WEBVTT NOTE Podcast: Logbuch:Netzpolitik Episode: LNP504 The politics of intellectual shame Publishing Date: 2024-10-17T02:32:15+02:00 Podcast URL: https://logbuch-netzpolitik.de Episode URL: https://logbuch-netzpolitik.de/lnp504-the-politics-of-intellectual-shame 00:00:00.005 --> 00:00:01.325 Good morning, Linus. 00:00:01.565 --> 00:00:02.325 Good morning, Tim. 00:00:02.985 --> 00:00:04.105 Linus, what watch? 00:00:04.565 --> 00:00:05.325 Ten watch. 00:00:05.725 --> 00:00:06.545 Such much? 00:00:06.865 --> 00:00:09.105 You'll get along beautifully in America. 00:00:30.245 --> 00:00:34.285 Logbuch Netzpolitik Nummer 503 vom 16. 00:00:34.465 --> 00:00:40.545 Oktober 2024 und ihr werdet es schon gemerkt haben. 00:00:41.385 --> 00:00:43.105 Wir sprechen eine andere Sprache. 00:00:43.585 --> 00:00:47.485 Wir sprechen eine andere Sprache heute, weil wir haben heute einen Gast und 00:00:47.485 --> 00:00:52.025 das ist dann ein Gespräch, was wir heute auf Englisch führen wollen. 00:00:52.285 --> 00:00:56.465 Bevor wir aber kurz unsere Episodennummer korrigieren, ist nämlich 504. 00:00:56.925 --> 00:00:57.645 Was habe ich gesagt? 00:00:57.765 --> 00:00:58.285 503. 00:00:58.405 --> 00:00:58.505 Oh. 00:01:01.145 --> 00:01:03.045 Ja, ausnahmsweise lebe ich mal nicht in der Zukunft. 00:01:03.205 --> 00:01:04.785 Man merkt, wir sind aufgeregt. 00:01:06.905 --> 00:01:11.865 Tim hat ja einen britischen Pass und muss jetzt mal seine Britischkenntnisse unter Beweis stellen. 00:01:11.985 --> 00:01:14.325 Was habe ich einen britischen Pass? Ich habe gar keinen britischen Pass, ich bin Deutscher. 00:01:14.885 --> 00:01:16.165 Okay, aber du hattest ja nicht einen britischen Pass. 00:01:16.205 --> 00:01:16.905 Das ist aber abgelaufen. 00:01:17.065 --> 00:01:19.125 Achso, der ist abgelaufen. Okay, dann das erklärt einige. 00:01:20.785 --> 00:01:25.925 Anyway, we say hello to Meredith. Meredith Whitaker. Hello, welcome to our podcast. 00:01:26.345 --> 00:01:28.405 I'm so happy to be here. Thank you for having me. 00:01:29.427 --> 00:01:37.187 Meredith, you are, in case somebody doesn't know you, the president of the Signal Foundation. 00:01:37.427 --> 00:01:38.667 I am. I am. 00:01:38.867 --> 00:01:42.467 And Signal is a messenger application? 00:01:42.767 --> 00:01:49.167 It is the world's most widely used, truly private messaging application. 00:01:49.647 --> 00:01:54.947 I always have to put the full sentence in there. A very special messaging application. 00:01:55.047 --> 00:01:55.787 What makes it special? 00:01:55.787 --> 00:01:58.847 Um we build against 00:01:58.847 --> 00:02:04.087 the grain to put it in kind of flowery language we go out of our way to build 00:02:04.087 --> 00:02:11.227 for privacy and to get as close to collecting no data at all as possible which 00:02:11.227 --> 00:02:15.967 as you and i assume many in your audience know is a lot more work. 00:02:15.967 --> 00:02:18.047 Than simply encrypting. 00:02:18.047 --> 00:02:22.327 Then well encrypting is part of that but then you need to make sure the encryption 00:02:22.327 --> 00:02:26.027 works and then encryption of message content is Das ist nicht möglich, 00:02:26.227 --> 00:02:30.987 weil du also hast Meta-Data. Du hast auch libraries, du hast du in. 00:02:31.107 --> 00:02:36.047 Du hast Core Services, du hast du mit. So, da ist viel ingenuity required zu 00:02:36.047 --> 00:02:43.487 actually create a system das ist rejecting den normative assumptions der Tech-Industrie heute. 00:02:43.767 --> 00:02:47.547 Wir alle wollen ein ton of Daten, und wir alle wollen monetize das Daten. 00:02:47.607 --> 00:02:52.087 Wir wollen Sie ads mit dem Daten. Wir wollen das zu anderen Kunden. 00:02:52.147 --> 00:02:55.467 Wir wollen das Daten-Broker. Wir wollen das AI-Modeln mit dem Daten. 00:02:56.747 --> 00:03:01.707 Most of the infrastructure in the tech ecosystem now assumes that as a given. 00:03:02.247 --> 00:03:04.327 And then we have to rewrite things too. 00:03:04.867 --> 00:03:10.627 And you negate all these assumptions and try to build a truly private messaging 00:03:10.627 --> 00:03:14.107 platform in a world that has a completely different business model. 00:03:14.427 --> 00:03:17.927 Yeah. Well, I'm going to just flip that to be a little bit rhetorical and say 00:03:17.927 --> 00:03:23.027 those assumptions negate the human right to private intimate communication. 00:03:23.027 --> 00:03:29.107 And we are trying to rebuild a tech ecosystem that actually honors those. 