Gentlemen, boys and girls, comrades and compadres, welcome to the Moynihan Report. I warned you last week that we'd be doing a couple of interviews this week, sort of longer form, interesting interviews. And I'm extremely excited about another one where the person sitting across from me is far more accomplished than I am, and I cannot go through your exhaustive resume, but Meredith Whittaker is the CEO of Signal, president of Signal. See, I didn't even—that's how dumb I am. That difference doesn't even make sense to me. You run... yeah, you run Signal. You were at NYU, right? The AI Institute there.
We were talking before we started about a lot of the negative coverage that you see of AI, a lot of the sort of doom and gloom, jobs are going away. And I just wanted to start with, you know, I find a lot of positive in it. You know, I use AI for a variety of things. It's taken me a long time to figure out how to appropriately use it and get good things out of it. But what would you say if somebody came to you and said, 'Well, this is a cataclysmic technology that's going to destroy the Earth.' Give me a positive reading on AI. And we can get to the negative stuff, which I'm sure you have a lot of that, too.
I mean, I'm going to slice this question a bit and just rewind it, right? Like, AI is not a technical term of art. AI is an aspiration. It's a marketing term.
It has meant a lot of things over time. I'll give a little of my little potted history here. In 1957, cognitive and computer scientist John McCarthy invented the term for two reasons, and he admits this in subsequent interviews. One, he wanted grant money, and you know, the mid-century US government was giving out a lot of money because technological supremacy was one way to... Cold War stuff, empire, etc., etc. And he wanted to exclude his rival, Norbert Wiener, from the convening he was looking for grant money for. And Norbert Wiener had coined the term cybernetics, and he was like, 'Well, if I use cybernetics, I'm Norbert's disciple. If I use AI, this is my thing. I can cook, right?' But at the time, McCarthy's version of AI meant something called the symbolic approach, symbolic systems, which was directly opposed to the deep learning approach, which McCulloch and Pitts had proposed or created in the early '40s. And so I offer that like tiny little perhaps sidebar of an explanation. Like, this term has meant many, many, many, many things.
Now, right now what we're talking about when we talk about AI is this deep learning-centric approach, sort of built on the neural approach that McCulloch and Pitts proposed. It's gone in and out of fashion largely because it is dependent on huge amounts of computational infrastructure. So this is the Nvidia, this is the massive, you know, the trillions, a quadrillions. We're making up numbers that sound like children invented them now. You know, the infrastructure investments that are happening and massive amounts of data, which sort of comes out of a very particular business model that emerged when the internet was commercialized in the 1990s.
So, my positive spin, and you may or may not like this, is this is not the only approach to AI. This massive, centralized, big data-centric, deep learning approach, which can do some things. It spits out interesting prompts. It gestures at approximate answers that may or may not be correct. It can find patterns in large amounts of data. And if that data is meaningful and your distribution matches the context you want to get answers about, those patterns could be very meaningful. But it's not the only way to do AI. And I think as we hit some of these thresholds, the energy use, you hit the issue of integrating these into sensitive systems where you need precise answers, not sort of approximate answers created by a predictive system, etc., etc. We're going to see more attention to the many other approaches to creating systems that behave in one way or another, or at least we perceive they behave in intelligent or human-like ways.
When people... I mean, the AI Institute, and you write and talk about this stuff. I like to sort of look at it from the viewer or listener's perspective, who maybe uses ChatGPT but doesn't know much more about it than that, Gemini, etc. And there is a sense amongst people like, you know, my mother for instance, to be like, 'Well, this is going to kill all the jobs, isn't it?' And I interviewed... I like almost everybody that I interview with one exception, and I'm never shy about saying...
Sam Altman, who I think is a loathsome person for a lot of reasons. He was just a real dick to me, too. Wow. No, it's on my Instagram. I posted videos of it. He's such an [__]. But the entire interview... Sam, shape up. Yeah, well, the entire interview, and this was like 2018 or '19, was about how he was in favor of UBI because the technology he was creating...
It's like Hayek's welfare.
Yeah, exactly. Well, this is what it is. It's like, I mean, this is Austrian school welfare. And he was like, 'It's going to destroy...' I mean, this is the guy who's running OpenAI sitting across from you telling me, 'Oh, the thing that I have created is going to be so disruptive, we're just going to have to pay people because their jobs will be eliminated by the technology that I've created.' Which was an interesting conversation. What do you make of that kind of vision of the future, that AI is going to replace so many jobs that there are going to be indigent people on the dole because all the stuff that they did previously can be done by computers?
