Paul Frazee 0:02 All right. 0:03 Is this guy on? 0:04 We good? 0:06 Okay. 0:07 Well, I hope y'all are ready for some cheerleading here because we got a lot to do here. 0:13 So first of all, it is so great to be back here for the Atmosphere Conference 2026. 0:18 Thank you to Boris, to Ted, to all the volunteers for such a great conference. 0:23 Yes, give it up. 0:23 Give them some applause. 0:29 All right. 0:30 Well, what a year it has been. 0:33 So just 1 year ago, I was giving this talk, "Where Did We Come From? 0:36 Where Will We Go?" And, well, it was a pretty crazy year in between. 0:42 In a little bit, I'm going to step through my roadmap that I gave in that talk and see how we did, which should be interesting. 0:49 But last year was hypothetical. 0:51 I was looking into the past. 0:53 I was looking into the future. 0:54 And this year, I'm going to talk about the present Because 2026 is the atmosphere, y'all. 1:00 Atmos year. 1:01 Atmos year. 1:05 Get your puns right. 1:08 So there are so many incredible things happening in the ecosystem, from BlackSky launching a fully independent App View. 1:14 That's right. 1:20 We got Tangled raising a Series A. 1:22 That is incredibly exciting. 1:24 Yeah. 1:27 And then I have this I just want to— because I love it— StandardSite, a collaboration between 3 different projects bringing blogging and web content into the atmosphere. 1:41 All right. 1:41 So there are a lot more shoutouts coming. 1:43 But let's start with this. 1:45 Last year after the conference, we were talking to Boris. 1:47 And we were talking about what we were hoping to see for the next year's conference. 1:52 And I said, you know what would be cool is if we could 2x the number of attendees from last year. 1:57 And last year was about 170. 1:58 So let's see how we did. 2:01 Hey! 2:02 Yeah, this community is growing and it's growing because it is a welcoming community, it is a smart and kind community. 2:12 I love being a part of it and I want to thank each and every one of you for making this such a wonderful place to be. 2:18 Thank you so much. 2:24 Okay, so let's look at some more numbers. 2:26 It's hard to get exact numbers in an open network, but we have good proxies for how things are going and so like one of them we like to look at is the Ad Proto API SDK. 2:35 And the downloads keep going up. 2:36 This is the weekly downloads steadily rising. 2:41 Then we have another interesting metric with the Bluesky PDSs. 2:45 Whenever we see people log in, we actually get a sense of how many other non-Bluesky applications are out there. 2:51 And we see about 1,000 weekly active applications other than Bluesky. 2:55 These are applications that have somebody logging in at least once a week, right? 2:59 There's 1,000 of them out there. 3:01 When I search for @proto on GitHub, I find about 5,600 different repositories. 3:09 And you know what's funny is that this number probably isn't complete either, either because we don't run all of the PDSs. 3:15 In fact, we have been seeing great adoption of new PDS services from BlackSky, EuroSky, NPMX, and a lot of others, including the self-hosting. 3:26 Now this is showing weekly users posting from non-BlueSky PDSs. 3:30 That's over 10,000 people. 3:31 Every week, but then I heard Rudy say that they have 30,000 accounts, so clearly, uh, independent PDSs are on the move, so that is great to see. 3:43 So what is the excitement all about? 3:46 Why are we here? 3:48 Well, we are here, I believe, for two good reasons. 3:51 One is that we believe in user freedom and building good software. 3:55 We are hackers, idealists, creators, and professionals, and Good software and user freedom is what we do. 4:02 Now, more to the point, the Atmosphere is a new network built on the AT protocol, and it does a lot of things. 4:11 But the core of the Atmosphere, I think, is the account. 4:16 Having one account that you can reuse and stop creating new accounts all over the place, stop having to enter thousands of passwords into— that's maybe exaggerating a bit, but a lot of passwords that everybody forgets across the internet. 4:29 You have one Atmosphere account that you share across your applications. 4:34 And when I mean shared, I mean truly shared. 4:38 This is your account that you are taking to the applications, or maybe better to say that the applications are coming to you. 4:44 Because this is your data, and you choose to share your data with it. 4:47 So when I log into Bluesky, a fully independent application— pause for applause— my entire social network is there when I log in. 4:59 I tried it out recently and I was like, "Oh, look at that. 5:02 There it all is. 5:02 I feel right at home." There is no lock-in in the Atmosphere. 5:07 If your account is hosted by Bluesky, you can migrate to a new account host and lose nothing. 5:13 Tech companies don't get to own your Atmosphere account, you do. 5:17 In fact, if you want to self-host, you can. 5:21 Technically the username isn't about that, but you get the idea. 5:25 The atmosphere is also crazy programmable. 5:29 Okay, not quite that easy, but it is very, very easy to work with. 5:34 Now, we have been talking a lot about phrases that we like to describe what this is all about, and Anuj has been saying something for a while that we've decided we like actually pretty well. 5:45 People, not platforms. 5:47 That's a shout out officially because we're going to steal that a lot coming forward. 