00:03:29.847 --> 00:03:31.407 That's an even higher goal. 00:03:31.627 --> 00:03:31.767 Yeah. 00:03:32.487 --> 00:03:38.387 Before we dig more into the Signal project, which in itself I think is quite 00:03:38.387 --> 00:03:45.267 interesting, we'd like to understand how did you get there, where you are right now. 00:03:45.367 --> 00:03:50.487 So what was your introduction into this digital world? When did it start? 00:03:53.027 --> 00:03:58.807 I don't have one of those stories where I'm like an Atari. I was four. I'm a hacker. 00:04:01.407 --> 00:04:05.327 I didn't care about tech, qua tech. I wasn't interested. 00:04:05.567 --> 00:04:09.507 You remember back in school, there were two kinds of nerds. 00:04:10.157 --> 00:04:12.777 There were the math nerds and the book nerds. 00:04:13.177 --> 00:04:14.117 You were the book nerd? 00:04:14.197 --> 00:04:18.137 I was the book nerd. But in the story we tell now, because nerddom has been 00:04:18.137 --> 00:04:22.877 overlapped with monetary success and sort of a career in math and science, 00:04:22.877 --> 00:04:26.357 I think we forget about the book nerds. But hello, I'm here to remind you. 00:04:26.877 --> 00:04:29.877 We exist. We're here. And that's who I was. 00:04:30.317 --> 00:04:34.257 And so I studied literature and rhetoric at Berkeley, which I just thought were, 00:04:34.857 --> 00:04:35.817 I mean, they're still beguiling. 00:04:36.017 --> 00:04:39.157 Like being able to read and write is pretty fundamental for anything. 00:04:39.677 --> 00:04:43.637 Und dann ich war poor so ich habe einen Job bei Google weil sie waren die ersten 00:04:43.637 --> 00:04:47.777 Menschen zu offeren mich ein Job und dann habe ich sehr fasziniert mit was auf 00:04:47.777 --> 00:04:49.757 der Erde was going on When war das? 00:04:50.017 --> 00:04:54.517 Wenn warst du bei Google als jemand der Studierende Literatur und Rhetorik? 00:04:54.977 --> 00:05:00.897 July 10, 2006 Ich habe mich über Berkeley und ich musste Geld und ich musste 00:05:00.897 --> 00:05:05.557 und ich put mein Resumé auf Monster.com which ist ein precursor zu LinkedIn, 00:05:06.177 --> 00:05:07.377 und ich habe mich überlegt out. 00:05:07.777 --> 00:05:11.717 And then I talked to my friend Michelangelo, because at the time you still needed 00:05:11.717 --> 00:05:14.057 to invite to join Gmail, as I recall. 00:05:14.237 --> 00:05:17.997 And I was like, hey, Michelangelo, can you get me an invite so I can make a 00:05:17.997 --> 00:05:20.777 Gmail address so I can appeal to this recruiter? 00:05:21.117 --> 00:05:25.257 And now my spam-filled Gmail address dates from that moment. 00:05:25.557 --> 00:05:29.117 So I was hired as something called a customer operations associate, 00:05:29.737 --> 00:05:31.677 and I didn't know what that was. 00:05:32.017 --> 00:05:34.677 No one knows what that is. It's just a bunch of words, But it sounded, 00:05:34.917 --> 00:05:37.837 I was like, that's a business job. It sounds like a business title. 00:05:38.037 --> 00:05:42.237 I had no idea. It was basically a customer support person who wrote technical 00:05:42.237 --> 00:05:46.597 documentation, who are kind of user documentation, some technical documentation, 00:05:46.597 --> 00:05:50.637 and answered inquiries about Google's free products. 00:05:50.897 --> 00:05:57.237 And I was doing that for Rightly, which was an acquisition that Google made 00:05:57.237 --> 00:06:00.897 that then they rebuilt to become Google Docs. Okay. 00:06:01.557 --> 00:06:06.397 Und das ist an der Zeit, die ganze Zeit, die Google-Businessen war, was nur Search. 00:06:06.557 --> 00:06:10.677 So Search war all auf der Main Campus. Es war wo die Money kam, 00:06:10.857 --> 00:06:17.757 und dann war das Dinky-little-Building, wo alles, Gmail, Blogger, Rightly, Reader, 00:06:19.197 --> 00:06:22.937 all of them sat und sie waren experimental projects. Und das war called Apps. 00:06:23.829 --> 00:06:28.009 Did you ever have exposure to actual customers or did you just write documentation? 00:06:28.309 --> 00:06:34.449 Well, they would email in. I only did this for a second because I kind of figured 00:06:34.449 --> 00:06:37.049 out I wanted to do other things. And that's part of the story. 00:06:38.169 --> 00:06:42.089 There's a part of the story, actually. So I did answer inquiries sometimes. 00:06:42.329 --> 00:06:47.269 And there were auto responses, like hotkeys you'd use in this really janky ticketing system. 00:06:47.409 --> 00:06:51.029 And this is back. We didn't have laptops. There's no such thing as a smartphone. 00:06:51.769 --> 00:06:54.529 Our desktops were chained to the desk. Und wenn du nach Hause bist, 00:06:54.529 --> 00:06:56.929 wenn du nach Hause bist, wenn du nach Hause bist, dann musst du nachdenken. 