That is a tangled nexus. Yeah, no, I... Sorry. I like because I want to approach a couple of different kind of claims and things that are going on...
That Sam Altman's an [__].
Yeah, yeah, yeah. No comment on that at this moment. But on the, you know, first, if you look at sort of the history of computer technology, and I've actually done primary research on this, it tracks back to desires for rationalization and labor control. Like, you know, Charles Babbage very admittedly created the archetypes for his analytical engine, which became the blueprints for classical computation, basically the template for what computers were up to the '70s with like punch cards, etc.
Well, he wanted to rationalize, to control, to standardize mathematical labor. And he was using techniques that were taken from sort of systems of factory management and worker surveillance, etc., to do that. So, like interesting tidbit, but you can sort of pull a lot of that through history and realize kind of the desire to not necessarily eliminate, but to control and degrade work in a certain way, whether it be the sort of assembly line speedups where you're asking the same individual to do more work faster because you have a machine that is now sort of calibrating the pace of labor, up to these sort of hyper-specialization that means no worker owns their skills in a way that they actually have leverage. So they're just sort of pushing a button or logging into an app in a way that they can't sort of demand a wage for. There's a whole history here that predates this modern deep learning, large-scale AI approach.
And I think what you get when you look at that, you broaden the aperture and look at that full picture, is a recognition of sort of what I would call the asymmetry between those who are actually the users of AI and the subjects of AI, the people on whom it's used. So, yeah, we can all pick up ChatGPT or Perplexity and do our little consumer app stuff, but that's not really the market. That's not where the revenue is coming from. We can almost see ChatGPT, at least the free consumer version, that's an advertisement for services and features that then can be used by those who are invested in...
I know I pay $30 a month because I find it useful, but, you know...
There's a lot of revenue. We aren't seeing profit yet. We're not seeing break even. Where is the revenue coming from? Where's the revenue coming from?
Beyond subscriptions, is there a...
Contracts, licensing APIs to third parties. So, GPT, you know, I don't have a number here and I think it'd be great for us to get an audit of this. But if you look at the AI startup industry, the large majority of the things we're calling AI startups are white-labeling an API for GPT. So they sign a contract with Microsoft, which, you know, OpenAI is sort of a part of Microsoft for all intents and purposes, to use their compute infrastructure, their Azure...
Yeah, for people who don't know, the API is the brain that you're kind of...
The plug, the socket you plug something into that then...
When it's created by OpenAI or Microsoft or something, yeah.
So like basically they are white-labeling, like re-skinning what OpenAI has created for a given market or what have you. But they're, you know, barnacles on the hull of the big AI player's ship. They are not independent actors sort of being market challengers there. And that's an important distinction. So it's, you know, this is why the push for massive government contracts. This is why the debt structuring is so weird because it's kind of circular because they haven't really found that killer consumer app. So it's, you know, there is a question of like, where does break even come from for an industry where the capex is so exceptional? And what we're doing is investing in a highly depreciable asset like these servers and chips that themselves are sort of updated every year or two in ways that mean, you know, we've seen data center build-outs that are stopped mid-build-out. The plans are raised. The construction is torn down because they realize, oh, that architecture for the data center isn't going to support the new H100s. It doesn't support the new breed of chips and we'll have to put in new cooling systems. We'll have to put in a new architecture for this building to do that. So let's just tear it down and build something else.
But how... I mean, you look at this stuff from a not a political perspective, but just from, you know, the administrations dealing with China and tariffs. And you read into these stories and if you just have a sort of broad knowledge of this, you say, oh, there's a big negotiation about access to Nvidia chips. And just in the sense that this is not... Oh, this is interesting. There is the AI technology and we kind of get it, it pervades almost every aspect of foreign policy, of domestic policy in a way that in the first kind of dot-com boom in the '90s where all these junk companies are getting tons of venture money and it's like, what does this all mean? I mean, that was a very separate thing, whereas now the integration of all this stuff with governments and with policy... I'm only realizing this, and this is my job, in the past couple of years of how important this stuff is in almost everything we do when it comes to governing and when it comes to foreign policy and tariffs and all this stuff. That's a big shift in a very short period of time.
