WEBVTT 00:00.000 --> 00:06.000 All right, I think we'll kick off. 00:06.000 --> 00:07.000 Thank you, David. 00:07.000 --> 00:09.000 Thanks a lot, everyone. 00:09.000 --> 00:10.000 I'm Danny Deiner. 00:10.000 --> 00:13.000 I'm presenting Bill your own timeline algorithm 00:13.000 --> 00:17.000 or as I like to call it, it's Deota. 00:17.000 --> 00:20.000 It also sounds like Bill your own travel agent. 00:20.000 --> 00:22.000 I think there was another project like that. 00:22.000 --> 00:23.000 Sorry about it. 00:23.000 --> 00:25.000 So, Bill, it's not a new algorithm. 00:25.000 --> 00:27.000 It's not an end user application. 00:27.000 --> 00:29.000 It's not a service. 00:29.000 --> 00:32.000 To me, it's a pacifist called for arms. 00:32.000 --> 00:36.000 This is why I chose the blues brothers instead of Uncle Sam. 00:36.000 --> 00:38.000 It's a personal project. 00:38.000 --> 00:40.000 I worked on in the last couple of years. 00:40.000 --> 00:44.000 It took a while, but I had to wait for the right tools to be around. 00:44.000 --> 00:47.000 And you will see it's just something I put together by it. 00:47.000 --> 00:49.000 Putting together great tools. 00:49.000 --> 00:51.000 How did this start? 00:51.000 --> 00:54.000 It all started with a sink. 00:54.000 --> 00:58.000 And you'll know the consequences of this sink, right? 00:58.000 --> 01:03.000 So, a lot of people moved from Twitter to other open social networks. 01:03.000 --> 01:05.000 And to me, it was a bit different. 01:05.000 --> 01:06.000 It started six months earlier. 01:06.000 --> 01:08.000 I was at Twitter employee. 01:08.000 --> 01:12.000 And I joined a smaller cohort and tried to look around 01:12.000 --> 01:16.000 for something that I felt was a bit more in control. 01:16.000 --> 01:20.000 I started to think about how to put something more in control of people. 01:20.000 --> 01:22.000 The networks were there already. 01:22.000 --> 01:24.000 They were already in control of people. 01:24.000 --> 01:27.000 The only thing which was missing in my opinion was an algorithm. 01:27.000 --> 01:29.000 A lot of people don't think we need one. 01:29.000 --> 01:31.000 Some of them might think about that. 01:31.000 --> 01:33.000 I will tell you what I think about this. 01:33.000 --> 01:35.000 So, timeline algorithms. 01:35.000 --> 01:37.000 We don't have a lot of time to delve deep into them. 01:37.000 --> 01:41.000 The only things that will leave you are a very, very short definition. 01:41.000 --> 01:46.000 They define weather and how posts appear on your timeline. 01:46.000 --> 01:52.000 I also put reverse chronological there because to me, that's an algorithm too. 01:52.000 --> 01:55.000 And if you want to delve deeper into, let's say, 01:55.000 --> 01:58.000 a platform recommendations and timeline algorithms. 01:58.000 --> 02:01.000 I suggest you this blog post, which in my opinion was great, 02:01.000 --> 02:05.000 written by Luke Thorv, who's in research on this specific stuff. 02:05.000 --> 02:10.000 The other thing I studied, and I worked on, were all the past projects 02:10.000 --> 02:13.000 that had been built in the Fedivers about this. 02:13.000 --> 02:15.000 I'm really standing on the shoulder of giants. 02:15.000 --> 02:19.000 I learned a lot just by seeing how people were trying to build new things 02:19.000 --> 02:23.000 and how these were accepted by the people on the Fedivers. 02:23.000 --> 02:27.000 So, by putting all of these together, I came up with some problems 02:27.000 --> 02:31.000 that timeline algorithms more or less all have. 02:31.000 --> 02:34.000 And they don't need to have all of these problems. 02:34.000 --> 02:37.000 Some of them have a few, some of them have more. 02:37.000 --> 02:40.000 And I think they are starting from an assumption, which was, 02:40.000 --> 02:45.000 we need to be something to allow users to make users come back to our platform. 02:45.000 --> 02:50.000 And I think if we drop that assumption, we can actually find many solutions to these problems. 02:50.000 --> 02:53.000 And I will not have deeper into each of them, but just tell you, 02:53.000 --> 02:57.000 putting together different tools, we can try and tackle this. 02:57.000 --> 03:00.000 So, Biotta is built on different kind of tools. 03:00.000 --> 03:04.000 I told you, one of them is called Lama File, another one is called Marimo. 