5:57 Yeah. 5:58 And actually, technically, the atmosphere as a term was invented by one of our members of our community, so. 6:03 Yeah, no, software for people, not software for platforms, not software for companies, but for people. 6:10 Now, OK, if you are new to the AT protocol, you might be asking, well, what can I do with it? 6:14 This all sounds really nice. 6:15 Well, that is a perfect opportunity for me to showcase some of what this amazing community has been doing. 6:20 So let's get into it. 6:21 And this is going to take a second. 6:23 I'm going to basically have 3 different categories of projects that we've been seeing out there. 6:27 Connected applications are just atmospheric applications. 6:30 Atmospheric websites, that's websites that are using the app protocol to enhance what you could do on the web. 6:35 And then infrastructure, this is tools and services and things like that. 6:39 So I will kick off with a crowd favorite here. 6:41 Leaflet is an atmospheric blogging application. 6:48 Comments from Bluesky integrate directly into Leaflet. 6:52 As if it's one network, which is funny because it is. 6:54 Okay, I'm going to skip a little forward here for a funny reason. 6:59 So you got 3 different applications on the Atmosphere. 7:01 Bluesky, which does short posts, Leaflet, which does blog posts, and Streamplace, which does video streams. 7:06 I decided to hit those 3 really fast because this crazy project happened. 7:12 Aether Operating System. 7:14 It's a web-based operating system that is a lot of fun. 7:18 It has access to the full firehose of data on the Atmosphere. 7:20 So it was able to replicate all 3 of those applications inside of it as a part of that experience. 7:27 Okay, next up. 7:29 All right, I swear, Rudy, it's the last time, but here we go. 7:31 Black Sky, completely independent fork and reimplementation of Blue Sky. 7:36 It has its own moderation, it has its own algorithms, and its own UI. 7:41 Now something that's really important about this is that our community Communities interact every day without even really being aware of the difference between the two. 7:49 They know that they're using BlackSky or Bluesky, but we're all a part of one network together. 7:55 All right, I briefly mentioned it. 7:57 StreamPlace, live streaming service that works really well. 8:01 That's how a lot of people are seeing me right now. 8:05 We've been using it a lot just because it works, which is, you know, something. 8:13 And one fun detail about it is that when you go live, you publish a Live Now indicator, which now shows up right inside of Bluesky. 8:21 And that has been great because Eli has been streaming about news that's been happening in the ecosystem. 8:26 I keep catching it because he's going live. 8:27 I see him on the profile, and I go and catch the thing while it's still happening. 8:32 All right, next up we have Tangled. 8:39 Yeah. 8:39 So Tangle is a GitHub replacement that puts the commits in Git and the issues on App Proto. 8:46 Next up, we have Germ, an end-to-end encrypted messenger that uses the new MLS encryption standard, and it uses App Protocol identities. 8:54 And you can start your conversations with the press of a button from a Bluesky profile. 8:59 Y'all are bringing the energy. 9:05 I love this. 9:07 What's that? 9:14 Whoo! 9:16 Please save the questions for after the talk. 9:23 Okay. 9:23 Next up we have Graze, Surf, and SkyFeeds, 3 incredible feed-building tools that integrate directly into Bluesky and really any application that wants to take advantage of generated feeds. 9:37 And that's it for the connected applications. 9:38 Let's now turn our attention to the atmospheric websites such as npmx. 9:44 Has anybody heard of npmx? 9:50 Anybody notice that all over your timeline? 9:53 Really cool community project. 9:55 I don't know how y'all are doing what you're doing. 9:57 But it's a browser for npm. 9:59 It has login and likes via app protos. 10:02 A little sprinkled in there. 10:04 In fact, if you like @proto/logins, in a classic dogfooding move, Boris put together the community discourse, and it uses @proto/login. 10:12 So we always know exactly who we're talking to. 10:14 We're not having to figure out, wait a minute, who's that? 10:15 It's got our usernames right there. 10:18 Next, like I mentioned before, we got standard sites. 10:21 This is a lexicon created by Leaflet, Pocket, and Offprint to enable a common language for publishing articles on the @protocol. 10:33 Now, why does that matter? 10:34 Well, let me use a personal favorite here, the @proto website. 10:38 It publishes its blog posts both on the web and on the app protocol using the standard site lexicon. 10:47 It also happens to have blue sky comets integrated directly in. 10:53 And so now that we have our blog posts getting published on the atmosphere, you can have sites like docs.surf. 10:59 For a Windows 95 flavor for reading all of your posts. 11:03 And if you've been missing Dan's posts about the— or his leaflets about the permission spaces, you can find them right there in that Windows 95 experience that I know everybody here is craving. 11:15 Is that '95 or is that XP? 11:17 That's XP. 11:18 Oh, guys, come on. 11:21 We'll get through this. 11:27 Of course, if you're a little more modern, you could use the Pocket.blog reader user experience. 11:32 So connected apps, atmospheric websites, they're all pushing on what you can do when you are part of one network together. 