It is. It is. I think you do have a legacy we're still living from the '90s in which the Clinton administration was one, very clear about its motivations for privatizing network computation, for sort of creating the internet industry in a regulatory way, right? And sort of economically, right? So they were coming off the backs of stagflation. They needed an economic miracle. Privatizing the internet and sort of reaping the economic gains of turning that infrastructure, sort of the pipes and the protocols and everything over to industry was meant to solve that economic hardship without redistribution, right? So there was a sort of tagline from that time, you know, it's the New Deal without the socialism.
And there was a very clear motivation within that that sort of played out around the crypto wars and some of these other debates around technological standards and who should be able to deploy and access tech, basically to ensure that the US was the center of the tech industry by exporting its networks, exporting its standards, exporting its internet companies and their reach beyond the US using soft power and other mechanisms to do that. So there has been an understanding that predates the internet, back to the telephone, back to the post office, that communication networks are inherently centralizing. They're sort of natural monopolies, right? You have to have one standard, they have to work together, and ultimately that's a sort of central node of control.
I think we built up the internet industry with the US as the center of gravity. China, having experienced a century of humiliation and kind of having a large enough internal market, was like, no way, you're not sort of getting a toehold here. So you have two poles geopolitically of the internet industry, the platform industry, the infrastructure industry on which the entire AI industry is built, and that's the US with their hyperscalers, and then that's to a lesser degree China. And if you look at what these same companies, these big tech companies provide, they're incredibly meaningful from a sort of soft power or even hard power geopolitical perspective. So you have the infrastructure that everything runs on, the servers, the clouds. And three US companies have about 70% of the global market. That's Amazon, Microsoft and Google. And then there's China sort of competing.
And we see what happens when the Amazon cloud goes down for half a day.
You can liken that to an on-off switch.
Yeah. Right? It's like, I mean, it was hard for me. I had things that I was trying to do that I couldn't do that day because the cloud went down.
That's like the service, but that's also our government. That's our core infrastructure. That's a really staggeringly profound choke point, single point of failure. Then you have built on top of that is our information environments. So these are the platforms. This is the social media, right? YouTube, Instagram, but also things like Google search, right? The, you know, what creates the shared reality that we swim in. Obviously since time immemorial, states and those in power want to control that type of information. So these are assets.
And then on top of that, I would characterize AI and these large-scale models and the APIs that are being licensed and the government contracts for use of AI as effectively our decision-making capabilities. So, you know, you sign a big shiny contract with Google, you integrate Gemini into your workflow at your corporation, but ultimately what you're doing is ceding decision-making authority to a company that can change that algorithm at any point.
Let me ask you a question, but yeah, no, and it's also, we're at that stage now where people are including AI in things that don't need it.
We've been at that stage.
I mean, I'm just noticing. Keep in mind I'm always saying I'm just noticing this stuff. Now you've noticed this for years. I call it verbose Clippy. It's everywhere. It's like, get out. I just need to send no as a response. I don't...
I love when these things get into the media. I was filming something in Russia with Vitalik Buterin, the guy who created Ethereum. And we're walking through a factory. This shows you how engaged I am. I use fiat currency. We were walking through a cable factory in Russia and Vitalik speaks Russian and they're going back and forth and I said, 'Vitalik, what are you talking about?' And he's a very odd guy and he said, well, he's talking about how he needs the blockchain for blah blah blah, but I don't have the heart to tell him that he just needs a spreadsheet. And I was like, yeah, no. Everyone's blockchain, blockchain, and that was for it. Now it's AI and that stuff is everywhere.
But to the point about private control of this stuff. It's very important for your company that governments do not have the keys and cannot seize it, and it's like Signal is the kind of gold standard for this stuff. But when it comes to, you know, Amazon cloud goes offline, it makes our life very difficult. The alternative would be what? Because I mean, we don't want the government controlling these things at all, right?
Well, I mean, I think that is a pretty profound question and I agree with you, right? Like I am wary of the type of centralized control, particularly the type of control that has resources like our entire infrastructure, all of our intimate data at their hands, right? Like whether it be corporate or government. That's my reflexive politics and I think it's...
And these, you know, you don't want 18 telephones on your wall. Each one you have to remember which group of friends to... You know, like it's like, you have iMessage, you have Signal, which of course is your primary messenger. You have, you know, like it's already annoying. And so...