00:06:57.069 --> 00:06:58.769 Du hast du nachdenken, das klunky Lenovo oder so. 00:06:59.289 --> 00:07:01.889 Ich weiß nicht genau, aber es war... 00:07:02.449 --> 00:07:07.929 Ich war mir, okay, ich werde rewarded in meinem Job, based auf wie viele Tickets ich close bin. 00:07:08.229 --> 00:07:10.849 Aber ich weiß nicht die Engenheer, die haben zu close diese Tickets, 00:07:10.869 --> 00:07:12.849 weil ich mich reporting bugs, und ich muss sie fix. 00:07:13.049 --> 00:07:15.549 Und wenn sie nicht fix die Bugs sind, dann bin ich nicht rewarded. 00:07:15.669 --> 00:07:17.549 Und ich war mir, das ist wirklich silly. 00:07:18.029 --> 00:07:21.049 So ich war über die Apps-Büte, weil ich in einer anderen Bühne bin bin. 00:07:21.049 --> 00:07:24.869 Und ich mit der engineering team und ich war, hey, ich bin das Person, 00:07:24.969 --> 00:07:26.749 die gibt es diese E-Mails, die Sie anwenden Sie. 00:07:27.249 --> 00:07:30.789 Und ich war able, zu convince einem einen zu gehen, zu Costco, 00:07:31.049 --> 00:07:35.189 die ist dieses große Warehouse-Store, ein Walmart-E-Store in den USA, 00:07:35.889 --> 00:07:41.369 und put ein giantes Kuchen auf seine eigene Karte, und bring es in den Apps-Bilden, 00:07:41.369 --> 00:07:42.569 so ich konnte sit mit dem. 00:07:42.809 --> 00:07:47.149 Weil wenn ich mit dem satte, dann könnte ich mit dem, und dann würde ich mich 00:07:47.149 --> 00:07:48.689 mehr möglich zu fixen, meine Bugs. 00:07:48.689 --> 00:07:53.249 Und so all my bugs started getting fixed but then my manager got really, 00:07:53.969 --> 00:07:57.309 upset because it was like obviously I had breached the hierarchy and I was like 00:07:57.309 --> 00:08:03.009 and so this meeting appeared on my calendar that was basically about me being 00:08:03.009 --> 00:08:06.129 insubordinate but I'm also like I didn't come from that class I had no I was 00:08:06.129 --> 00:08:08.309 like what the fuck did I do wrong I don't know, 00:08:08.849 --> 00:08:12.309 and I was like oh god I like I did something and then. 00:08:13.747 --> 00:08:17.027 The day before I think that meeting was going to happen, this email hit, 00:08:17.147 --> 00:08:22.327 we had this all, like some group, I don't know, it was like the group to the 00:08:22.327 --> 00:08:23.687 consumer operations team. 00:08:24.347 --> 00:08:28.347 And one of the engineering directors in Apps must have had like a beer or something 00:08:28.347 --> 00:08:31.207 during one of the many, many, many drinking parties that happened during the 00:08:31.207 --> 00:08:32.087 day at Google in that era. 00:08:33.887 --> 00:08:39.207 And sent some email that was like, Meredith, this couch is a model for collaboration. 00:08:40.427 --> 00:08:43.367 I was going to say, that's why you hired somebody from Berkeley. 00:08:44.627 --> 00:08:47.347 To write your tickets because they come up with ideas like that. 00:08:47.667 --> 00:08:51.007 And then that meeting just disappeared and I was like, well shit, 00:08:51.147 --> 00:08:54.087 I need another job because I just burned someone who's going to figure out how 00:08:54.087 --> 00:08:56.467 to get back at me and that's where my street sense comes in. 00:08:56.527 --> 00:09:02.107 So I took a bike around campus for a couple months and asked every VP during 00:09:02.107 --> 00:09:05.307 their office hours, which is a thing they used to have like, hey, I want another job. 00:09:05.447 --> 00:09:08.967 And I would just drop in and be like, hi, you don't know me. I'd like a job. 00:09:09.327 --> 00:09:16.467 And then I got a job doing basically Standards work, like trying to push document interop standards. 00:09:17.087 --> 00:09:21.267 And so I started in standards and then standards parlays into measurement pretty 00:09:21.267 --> 00:09:24.047 quickly because measurement is ultimately the methodological standard. 00:09:25.307 --> 00:09:30.467 And co-founded MLab and then that became the nucleus around which the research 00:09:30.467 --> 00:09:34.207 group I founded existed and then that was the... 00:09:35.242 --> 00:09:38.882 That was when the fresh air of sort of these political and social complexities 00:09:38.882 --> 00:09:42.962 started hitting my technical work. And I was like, oh, standardization is power. 00:09:43.282 --> 00:09:47.882 Oh, creating data means you own the narrative. Oh, shit, none of this is neutral. 00:09:47.882 --> 00:09:49.182 All of this is contingent. 00:09:49.302 --> 00:09:52.722 I started seeing the balance sheet. I started seeing the capital that was involved 00:09:52.722 --> 00:09:53.902 in running infrastructure. 