11:39 And maybe you, the person new to app proto, don't really like making apps or websites, but you like making infrastructure, like giant crazy services and indexes of lots of data. 11:50 Well, then maybe you'd take an interest in Vig's work. 11:54 [APPLAUSE] And actually, But if you're building, you probably want to give microcosm a look too, because these are a collection of very, very helpful ad protocol services that make development a lot easier. 12:11 There's Constellation, which lets you query the network by looking at backlinks, which is a shockingly intuitive way to do it. 12:18 There's Spacedust that essentially watches for, as far as I can tell, this may not be a great description, but new links between users, like follows or likes, and streams them to you. 12:26 I think it can be used to make, like, notifications essentially, or activity streams. 12:30 There's Slingshot, which is a giant cache of user IDs, and I think also records, And now there is Hubble, a full network mirror. 12:39 So that is Microcosm and there's also a lot more that Fig has been doing, so definitely check that one out. 12:45 There is also PDSLS. 12:47 I'm counting tools as infrastructure here. 12:49 This is an app proto data explorer that has become a regular part of development for everybody. 12:55 And I actually am gonna just sneak in something that happened between me writing the talk and now, which is Taproot, which has developed an insane number of tools in this spirit. 13:04 So definitely check out Taproot. 13:06 You've got Lexicon Garden, which aggregates all of the app protocol schemas being published across the network. 13:12 I've been using this quite a bit myself. 13:14 Next, you've got PDS Mover, which is an account migration tool. 13:19 Yeah. 13:24 With genuinely one of my favorite logos. 13:26 And it's been seeing huge utility for people looking to switch their hosting infrastructure. 13:31 And speaking of that, you got EuroSky, which has been making moves with their PDS service. 13:35 That's right. 13:40 I believe both EuroSky and NPMX are located in the EU, so we've been actually seeing both pick up. 13:46 EuroSky also has their own migration tool they've been working on with the best name ever, EU-Haul. 13:51 It's good. 13:54 It's good. 13:56 All right. 13:56 And one more. 13:57 We got BridgyFed. 13:59 Like I said. 14:04 Successfully bridging ActivityPub and App Proto for over a year now, I think. 14:10 3 years. 14:12 Where does the time go? 14:15 Yeah. 14:15 And like I said in the talk previously, I have to give this one a lot of credit because I thought that would never work. 14:24 OK, well, are you inspired yet? 14:31 I am sure that I missed some really great shoutouts here because there is so much happening. 14:36 And I know that y'all are probably tired of applauding, but let's give one more round of applause to this community because this is what happens when good people build good software. 14:53 Okay, so I mentioned earlier that last year in my talk I shared a roadmap for the protocol and now it's time for the report card. 14:59 So how'd we do? 15:00 Did we follow through? 15:02 Let's find out. 15:02 Sync 1.1. 15:04 Yes! 15:06 Hot dog! 15:06 We did it. 15:11 So now the talk gets a little drier. 15:17 Sync 1.1 is spec'd, published, and deployed. 15:20 This update made relays much cheaper to run and improved sync reliability in general. 15:25 Fig demonstrated a working relay at less than $5 a month. 15:29 Which is great. 15:31 And there are now over a dozen relays running. 15:33 Okay, auth scopes and permissions. 15:35 Did we get that done? 15:37 Yes. 15:38 Apps can now choose exactly which permissions they request from users, and you can use permission sets to condense that down into a much more readable user experience so you're not seeing a long list of NSIDs and wondering what horrible thing you've gotten yourself into. 15:51 Lexicon resolution. 15:52 Did we get that done? 15:54 Yes. 15:54 Lexicons are now published on protocol. 15:57 This is why things like lexicon.garden can aggregate them together to form nice live documentation. 16:02 And this has ended up being a really fortuitous development in the protocol, which we'll talk more about later. 16:08 Next up, permission data. 16:10 Oh, it's close. 16:15 We are very excited about this. 16:17 The Bluesky team has been working on our proposal and has been sharing the informal write-ups of that. 16:21 A formal write-up will be coming soon. 16:23 There's also been a number of people in the ecosystem working on their own proposals, which we are going to be bringing together to to make sure that we are getting all the great ideas possible for this and working with y'all so that we can have a great permission data solution. 16:36 So I'm feeling very good about how this is coming together. 16:38 I'm also very— feeling very good about letting Daniel take this. 16:42 So if you have any questions, take them to Daniel. 16:46 Okay. 16:47 Now, end-to-end encrypted messaging. 16:48 We did actually say this one last time and not exactly. 16:52 We have officially not picked this one up. 16:54 That said, Like I said previously, Germ has picked up the mantle on this and deployed end-to-end encrypted messaging using Ad Proto for Identities. 17:01 And I'm sure that as permissioned spaces roll out, there will be some interesting opportunities to look at merging those kinds of technologies together. 17:09 Okay, so what else? 17:10 That's not bad, but that's also not all because that's just what I happened to put on the slides for last year. 17:16 Let's see what else has rolled out. 17:18 PLC replication. 