Communication networks, these types of networks are centralizing inherently via the need for interoperable standards and some sort of way of making them all work together. I think the question then is a question of governance, a question of transparency, a question of what are the norms for decision-making around how these things operate and how they are maintained and what is the objective function that is sort of calibrating the actions and decisions of those who ultimately control that.
And right now there are a handful of US companies that sort of swing in the wind, one administration they go one way, the next administration they go the other way. I worked at Google 2006 to 2019, so I saw the Obama elections. I saw that, you know, like you got your policy shop, you hire in the people who are as close as possible to the new administration. Hopefully they're like dating them or something. So you got all that. You sort of refresh your messaging to appeal to those in power. And why are they doing that? Really simple, because ultimately their goal at the end of the day is growth and revenue. And other metrics forever, indefinitely.
Yeah, maybe I'm framing it wrong because it's a really interesting point that in 2014, I was working at Newsweek Daily Beast. I remember pitching to Tina Brown a story because I had gone to see a friend on the Google campus and he was running the policy shop, Swedish guy.
Yeah. Great guy. One of the smartest guys I've ever met.
You know, it's... but I'm not going to say it, but...
Yeah, you can say that to her. He's very smart. I will say incredibly...
I know. Yeah, well, yeah, yeah, yeah. And for you, is that neoliberal's not... it's fine. She thinks it's bad, she went to Berkeley. But...
No, no, no. I think it's bad because I don't think capital flight should be able to discipline democratic governance, but anyway.
That's the last half of the show.
But it was... I was saying to her, I was like, you know, they have a State Department. You know, this stuff is happening when it came to, you know, this Innocence of Muslims film that was blamed, the stuff that happened in Benghazi, and the government is interfacing all day with Google. With their own... and you would say this to the Google people, and not Nicholas by the way, and then wink and say, we don't say that we have a State Department that we interface in that way. But I'm framing this to you as there's private companies and then there's government.
We're kind of at a point where those have merged in a lot of ways, right? I mean, I think Citizens United kind of structurally merged them in a way that can't be...
Sure, but I mean...
But in a tech way of like the data, the stuff that the government doesn't have the capability because they can't get the goddamn DMV website to work in this city. That the interfacing with tech companies is, you know, you have this thing with Elon Musk buying Twitter and then the Twitter files and then this kind of sensationalized thing. It's like, is anyone...
But that flop sweat failure?
Yeah, I mean, it's just like, is this that interaction of government and private companies has been going on for a long time but it's really, really significant now. Am I wrong? Because you know this stuff and I don't.
I don't think... I mean, I think there's more attention to it because the geopolitical significance of the type of control these companies have globally can't be ignored. It is not like Snowden proved this is not new. That was the government, you know, like... and if you go back to, you know, the Crypto Wars describes this debate that went on through the 1990s, it sort of tracked alongside the effort to commercially... moves that were made to commercialize the internet.
And it was basically, encryption is the technology we have to protect privacy and authenticate people online to ensure transactions can happen, right? It's this complicated math. It's very precious. And the debate, you know, before the end of the '90s, it was classified as a munition, and in order to create strong crypto and deploy it, you had to get permission from the State Department. You couldn't export. It was this whole pain in the ass, right?
Like an export license was a...
Yeah. And so there was this debate that was basically like, okay, we can't actually commercialize the internet if transactions can't happen, if confidential information can't be sent. But... and then this is the sort of intelligence agency and others are like, but you can't have regular people having privacy from us. We need interception points. We need to be able to undermine the privacy of communications to, like it's always the pretext, get the terrorist, get the bad guys...
The ticking time bomb scenario and things like that. This is like playing out between kind of like, you know... Wait, is that still playing out with you though?
It plays out every [__] day. It's like the zone we'll get into that, but let me walk you through this historical thing because I think it's actually important to sort of pin down what we're talking about. Basically there's an acknowledgement we need encryption because we can't reap the benefits of this without it. We need to be able to export it because we can't...
No one's buying that stuff if there's...
Yeah, yeah. We can't have the US as the center of power of the tech if we can't export encryption and have people adopt our standards. No one's going to buy software from a US company that has a backdoor for a foreign government. Like, full stop. And then like, well, how do we adulterate it? And there's this sort of thing that plays out, that's still playing out really, this kind of magical thinking around like, well, we'll just get a golden key for the good guys, and obviously that's not possible.