00:09:54.462 --> 00:09:59.342 And I remember around the time I met you, Linus, like maybe over a decade ago. 00:09:59.542 --> 00:10:02.702 Yeah, it must have been about a decade or longer. Yeah, probably a bit longer even. 00:10:02.782 --> 00:10:04.542 Yeah, aging is weird. 00:10:06.762 --> 00:10:12.262 Aber ich habe diese schluss-Ballanz-Ballanz-Sheets. Ich habe aus Google gepritten, 00:10:12.942 --> 00:10:19.882 die gezeigt, wie viel die Uplink-Bandwidth und die Infrastruktur und Power-Costs für MLAB war. 00:10:19.882 --> 00:10:22.782 Und ich würde das zeigen, das zu all diese Civil Society funders, 00:10:22.782 --> 00:10:24.922 wie gesagt, Sie können 250k verdienen. 00:10:25.102 --> 00:10:29.562 Es kostet 40 Millionen a Jahr in Bandwidth, nur zu MLAB. 00:10:29.582 --> 00:10:32.962 Du wirst nicht verstehen, die economics of this, denn wir sind in der Era der 00:10:32.962 --> 00:10:35.962 Civic-Technik, All we need is a good idea, right? 00:10:36.302 --> 00:10:41.102 And I think I was really lucky to get sensitized to the political economy and 00:10:41.102 --> 00:10:43.022 the fact that we're talking about infrastructure, 00:10:43.442 --> 00:10:49.542 capital, network effects, economies of scale and not some kind of brilliant 00:10:49.542 --> 00:10:53.622 idea that just ephemerally transformed our world, 00:10:54.142 --> 00:10:58.142 and that our side just needs to wait to have one of those to get our turn. 00:10:58.422 --> 00:11:02.342 That sounds like a quick upgrade from a customer support person. 00:11:02.882 --> 00:11:08.102 Yeah, yeah, yeah. So, you mentioned MLab, which stands for Measurement Lab. 00:11:08.382 --> 00:11:08.962 It does, yeah. 00:11:09.022 --> 00:11:13.662 Can you explain what this is all about and how it came alive and why you were involved? 00:11:13.982 --> 00:11:22.322 Yeah. It was a project that was Derek Slater, Vince Cerf, Stephen Stewart, 00:11:22.982 --> 00:11:27.062 Sasha Meinrath at Open Tech Institute. 00:11:29.005 --> 00:11:30.345 The elders of the internet. 00:11:30.825 --> 00:11:33.285 Yeah, some old internet guys and me. 00:11:35.585 --> 00:11:37.165 Let's say the elders of the internet. 00:11:38.705 --> 00:11:39.965 You're one of them now. 00:11:41.305 --> 00:11:47.745 And the lady with the couch. And the conceit there, which felt really simple 00:11:47.745 --> 00:11:50.725 to me at the time, is everyone is buzzing about net neutrality. 00:11:51.285 --> 00:11:56.545 And I was a Kool-Aid drinker. I still think the value underlying that kind of 00:11:56.545 --> 00:12:02.445 mythology, let's say, like, yes, we should not have one gatekeeper deciding 00:12:02.445 --> 00:12:06.125 which news source, right? This is old school common sense. 00:12:06.245 --> 00:12:09.025 It goes back to Western Union, who there's a little, you know, 00:12:09.185 --> 00:12:15.905 in the US would refuse to carry telegrams from political candidates the company didn't support, right? 00:12:15.965 --> 00:12:19.085 So there's a real bedrock precedent here. 00:12:19.785 --> 00:12:22.145 And I was like, yeah, of course we need net neutrality. 00:12:23.205 --> 00:12:27.625 Neutrality itself is a really wafty, it's a loose concept. 00:12:27.805 --> 00:12:32.605 It needs to be augured in some type of benchmark against which we can assess is this neutral or not. 00:12:32.605 --> 00:12:33.365 You need to quantify it. 00:12:33.365 --> 00:12:39.205 Yes. And then you just flung into the abyss of philosophy the second you try to do that. 00:12:39.285 --> 00:12:45.905 But we started and I would say it was not necessarily built to succeed. 00:12:45.905 --> 00:12:49.785 It was built as a sort of hypothesis project where 00:12:49.785 --> 00:12:52.905 we stood up three servers that hosted 00:12:52.905 --> 00:12:56.865 open source measurement clients and the thing 00:12:56.865 --> 00:13:02.005 that we were doing was putting servers that were all configured the same that 00:13:02.005 --> 00:13:08.865 gave each client a dedicated slice of resources and then way over-provisioned 00:13:08.865 --> 00:13:13.905 the uplink between the server and the switch so that we could all, 00:13:15.005 --> 00:13:20.145 for all intents and purposes, guarantee that any artifacts that were detected 00:13:20.145 --> 00:13:24.045 through the measurement methodology like some TCP, 00:13:24.405 --> 00:13:29.625 RTT or something, were not interfered with by our infrastructure and weren't 00:13:29.625 --> 00:13:30.925 suffering for bandwidth. 00:13:31.205 --> 00:13:36.625 So for practical reasons you had unlimited bandwidth and zero latency to everywhere. 