17:20 Uh, this is actually pretty, darn useful, not just ethos-wise, but practically speaking, the PLC directory now supports replication. 17:28 So you can sync the entire PLC directory into your local infrastructure, which is very nice if you have to do a lot of identity lookups. 17:34 So if that's you, take a look at the PLC replication. 17:39 There's also a new TypeScript SDK, @proto/lex. 17:42 This makes it a lot easier to install and work with lexicons that are published on the network. 17:47 You were previously stuck with whatever we were compiling into our API. 17:51 Now there is a very nice set of tools and flows to be able to build the SDK that you need for whatever lexicons you're working with. 17:59 The tap tool now makes it much easier to backfill and sync data from the network. 18:03 It kind of works magically, handles the cursor for you, and switches from backfill, searching out what's already out there, straight into live replication so that you get events as they occur. 18:12 Definitely check out tap if you've been wanting to tap into the network. 18:18 We put out a refreshed website and completely reworked the docs. 18:22 And, you know, relatedly, I have to give a shout out to our first two DevRel, Jim and Alex. 18:27 Yeah, they are here and they are happy to talk to you. 18:37 So if you haven't met them yet, be sure to say hi. 18:39 There's also one more nice little bit of news that I am pleased to share. 18:45 The @protocol/iETF working group has officially been formed. 18:54 I'm just going to give a huge thank you to the IETF for welcoming us and a huge thank you to this community for helping to make that happen, for showing up and helping to explain to the IETF why this is something that is worth investing in. 19:08 The AT protocol is a long-term vision for the internet. 19:11 The IETF is a great home for the standards governance. 19:15 So I've been over the moon about that. 19:19 Okay. 19:19 So what is next? 19:21 As I said before, permission data is going to be, I think, probably the banner update for this year. 19:27 This is the tools for creating records that are either totally private or gated to a selected group of people. 19:34 This is going to open up a huge number of use cases for the app protocol, so I cannot wait for this. 19:39 We are also going to be spending some time working on the account experience. 19:42 This is login and account management. 19:44 I know that there are some people in this room who know exactly what I'm talking about. 19:48 And are excited for this. 19:49 But this will include things like, believe it or not, we're going to at least claim two-factor auth will finally make its way in. 19:56 We are also going to be helping to work with the FedCM standard, hopefully making some progress on that to make logins faster and easier. 20:02 Very excited about that. 20:04 There's also some branding questions about what kind of login experience we should all be using, and we want to do our best to help answer those questions with everybody. 20:12 All right. 20:13 What else will we be doing this year? 20:14 There's going to be a lot of follow-through about governance. 20:17 In fact, If you are interested in that, you should catch Daniel's talk, which, unless he has changed it, is supposed to be about governance. 20:24 That'll be about the ITF and also about DID PLC, which has some very interesting developments on the way. 20:29 So be sure to check out that talk if this is something that you are interested in. 20:33 And then there are just, frankly, continued improvements, quality of life, more tools, more SDKs, more services, fixing bugs, fixing docs, and making the app protocol a joy to work with. 20:44 So that is our plan moving forward. 20:48 Whoo! 20:49 All right. 20:59 So this is the moment that you realize that this talk is for an hour. 21:03 And I didn't bring a guitar, but maybe later. 21:12 I think there was some talk of karaoke later. 21:14 Yeah. 21:16 You know I'm in. 21:18 Well, we live in some crazy times, don't we? 21:21 Slay the Spire 2 came out. 21:25 Ozempic is a thing now. 21:29 Oh yeah. 21:31 Also, AI is a thing. 21:34 That really happened to me. 21:35 Yeah. 21:42 Thank God, we're finally there, you know. 21:47 I posted that picture and somebody naturally asked me, "Which program is that?" And I suppose you're wondering that now too. 21:56 Okay. 21:57 Well, we are living in a radically different world than we were even one year ago and it has raised some big questions. 22:03 What is this future that we are creating? 22:06 How should we react to this new technology? 22:08 What role does decentralization have in shaping this world? 22:12 And what role should we be playing both as individuals and as a community and as an ecosystem? 22:18 Well, as you all know, Jay, the founder of Bluesky, has recently changed her role to the CIO, which is the chief innovation officer. 22:25 And this is why Tony has joined us as a CEO so that Jay can make questions about the future of App Proto questions like this her main focus. 22:33 So please join me in welcoming Jay to the stage. 22:52 Hey everyone. Jay Graber 22:55 Hi, how are you all doing? Paul Frazee 22:57 Good. Jay Graber 23:00 So I want to start by sharing a quote that I really like. 23:03 This is by a historian of technology that I keep coming back to. 23:07 Technology is not good nor bad, nor is it neutral. 23:13 This sounds like a riddle, but I think it actually contains an important insight. 