But what you see at the end of the '90s, and Karina Rider is actually the person who's done really good work on this, is basically it was like, okay, this is the detente we reached. We're going to liberalize encryption. Anyone can make encryption. Anyone can ship encryption. You don't need permission. Basically the Signal protocol is an example of liberalized encryption. Just two virtuosic cryptographers wrote it, published it, used it. It's the gold standard. But a recognition that we're also going to endorse a business model that relies on surveillance.
This sort of advertising business model where it's like collect all the data in the hands of private companies. There's no restrictions on privacy, and because you're doing advertising, you want to know your customers, so you're gathering more and more and more data. That sort of set in stone the business model, this surveillance business model we're in now. The government recognizing like, we don't have to have a public fight about this. We can just come around the backdoor, get data that we would not have permission to collect as government because there's sort of laws around that, from private companies, and you know, kind of we're all okay. We reap the benefits. We get the data. We don't have to sort of look like fools debating the nuances of encryption in the public.
And so, you know, I think it was Snowden that really sort of blew up the collateral consequences of those relationships, which is a long but I hope interesting enough way of saying like, I don't think this is new. I think what we're seeing is the dependencies on a handful of US-based and less prominently Chinese-based hyperscalers who have monopoly controls, economies of scale where you can't really just spin up a competitor, are now clear enough that the geopolitical consequences of that can't be ignored.
When you ask the average person, you know, it's always that, well, I don't care what they see. I don't have anything to hide. You know, that is the most common...
I do. I've heard the average person give that line a handful of times.
Yeah, a handful of times. I have plenty to hide, so please don't hack me, but they're... I don't have the time for that.
You know, look, for most people the trade-off is great. I have free services all over the place. I watch an ad, I give them my information. I, you know, I don't have anything to hide from anyone. I get free...
Gmail. I get, you know, I can watch movies on YouTube for the cost of a 45-second ad. Pretty good. You guys are out there talking about privacy all the time. Who gives a [__]? I've got nothing to hide. You know, this is a good deal for me. What do you say to people who say that?
I mean, I have a like, I have a little exercise I will do sometimes, which is, you know, close your eyes.
A mentalist now? I got to close my... Okay, okay, okay.
You'll wake up hypnotized. You won't know.
Eyes are closed and, you know, like, get in your body. Think about this. Okay. Every text message you've ever sent. Every email you've ever sent. Every Grindr message you've ever sent.
What are you implying? I'm a Hinge, whatever.
I, you know, like, all of it.
I'm implying nothing. I'm saying all of it.
All of it is now on a database. Someone put it, they found it, they put it on a database. That ill-advised text you sent your friend about the thing, the like shitty little comment you made about your girlfriend, the like weird flirt you did that you then regretted.
My girlfriend would be very upset when she sees the Grindr messages, but continue.
Well, yeah, but she's going to see it because this is all on a database. And it is, you're sitting in a room, right? And there's a person there who's like, hey, we put it in a database and we're going to send the link to that database. It's searchable. It's summarized by AI. It's really elegant. You can find anything. We're going to send it to your employer. We're going to send it to your entire family. That's, you know, your in-laws, your cousins, everyone. We're going to send it to your church, to your pastor. We're going to send it to your pickleball league. I don't know. All of them. Your best friend, your high school. Everyone is going to get a link to this database where they can immediately control-F and find their name. They can go through everything you've ever said. How many of your relationships would survive that?
And what would you do to prevent that?
Okay, can I open my eyes now?
So you do care. You do care, my friend.
I just QED. Yeah, I just want you to know that the eye-closing thing was useful. I think they're coming for me now because of what I was thinking.
Sorry about the cortisol, but it does work, right?
It works, and I was thinking about, I narrated what I had texted today. And I'm like, oh, I think I'm [__] just from today. I know. Well, you... There you are. Like, it's bad because, you know, it's funny because there was a story this morning, an Epstein story in the New York Times about these messages between Epstein.
No, and it's just like, I don't know, because there was one message that I was texting a friend of mine who was also a journalist, and I said, well, I don't really know what this means contextually, and it didn't make a ton of sense. And I think of like every text that I've sent that I'd need to explain the context of and why it sounds so terrible.