00:13:36.925 --> 00:13:41.705 Yeah, and let's, you know, and that's very expensive. But then, 00:13:42.465 --> 00:13:47.805 we immediately DDoS the servers or like, oh shoot, because Vint's put a blog 00:13:47.805 --> 00:13:49.745 post out and he was you know he's a big name, 00:13:50.225 --> 00:13:55.105 and then it was like this job of years and years of getting servers in different 00:13:55.105 --> 00:13:59.145 interconnection points because we wanted to measure across consumer you know 00:13:59.145 --> 00:14:02.745 we're not just measuring to a server in the network exactly and we need to cross 00:14:02.745 --> 00:14:06.085 interconnection points to do that because that's where you begin to see interesting 00:14:06.085 --> 00:14:08.225 business relationships and kind of feuds, 00:14:09.285 --> 00:14:15.005 and it was just it was very it taught me about like how difficult and contingent and, 00:14:15.890 --> 00:14:23.610 ultimately subjective, creating data is, and the political process of then sort 00:14:23.610 --> 00:14:27.750 of defending that data as a reliable proxy for reality. 00:14:27.910 --> 00:14:31.810 Because I would be at the Federal Communications Commission and I had Comcast 00:14:31.810 --> 00:14:37.710 across from me being like, no, we want to measure it multi-threaded TCP versus 00:14:37.710 --> 00:14:41.590 single-threaded because you get a, let's just say, higher number that way. 00:14:41.590 --> 00:14:45.750 It's more forgiving. and we would be sort of defending the methodology of our 00:14:45.750 --> 00:14:48.550 tests and the openness principles. 00:14:48.830 --> 00:14:53.650 So it was open source code, the database architecture was open. 00:14:53.650 --> 00:14:58.970 I don't remember all the server architecture, it was old school, 00:14:59.810 --> 00:15:00.990 old school ideological. 00:15:00.990 --> 00:15:05.610 You really don't sound like a book nerd now. You've become a real network nerd. 00:15:05.710 --> 00:15:09.730 Let's just say book nerds read and network stuff is written down. 00:15:13.190 --> 00:15:19.070 I read the fucking menu of it. I'm like, no one told me I was allowed to not 00:15:19.070 --> 00:15:20.590 read it and still have an opinion. 00:15:24.710 --> 00:15:26.250 I love networking. 00:15:26.250 --> 00:15:31.050 It taught me everything about so much because you get down to the network layer 00:15:31.050 --> 00:15:34.910 and you're like, fuck, everything is broken. It's all tied together with string. 00:15:35.350 --> 00:15:41.610 What's going on? Okay, so your promises in the pitch meeting are guaranteed by this. 00:15:42.530 --> 00:15:47.010 This crusty old library that's maintained by a guy on a boat with a BitLocker 00:15:47.010 --> 00:15:48.470 account. Like, what is going on? 00:15:51.590 --> 00:15:52.810 It's scary, isn't it? 00:15:53.090 --> 00:15:57.970 It is. I mean, it's a catalyst into today for my work. 00:15:59.530 --> 00:16:03.570 You did that for, I think, a couple of years, the MLab stuff. 00:16:03.810 --> 00:16:08.830 And you also, at this point in time, I think, began, you know, the ties to the, 00:16:09.743 --> 00:16:16.723 Let's say broader internet freedom community. I remember Uniprobe at the time, I think, was another. 00:16:17.683 --> 00:16:19.723 They still exist, right? They do. 00:16:19.923 --> 00:16:23.383 And we actually talked to them at Signal. They'll detect signal blocking. 00:16:24.003 --> 00:16:29.283 And I believe, so one of the things MLab did was provide backend infrastructure 00:16:29.283 --> 00:16:30.923 for projects like Uniprobe. 00:16:31.023 --> 00:16:33.703 And I don't actually know what that relationship is today. 00:16:35.743 --> 00:16:42.323 So academics and hackers and developers could write a test or kind of a measurement 00:16:42.323 --> 00:16:44.363 methodology deploy a client, 00:16:45.083 --> 00:16:49.883 to consumers like you test from your laptop or whatever and then we would be 00:16:49.883 --> 00:16:54.823 the ones paying for the bandwidth and infrastructure that would allow that to scale Uni. 00:16:54.823 --> 00:16:58.263 Is the Open Observatory of Network Interference. 00:16:58.263 --> 00:17:01.683 If I remember the acronym Hi Arturo So... 00:17:04.883 --> 00:17:10.143 Disclosing a whole social network here voluntarily. So, but eventually something 00:17:10.143 --> 00:17:14.283 happened that is known as the Google Walkouts. 00:17:14.503 --> 00:17:19.563 Was that following your MLab activities or during your MLab activities or did 00:17:19.563 --> 00:17:25.303 you rise even higher in the Google hierarchies before that happened? 00:17:25.563 --> 00:17:29.823 Well, I mean, there are many. I think, so... 00:17:30.806 --> 00:17:34.146 I was not working on MLAB by that point. 