23:17 Technology is a force multiplier. 23:20 It multiplies the impact of whatever you apply it to. 23:23 And the outcomes depend on how it's used, by who, and towards what end. 23:29 Our mission is to develop and drive adoption of large-scale technol— large-scale adoption of technologies for open and decentralized public conversation. 23:38 I wrote this into our founding charter because I believe the internet should be built to serve the people who use it. 23:44 Right now, AI is starting to undermine human agency at the same time as it's enhancing it. 23:51 The spread of low-quality AI-generated content is making public social networks noisier and less trustworthy at a time when we need accurate information more than ever. 24:01 The signal is getting harder to find right when we need it most. 24:06 And the major platforms aren't exactly trying to fix this. 24:10 They're using AI to increase the time users spend on platform, to harvest training data from you, and to shape what users see and believe through systems that they can't inspect, and can't control. 24:21 We think technology should serve people, not platforms. 24:24 [APPLAUSE] So an open protocol, the purpose of it is to put the power to shape your own experience in your hands. 24:37 You build your own feeds, you can create software that works the way you want it to, and you can find signal in the noise. 24:43 We built the app protocol so anyone could build any app they imagine on top of it. 24:48 But up until recently, anyone really meant anyone who can code, which might be all great for people in this room, but really there's a lot more potential here. 24:58 Agentic coding tools are starting to make this a lot more accessible. 25:02 For the first time, an open protocol can be truly open to everyone. 25:06 It's increasingly possible to personalize software with no coding experience at all. 25:11 The Atmosphere is an open data layer with a clearly defined schema for applications, which actually makes it uniquely well-suited for coding agents to build on. 25:19 So we started asking, what happens when you can just describe the social experience you want and have it built for you? 25:26 So in my new role as CIO, I've started an exploration team within Bluesky to find out. 25:32 The first thing that we've built is called Atty, named after the protocol. 25:36 @hee is an agentic social app and custom feed builder. 25:39 So it feels more like having a conversation than configuring software. 25:43 You just describe the kind of posts you want to see, and the agent underneath the hood builds the feed for you. 25:48 But that's just where we started. 25:50 Paul's going to give you a sneak peek today of what we've been exploring at the frontier of what software on an open network can look like. 25:57 Technology is an accelerant on whatever you apply it to. 26:00 And I want this to accelerate decentralizing social and putting power back in users' hands. 26:05 But honestly, I don't think the most interesting things built on @proto are going to come from us. 26:10 They're going to come from everyone out here who picks up these tools and starts building. 26:14 So we've just started experimenting, but we're sharing what we have with you now so that you can build alongside us. 26:19 I've watched people use Addy for the first time over the last couple of weeks, and there's a moment where they realize they just described something and suddenly it exists. 26:28 That moment's what this is all about. 26:30 I don't have all the answers for what the future of the open web looks like, but I can't think of a better group of people to be building it with. 26:36 So if you want to do that literally, we're hiring. 26:38 Come talk to me after or visit our jobs page. 26:41 And Paul is going to show you what we've been working on. Paul Frazee 26:52 All right. 26:53 Well, who's ready for a demo? 26:58 Well, this is Addy. 27:01 Can we see that okay? 27:02 All right. 27:03 It's pretty good. 27:04 Yeah, this is the result of the last 2 months of some pretty crazy coding. 27:09 One of Bluesky's first ideas on the atmosphere was the custom feed generators. 27:22 Now, we added feed generators because one of the most fundamental problems with social networks is the algorithm. 27:29 How do I control my algorithm? 27:32 And the general sense that we have here is, wouldn't it be nice if you could just ask? 27:38 And so what we're seeing here is that the agent is running some tool calls, searching the network based on my query here. 27:46 And it seems to have created the feed. 27:47 Now it is testing the feed to make sure that it has the right fit. 27:51 And in a second, we're going to see the actual output of this. 27:53 So let's Hey, there we go. 27:55 Oh, look at that. 27:58 It's kind of recursive. 28:04 And we have here a feed that is about the Atmosphere conference. 28:07 Now, it's— in a second, we'll figure out exactly how it did this. 28:10 But you can see that it's doing a pretty good job. 28:12 Sometimes it's getting @proto, Atmosphere conference. 28:14 It's going by tags probably frequently. 28:16 But in this case, it probably went off the alt text. 28:20 You can also see over on the side there's a nice little summary of everything that's in the feed that we're looking at here. 28:25 That's particularly nice for kind of a breaking news feed to quickly digest what's in there. 