it is not your willpower versus the tech companies, because eventually you'll lose. Find some other things.

Siara (1:04:47)
Mm.

Mmm. Please make that widely available. That's genius.

Aza (1:05:05)
Thanks. I've

tried. I've tried to work on making the app. just, well, I'll put it out there. If there's somebody out there that wants to take that idea and run with it, I don't have time to do that. But I would love somebody out there to make it. Please turn it into a business and I will be a huge supporter.

Siara (1:05:15)
Perfect.

That's awesome. so coming off the back of that, I ask every single guest this, how you interpret it is completely up to you. Is there something that you are logging out of this year and what are you logging into this year?

Aza (1:05:42)
I logging out of an

For the last, I don't know, seven, eight years, I sort of have woken up with the mentality of, want to do everything I can to help save the world. And that's often come at the cost of getting to savor the world, spending time with the people I love most, because I'm out there doing stuff. And I think a big realization I've had, besides for the fact that, just to state the obvious, like,

No one person can save the world. Like that's not a thing. But there's an orientation to like showing up in service that I mean when I say that. And I think I'm going to log out of that kind of like dogged pursuit of save the world over savoring the world. And instead, I think the thing I'm logging into is finding the synthesis of the two that

Siara (1:06:18)
Yeah.

Aza (1:06:45)
There is some way of acting that is at the same time both saving and savoring. And the thing I'm really trying to sit in is that in fact, to do the work of showing up in service, well, you have to do both save and savor at the same time because it is not just enough to think about the better world we want to live in. We have to be living in being the better world we want to live in.

Otherwise we're just doing intellectual gymnastics. It's like trying to like make the better world with salad tongs. It's just not going to be very effective. And so I think there is a new orientation I have, which is trying to find the place where I am authentically coming from getting to live in that more beautiful connected space, including like by building technology so that when I show up in service, it's like fully integrated.

Siara (1:07:21)
Mm.

Mmm.

Aza (1:07:43)
Because I think if we are ungrounded as we try to do our work, we will make the world more ungrounded. And that's what I mean. There's a full synthesis of how we have to show up.

Siara (1:07:53)
Mm-hmm.

Yes, absolutely. That is beautiful. Thank you so much, Aza. We have all really appreciated your wisdom.

Aza (1:08:03)
Thank you so much, Siara It was such a pleasure talking with you.

‍