00:17:34.326 --> 00:17:40.926 MLAB, there's a part of this story that I didn't cover, which is around, 00:17:40.926 --> 00:17:43.226 I don't know, 2014, 2015, 00:17:45.086 --> 00:17:48.586 we started seeing really interesting artifacts in the data, 00:17:48.906 --> 00:17:55.066 which showed essentially that at particular interconnection points between particular 00:17:55.066 --> 00:18:00.106 telcos, we were seeing drastic drops in performance. 00:18:00.866 --> 00:18:06.306 And what we were able to do was look at, you know, every intersection of Telco 00:18:06.306 --> 00:18:09.146 1 and Telco 2, we see this drop. 00:18:10.026 --> 00:18:13.766 Intersections of Telco 3 and Telco 2 don't see this drop. 00:18:14.266 --> 00:18:17.906 And so what we're able to do is say, like, there's actually a business feud 00:18:17.906 --> 00:18:22.026 going on. And the interconnection point is the locus of that feud. 00:18:22.186 --> 00:18:24.846 And using traceroute data, we were able to say, like, look, these aren't, 00:18:24.986 --> 00:18:30.146 it's not just that they're sharing a path in some other region of the network that's slowing it down. 00:18:30.806 --> 00:18:32.506 They're actively throttling. 00:18:32.766 --> 00:18:37.566 They're actively throttling. And we put together a report on that. 00:18:37.786 --> 00:18:42.486 And this is where I was sort of, I think, kind of sideloading myself into some 00:18:42.486 --> 00:18:44.966 academic research work. I was like, okay, how do we do this? 00:18:45.786 --> 00:18:49.806 Document our methodology. Document sort of everything was open. 00:18:49.826 --> 00:18:51.066 So we pulled that together. 00:18:51.246 --> 00:18:56.566 And it showed, I don't know if you remember, there was a sort of Comcast Cogent Netflix. 00:18:56.806 --> 00:18:57.166 Oh, yeah. 00:18:57.166 --> 00:19:01.166 Netflix was shoving all its traffic through Cogent. What we had done is expose 00:19:01.166 --> 00:19:05.686 that and exposing that it exposed the principle that you have to if you wanted 00:19:05.686 --> 00:19:08.966 to ensure net neutrality, you have to take the interconnection points and the 00:19:08.966 --> 00:19:11.326 interconnection agreements into account. 00:19:11.566 --> 00:19:16.326 And that led Obama to add interconnection to the reclassification under Title 00:19:16.326 --> 00:19:21.046 II of these, you know, which kind of moved toward net neutrality that was nullified. 00:19:21.046 --> 00:19:27.166 But, you know, that was kind of the swan song, let's say, for my MLab time, 00:19:27.346 --> 00:19:29.546 because, of course, that was a huge deal, right? 00:19:29.586 --> 00:19:32.826 Like, that's where the business model rubber hits the road. 00:19:33.586 --> 00:19:37.006 And, you know, I'll just shorthand it. Like, that bought me a lot of capital at Google. 00:19:37.786 --> 00:19:41.246 And with that, I was, you know, I was already interested in a lot of things. 00:19:41.406 --> 00:19:44.926 MLab was sort of humming along and, like, it had grown from a hypothesis project 00:19:44.926 --> 00:19:48.546 to, like, a global thing that was working and doing stuff. 00:19:48.546 --> 00:19:53.726 And I started getting agitated by AI and a lot of these privacy and security 00:19:53.726 --> 00:19:56.866 concerns, being part of the community that you were part of, 00:19:57.046 --> 00:19:58.186 you know, kind of thinking around like 00:19:58.186 --> 00:20:01.986 tech alternatives and getting less comfortable with the business model. 00:20:03.226 --> 00:20:07.886 And from there I always had like eight different projects going on but I went 00:20:07.886 --> 00:20:13.586 on to found, to co-found the AI Now Institute which was really trying to bring the conversation, 00:20:14.326 --> 00:20:18.286 on what we called machine learning back then but like AI was like a flashier 00:20:18.286 --> 00:20:23.326 term, like bring it down to the ground a little bit and stop talking about superintelligence 00:20:23.326 --> 00:20:24.826 and stop talking about, you know, 00:20:25.826 --> 00:20:29.866 Political economy, what are these technologies, how are they being deployed, 00:20:30.046 --> 00:20:34.426 how do you oversee them, who uses them on whom, and what are the social and 00:20:34.426 --> 00:20:35.706 political dynamics of that. 00:20:36.106 --> 00:20:41.046 But the AI Now Institute is not hosted at Google, is it? No. 00:20:41.586 --> 00:20:43.046 Which university is it again? 00:20:43.246 --> 00:20:48.226 It was at NYU, and it's moved out of NYU for reasons of that being easier. 00:20:48.546 --> 00:20:49.026 Okay. 00:20:49.446 --> 00:20:52.506 And cheaper, because they take like 40%, those universities. 