28:31 But this is about talking to it. 28:33 Let's talk a little bit more. 28:38 Likes and repos, y'all. 28:39 Come on. 28:46 Addy works by building a kind of a pipeline program, and we can actually examine it. 28:51 It'll probably still— there it goes, okay, so it's done. 28:55 I'll clear that session. 28:56 All right, there we go, ranked by hotness. 28:59 Yeah! 29:14 Right on. 29:19 Okay. 29:20 So, Addy built a kind of a pipeline program to do this. 29:23 It specifies its sources like keyword searches, which is what it's doing here. 29:27 It's looking for the Atmosphere conference hashtag and mentions of the Atmosphere conference. 29:31 So, let's see. 29:32 Take a look there. 29:32 So, yeah, looking for the text. 29:34 So, it's actually just kind of running probably Elasticsearch initial queries. 29:39 Then it goes down the pipeline. 29:40 You've got a pre-filter, which is actually TypeScript. 29:42 That it's executing on the server. 29:45 Then it's going to do a labeling pass on every one of those things. 29:47 You were evaluating posts about Atmosphere conference. 29:49 And what it's trying to figure out is, is it actually about the conference or not? 29:52 And it emits these labels, is about conference, yes or no. 29:55 And then finally, it runs a TypeScript pass down beneath where it is filtering out based on the AI labeling. 30:02 It is removing duplicates from authors. 30:04 And then it is applying that ranking algorithm that I asked it to do. 30:08 It wrote all that code on request. 30:11 This is a program-building agent. 30:16 Oh, you know what else I could do, which is rather neat, is that I can explore each of the steps in the debugger here. 30:21 So I could see it got 107 initial posts, ran that pre-filter, that didn't filter, got the labels onto it and was able to determine which ones are about the conference or not, and then ran the curator, which is pretty helpful whenever it's not doing what it's supposed to do. 30:36 Okay, well, that's pretty handy. 30:38 Why don't I go ahead and publish that? 30:43 Okay, there we go. 30:44 And I'll just refresh here. 30:47 Go to my feeds. 30:49 Atmosphere Conference 2026. 30:52 Isn't interoperability cool? 30:59 And maybe you're thinking, well, of course Bluesky can do this. 31:01 They have privileged access to the network. 31:04 Well, it would be reasonable to think that because that's how software has always worked, but not here. 31:09 ADDIE was built using the exact same access that everybody has to the atmosphere. 31:14 Put another way, these ideas can be built by anyone here or not here in this room. 31:19 And we plan to help share those ideas, those techniques. 31:21 We will share what we have been learning with the— with ADDIE. 31:27 Now, let's do a little something here. 31:31 I should first kind of say this is what we will be sharing with everybody after the conference. 31:36 But I have a little tool here. 31:41 And I'm just going to give myself some more features. 31:46 Hey, that's better. 31:51 So I'm going to show you some of the stuff that isn't quite ready yet for release, but we really wanted to show it to to show you here because we think it illustrates the future that's ahead of us. 32:01 I'm going to open up this fake Bluesky interface here. 32:06 It's not fake, actually. 32:07 It's pretty real. 32:08 But one of the things that has always bugged me about the Bluesky interface— I know I probably shouldn't say something like that, but I know the people that are in my following feed, and the handles start to just feel like noise after some point. 32:20 So why don't we just go ahead and get rid of those things? 32:23 So let's go with— don't show the handles in the posts. 32:35 What is the core idea behind personal computing? 32:38 It's that software ought to serve the user. 32:41 And for years, there's been a divide between the coders and the users. 32:46 Now, maybe we will never fully erase that divide, but in the past year, the laws of physics has changed for software. 32:55 We can now tell the computer what we want in plain English and it will do it. 33:01 It will make or modify applications on our behalf. 33:06 Perhaps the big question is, will it do it fast enough for my demo? 33:13 Ah, perfect timing. 33:14 Isn't that cleaner? 33:34 When we started, we knew we weren't interested in generating content for people. 33:39 If anything, we kind of want the opposite. 33:41 We want to enhance the atmosphere and not pollute it. 33:45 So I started pitching a constructor. 33:48 This is a tool that builds the atmosphere instead of just browsing it. 33:54 And Jay said, well, that term is terrible. 33:58 Why don't we call it a liquid interface? 34:01 Because after all, we're liquefying software. 34:04 It is dynamic. 34:05 It can change. 34:06 And it is the opposite of the shrink-wrap software that we've been using for most of our lives. 34:12 So we built a Liquid interface, a live programmable atmospheric environment. 34:17 Now watch what it does. 34:20 It finds the lexicons. 34:23 You can see here it is searching through the leaflet lexicons for rich text. 34:30 We made the app protocol self-documenting. 34:33 The APIs and the data models all have live documentation in the form of lexicons. 34:38 And if you're not familiar with Lexicon, it's a lot like JSON schema. 34:43 It describes the APIs and the data models, as I said. 34:45 And we made these Lexicons to help software engineers build together in an open environment, which is a challenging coordination challenge. 34:54 But a lucky side effect is that we made something remarkably similar to MCP, which the coding agents can use. 35:03 And because the Atmosphere is an open network, the agent is able to build live software against it. 35:09 Now this time I knew that it was going to take a little bit of time. 35:12 It's read the documentation it needs to, it's cooking on the code. 35:15 I'm just gonna jump to one that I had it build previously. 35:19 And here is my leaflet rendering, and this is the result of asking. 35:36 Still thinking. 35:39 So Addy is an agentic social environment, or perhaps more accurately, it's our playground for experimenting with AI on the atmosphere. 35:49 Or perhaps more accurately, it's the result of 2 months of deep AI psychosis as we have been working on this thing together. 35:58 But I do want to give a shout out to the team that has been working on this. 36:00 This is Jeremy Haley, Jim Collabro, and Jay for being a part of making Addie v1 possible for this conference. 36:12 So this is online at addie.ai. 36:21 It is a closed beta. 36:23 And guess what, y'all? 36:25 Guess what that means? 36:27 Invite codes! 36:29 Our greatest innovation yet. 36:31 We've brought AI to invite codes. 36:36 There is still a lot to learn about what it means to actually deploy AI in the atmosphere or just in general for personal computing. 36:45 We're learning a lot about the performance. 36:46 We're learning a lot about the cost, learning a lot about correctness and security. 36:52 As Jay said, the atmosphere is here so that anybody can build. 36:57 We wanna make anyone everyone. 37:00 And we want to figure this out with all of you. 37:03 So thank you all so much for being here. 37:05 Thank you for your time. 37:07 I'm excited about what's coming up this next year and to be here with you. 37:10 So thank you very, very much. 37:22 And you should get an email with an invite code if you're at the conference, so if you don't get it, hit me up after. 37:33 Do you want to take some questions? 37:35 Do you want to have people— we've got 5 or 10 minutes. 37:39 I think we can do that. 37:40 Okay. 37:44 Let me run mics. Jay Graber 37:49 Thank you so much. 37:51 My question is, as I'm worried about the people outside of the conference that may not be very eager to let their posts be read by an AI agent, so is there a way for them to To be like you have the settings to turn off people not logged in in BlueSky to see your account. 38:25 Something like that for these AI agents would be very much appreciated, I think. Paul Frazee 38:33 I think that conversations like that absolutely should happen after this. 38:39 I think that, well, you know, we have to meet people where they're comfortable with. 38:44 One of the things that we want to make sure that people get out of AI is something that actually serves people and finding out what it means to actually serve people. 38:51 So far, a lot of what we've experienced with AI has felt more like it's serving slop. 38:57 And I think we've all had enough Shrimp Jesus in our lives, and it doesn't feel great to help that happen. 39:02 So hopefully, I want to believe that this thing, this kind of direction, finding ways for AI to actually help people will make that less scary as a premise, but I think that what you're describing is something that we should have that conversation about. 39:17 Yeah. 39:20 You said you like described it as like a playground. 39:23 Do you see this as like an experiment or like something you want to commit to in the future? 39:29 And if it's more of an experiment, do you see yourself open sourcing it? 39:35 We are definitely interested in We haven't open-sourced this one yet because it's a bit of a mess, and we have some features in there that we're not quite comfortable— we don't feel like we're ready to share with people. 39:46 So we will absolutely be open-sourcing some software that is derived from this, at minimum, to make sure that people are able to kind of pick up on what we learn through this. 39:57 And we have been learning quite a bit about what it means to, for instance, orchestrate the context window to make sure that it's able to interact with the network properly. 40:03 Things like that. 40:06 But is this an experiment, or is this like a product? 40:09 We actually don't know yet. 40:10 We're kind of interested to hear what people think. 40:12 And if it ends up being just a place for us to learn things and then export so that somebody else can do it better, then that's one pathway. 40:18 And if it ends up actually having utility for folks, we'll stick with it. 40:23 Yeah. Jay Graber 40:24 I'm going to be selfish and have two questions. 40:27 One, sorry if I missed this, but what's the underlying model behind Addy? 40:31 And then second, you know, obviously this changes whether you open source it or not, but there are obviously going to be some safety and moderation implications when you introduce genetic workflows like this. 40:42 So you know how to find us. Paul Frazee 40:46 More of a comment, really. 40:52 Yes. 40:52 Okay. 40:53 Which model? 40:55 We actually, I, you know, Forgot this was going to be a story beat, but we have a model selector that you can configure in your own model. 41:05 At the moment, it's a mix. 41:07 And actually, one of the things we've been learning quite a bit about is when you want to switch between different models for different use cases, which kind of makes configuring it even more of a pain. 41:15 But primarily, we're using Anthropic in the demo. 41:19 And a lot of that was Sonnet. 41:21 Yeah. 41:23 Let's see. 41:24 Oh yeah, well, correct. 