00:20:53.106 --> 00:20:53.346 Okay. 00:20:54.666 --> 00:20:56.986 Word to the wise, your money. 00:20:59.506 --> 00:21:00.586 What do you think? 00:21:00.786 --> 00:21:03.106 Like whatever donations you get. Okay. 00:21:03.426 --> 00:21:06.866 Yeah. So it's an expensive brand. 00:21:10.046 --> 00:21:17.106 Well, for a reason, right? I mean, NYU Institute, that's something to have, I think, co-founded. 00:21:17.746 --> 00:21:22.366 Co-founded, yeah. I mean, the work was really good. And this was, 00:21:22.446 --> 00:21:26.026 again, gaining the capital at Google, cementing a reputation, 00:21:26.206 --> 00:21:28.366 being able to get to a level where I had a budget. 00:21:28.646 --> 00:21:33.946 And then part of what I was always trying to do is how much can I pull out of 00:21:33.946 --> 00:21:37.806 Google and get into the hands of people doing work that I think is cool. 00:21:38.006 --> 00:21:42.946 How do we carve tributaries in the massive river of this huge, 00:21:42.946 --> 00:21:46.026 rich company and get it out? 00:21:46.266 --> 00:21:50.386 So you started that institute at the end of 2017. 00:21:51.906 --> 00:21:55.026 Well, beginnings are sometimes hard to date. 00:21:55.246 --> 00:22:03.226 It was born out of a request from the Obama administration to be the host of one of his AI summits. 00:22:03.286 --> 00:22:06.606 I don't remember exactly the contours there, but it was in 2016. 00:22:07.166 --> 00:22:13.506 And so it started as like, like the idea then was like, let's do this and let's do it big. 00:22:13.646 --> 00:22:18.626 Like, let's make this the most polished, flashiest, like hard to ignore kind 00:22:18.626 --> 00:22:23.146 of spectacle using all the tools we can get from like hiring an events agency, 00:22:23.326 --> 00:22:24.666 doing good press, all of that. 00:22:24.666 --> 00:22:31.586 But let's also make this the one that is the most engaged with these political topics, 00:22:31.586 --> 00:22:36.086 that's actually forcing the debate in that direction and kind of making it face 00:22:36.086 --> 00:22:39.886 these questions that are much more grounded and, you know, 00:22:40.046 --> 00:22:43.386 hopefully much more healthy for our population. 00:22:44.401 --> 00:22:50.561 So it was very successful and from there we got offers of funding and a lot 00:22:50.561 --> 00:22:53.661 of encouragement and the work just kept going. 00:22:53.661 --> 00:23:00.141 Well, we in Europe, we are used to be a bit behind, but if I can recall 2016, 00:23:00.141 --> 00:23:04.321 there was no discussion about AI in Europe at all. 00:23:04.741 --> 00:23:10.061 And so, it's quite interesting to see that Obama has actually, 00:23:10.281 --> 00:23:16.361 you know, decided that it's finally time to do something about this topic nobody 00:23:16.361 --> 00:23:18.101 has ever really talked about. 00:23:18.321 --> 00:23:21.401 Germans are currently making up their mind whether that is a topic. 00:23:21.401 --> 00:23:25.081 Ja, wir sprechen viel über Europa heute. 00:23:25.341 --> 00:23:33.801 Aber mein Frage ist, kann man describe was kind of discussions going on 10 Jahre 00:23:33.801 --> 00:23:41.781 ago in American society that this has come up to be something of a topic for the near future, 00:23:41.921 --> 00:23:44.441 which actually became a big one. 00:23:45.941 --> 00:23:48.721 Where did this discussion take place? 00:23:49.681 --> 00:23:54.901 Yeah, to answer this question, I'm going to be drawing on a lot of the research 00:23:54.901 --> 00:23:58.841 and the historical work, the work I've done since then. 00:23:58.961 --> 00:24:03.341 Because when this dawned in my life, when it started being a thing, 00:24:03.681 --> 00:24:07.341 I had basically that same question. What is this stuff? 00:24:07.481 --> 00:24:11.341 Why is it kind of... At Google, you would see... 00:24:12.464 --> 00:24:18.264 A shift toward a new paradigm or a new trend by the incentives that were structured 00:24:18.264 --> 00:24:20.444 into the OKRs, the quarterly goals, 00:24:21.024 --> 00:24:23.924 that kind of, you know, there'd be all these training modules that would pop 00:24:23.924 --> 00:24:28.024 up and it's like make your software engineer into an AI developer, 00:24:28.244 --> 00:24:30.384 you know, and you'd be like, there's an incentive here. 00:24:30.504 --> 00:24:32.744 Seems like they weren't too successful though, right? I mean, 00:24:32.784 --> 00:24:36.984 that's not like the biggest AI wave at Google. 00:24:37.464 --> 00:24:40.444 Well, they were, that was DeepMind era. So, okay, okay, good. 00:24:40.584 --> 00:24:40.704 Sorry. 