41:26 But I'm also very interested in using the potential for using this sort of thing to build more personal labelers, actually. 41:34 So I think this could end up having great value in that direction. 41:38 Yeah. 41:41 Other questions? 41:52 Hey, I'm Stan. 41:54 I'm just wanting to ask, I just received the email with the invitation. 41:57 Hey, it worked! 41:58 It worked. 41:59 Thanks, Jim. 42:00 And I just noticed at the end of the email that it was only available in the United States and Canada. 42:07 That's right. 42:07 For 18 and more users. 42:10 Can you explain why you had to add this restriction? 42:15 Because it doesn't seem to manipulate user data. 42:20 It's only creating programs. 42:21 So I'm kind of wondering, why did you have to do that? 42:25 And maybe it can highlight, you know, stuff that Victoria talked before. 42:30 Thanks. 42:31 If Matt is watching, are you panicking right now about what I'm about to say? 42:34 You should talk to our head of legal, Matt. 42:37 Yeah, there. 42:39 And I will leave it at that. 42:41 He'll be able to I'll give you a much better answer. 42:47 So I had a question regarding that demo that you did regarding the UI. 42:53 So is the idea that any user could make potentially any change to the UI and have that be their experience? 43:02 Because one of my concerns is that especially once you start introducing communities information data, I feel like I feel like the UI is a very important part of— it's a very important part to show different social signals that might be important. 43:19 And having a standardized way of displaying that, that the user might not be able to easily control, I feel like is an important part to having a coherent social experience sometimes. 43:32 Sure. 43:34 Is the idea just that anyone can change the UI however they want, essentially, or? 43:40 Well, there's an interesting immediate question that includes what you're bringing up, but also how do we live in a consensus reality of user experience? 43:48 And I think that, well, first of all, we're going to be finding out together in a big way. 43:53 But Lexicon is kind of the tool for that. 43:56 I think that we should be Using the— if we need to add more primitives into lexicon to define this, then we should. 44:05 Getting to a point where we can kind of set mutual expectations across these software application boundaries, you know, so that whenever I customize, I don't do something that suddenly doesn't work on your machine because you use different schemas and have different expectations. 44:19 That's actually a pretty big part of it. 44:21 Another interesting thing that'll come into play in general is just How are we choosing which lexicons and which descriptions of how the software should work to trust? 44:30 And so attestation and trust layers for the inputs into the agents is going to be a really compelling research area throughout all of this. 44:39 Hello. 44:40 I have a question, which is, does this tool have a set of tools on its own that make it easier to interact with App Proto? 44:51 Lexicons and such, or is it just sort of using the lexicons as like an MCP? 44:57 Could you hit me with the question one more time? 44:59 Does— it's called Addy, right? 45:02 Yeah, yeah. 45:02 Does Addy have like a toolkit that is providing like guidance for how it should be interacting with this sort of content, or is it kind of just like, you know, because you can just ask Claude to build you a feed or whatever, and it will. 45:18 Right. 45:18 So is it like, does it have a set of tools that is providing guidance or help to make it sort of on a better track to have that happen? 45:28 Yes. 45:29 So first of all, it has its own set of tools that we custom-built. 45:33 It actually uses a kind of a generated documentation structure because we found that it seems to like like the navigational style of trees. 45:45 And so there's some of it which is just fixed and baked in explaining the environment that it's developing against. 45:51 There's a lexicons directory which is actually doing lexicon resolution and then generating Markdown in the response that embeds the lexicon and then gives it like usage examples and things like this. 46:02 Now some of that stuff we absolutely should be producing as like Claude skills or MCP because the same utility exists lists generally. 46:11 And that's one of the things that we're also quite keen to get out so that, you know, we put together the TypeScript SDK recently with better code generation. 46:18 Well, these things are incredible code generators. 46:20 So making sure that they can get access to the lexicons I think will actually really help speed up development just in all cases. 46:27 But there's something quite nice about the live environment, having both kind of a more constrained target, one that's built for specific use cases, and which we can tell it, tell the coding agent things about the configuration, what services are available or what like endpoint the user might prefer to, for like an app view versus one versus the other. 46:49 So we can, I think the live environment has some interesting possible advantages in addition to just being a live environment that doesn't, you know, takes away the pain of deployment and setup and having to pick your API keys and things like that. 47:02 And I think we'll leave it there. 47:03 You can mob Paul and the rest of the team later. 47:07 Thank you so much for giving us the atmosphere report. 47:10 Thank you, everybody.