00:24:40.704 --> 00:24:46.884 I think they were ahead. It was them and Meta for a long time. And the history... 00:24:46.884 --> 00:24:48.684 Didn't make it to the business though at the time, right? 00:24:48.884 --> 00:24:51.684 No, but they're chaotic. It's a kind of court in decline. So that, 00:24:52.164 --> 00:24:56.764 you know, actually where the... Actually the business model part has never been 00:24:56.764 --> 00:24:58.644 their strong suit beyond search. 00:24:58.944 --> 00:25:03.964 To be real, you know, like cloud is like the best technology presented confusingly 00:25:03.964 --> 00:25:07.204 with 18 versions all deprecated, right? Like that's... Um, 00:25:09.244 --> 00:25:16.444 um, but this, the, the AI stuff was, so if you look at the, if you look at the 00:25:16.444 --> 00:25:19.404 sort of recent history, which is something I'm spending, I've spent a lot of time on. 00:25:19.504 --> 00:25:22.644 I spend a lot of time on because I think we, you know, that gives us a really 00:25:22.644 --> 00:25:27.244 different picture than the Elon Musk narrative or the kind of popular narrative. 00:25:28.224 --> 00:25:33.784 There's a very important paper that was published in 2012 that introduced the 00:25:33.784 --> 00:25:39.584 AlexNet algorithm. und das war Jeff Hinton und his students, you may recognize. 00:25:39.704 --> 00:25:43.264 Who just got the Nobel Prize. Physics, interestingly enough. 00:25:43.864 --> 00:25:45.184 Yeah, because there is no AI. 00:25:45.444 --> 00:25:48.124 Yeah, because there is no computer Nobel Prize. 00:25:48.284 --> 00:25:51.944 Well, if you claim that your technology is everything, then you can get a prize in anything. 00:25:56.403 --> 00:25:58.203 Which is a proven point now, yeah. 00:25:59.043 --> 00:26:05.763 And it was Ilya Stutzkever and then Alex. I'm sorry, Alex, I'm not grabbing 00:26:05.763 --> 00:26:08.123 your last name from the ether right now. 00:26:08.223 --> 00:26:12.323 But nonetheless, this was a paper that kind of pulled together key ingredients 00:26:12.323 --> 00:26:16.703 that became the foundation of the AI boom now. 00:26:16.843 --> 00:26:20.523 So this is deep learning algorithms, which is the paradigm we're still in. 00:26:20.603 --> 00:26:23.523 It doesn't matter. There's architectural sort of rejiggering, 00:26:23.623 --> 00:26:25.303 but nonetheless, it's deep learning. 00:26:26.323 --> 00:26:31.523 Huge amounts of data. So what I've called the derivatives of the surveillance business model. 00:26:31.603 --> 00:26:33.503 He found all the cats on YouTube. 00:26:34.083 --> 00:26:36.043 I mean, yeah, that was Jeff Dean, I think. 00:26:38.183 --> 00:26:41.503 And then, you know, powerful compute, right? 00:26:41.623 --> 00:26:45.263 And they showed that sort of using gaming chips and a lot of data, 00:26:45.463 --> 00:26:47.083 you could beat the benchmark. 00:26:47.083 --> 00:26:53.163 So score much better against standard evaluations than past models and thus 00:26:53.163 --> 00:26:59.703 sort of catalyzed the industrial industry interest in AI and why were they interested? 00:26:59.943 --> 00:27:06.123 I think this is a key point because these optimist-seeking algorithms are really 00:27:06.123 --> 00:27:09.243 good at curating news feeds. 00:27:09.503 --> 00:27:13.903 They're really good at figuring out algorithms, right? 00:27:14.383 --> 00:27:18.783 I don't think it's an accident that Jeff was immediately hired, 00:27:20.063 --> 00:27:28.363 dass Jan LeCun, die die deep-learning-Algorithmen, die das Seed der Moment in 00:27:28.363 --> 00:27:32.043 AI in den 80ern, in den letzten 80ern, war er mit Meta. 00:27:32.063 --> 00:27:37.883 Und es war die Plattform-Companien mit der realen Investition in squeezing mehr 00:27:37.883 --> 00:27:43.263 Ad-Dollars aus der Daten, better-serving, all of das. 00:27:45.725 --> 00:27:51.125 Das ist Google mit DeepMind. Du sehen Meta und Google die leaders in das. 00:27:52.645 --> 00:27:57.025 Als measured by different evaluation standards, the measurement question hier 00:27:57.025 --> 00:27:59.065 ist eigentlich wirklich interesting und troubling. 00:27:59.665 --> 00:28:06.945 Until this generative moment where I think the ChatGPT Microsoft products shifted 00:28:06.945 --> 00:28:12.125 people's perception of AI and what it can do and just rearrange the leaderboard. 00:28:12.265 --> 00:28:16.405 But the paradigm is still the same And the paradigm is still that AI is applying 00:28:16.405 --> 00:28:22.225 old algorithms on top of the sort of massive platform monopoly business model, 00:28:22.585 --> 00:28:27.305 availing itself of huge amounts of data, which is produced via this surveillance 00:28:27.305 --> 00:28:32.165