03:04.000 --> 03:06.000 Another one you already know, it's master. 03:06.000 --> 03:08.000 I actually mastered on the pie library. 03:08.000 --> 03:11.000 So, another talk about pie-tonal libraries for clients, 03:11.000 --> 03:15.000 because you can actually find plenty, any social network, 03:15.000 --> 03:17.000 which is not so stupid to remove. 03:17.000 --> 03:21.000 Sorry, I will rephrase it, which is intelligent enough to give you an API. 03:21.000 --> 03:25.000 We'll allow you to get your posts. 03:25.000 --> 03:29.000 You can also use summaries from an RSS feed, if you want. 03:29.000 --> 03:32.000 As long as you get text, it's going to work. 03:32.000 --> 03:36.000 Lama File is a single file language model. 03:36.000 --> 03:40.000 I purposely removed LLM, I just wrote language model, 03:40.000 --> 03:42.000 because it doesn't need to be large. 03:42.000 --> 03:46.000 There's one that I'm using here, which is called Old Minielem, 03:46.000 --> 03:48.000 which is 46 megabytes. 03:48.000 --> 03:51.000 It has a valuable code, it's an academic project, 03:51.000 --> 03:54.000 and it has information about the data sets that have been used 03:54.000 --> 03:56.000 to generate its own embeddings. 03:56.000 --> 04:00.000 It's Lama File, runs it, it's 100% local. 04:00.000 --> 04:04.000 It's optimized to run on slower and older machines, 04:04.000 --> 04:06.000 and it's used to calculate these status embeddings. 04:06.000 --> 04:08.000 If you don't know what's an embedding is, 04:08.000 --> 04:10.000 let's say it's numerical descriptors 04:10.000 --> 04:13.000 that are the most similar, the most close, 04:13.000 --> 04:16.000 the more the posts that you're trying to embed 04:16.000 --> 04:19.000 are similar, semantically similar, in terms of content. 04:19.000 --> 04:23.000 Marimo is a reactive notebook in Python. 04:23.000 --> 04:25.000 It's shareable as an application, 04:25.000 --> 04:28.000 and running your browser as it wasn't powered, 04:28.000 --> 04:30.000 standalone HTML file. 04:30.000 --> 04:33.000 Let me pose a second, tell you again. 04:33.000 --> 04:37.000 You can clone my codes. 04:37.000 --> 04:40.000 You can customize the algorithm right your own. 04:40.000 --> 04:43.000 You can serve it on a solar powered, 04:43.000 --> 04:46.000 Raspberry Pi HTML server, 04:46.000 --> 04:49.000 HTTP server, maybe that's a bit too much, 04:49.000 --> 04:50.000 but you can do that. 04:50.000 --> 04:52.000 The code is not going to run there, 04:52.000 --> 04:54.000 somebody else is going to download this file, 04:54.000 --> 04:57.000 run all the code on their own laptop locally, 04:57.000 --> 05:00.000 and it will run your own algorithm. 05:00.000 --> 05:02.000 This is how Marimo looks like. 05:02.000 --> 05:05.000 It's like a typical Python notebook. 05:05.000 --> 05:08.000 It's not the ideal user interface. 05:08.000 --> 05:11.000 This is how the application version looks like. 05:11.000 --> 05:14.000 All the code goes away and just the UI elements are there. 05:14.000 --> 05:17.000 So you start making something available to people. 05:17.000 --> 05:20.000 This is an extra feature, the green feature, 05:20.000 --> 05:23.000 well, people can compose and move different parts of the UI 05:23.000 --> 05:27.000 so that they can use them differently for their own use cases. 05:27.000 --> 05:30.000 In practice, what does the tool do? 05:30.000 --> 05:32.000 Embadding visualization. 05:32.000 --> 05:34.000 Here you can see embeddings from four different timelines. 05:34.000 --> 05:37.000 I took the embeddings, plotted them into the, 05:37.000 --> 05:40.000 this is my home timeline, the people I follow. 05:40.000 --> 05:42.000 This one is the local timeline. 05:42.000 --> 05:44.000 This is the federated one, 05:44.000 --> 05:48.000 and this is the one that I get when I look for the go for tag. 05:48.000 --> 05:50.000 Yes, I'm an other go for fun. 05:50.000 --> 05:52.000 So if you put all of these together, 05:52.000 --> 05:54.000 you can generate this map. 05:54.000 --> 05:56.000 It's kind of geographical. 05:56.000 --> 05:58.000 This is how I call the different parts. 05:58.000 --> 06:01.000 It's fun and nice and I spent a lot of time preparing that. 06:01.000 --> 06:06.000 The nice thing that it shows how posts from different timelines 06:06.000 --> 06:08.000 can actually be grouped together, 06:08.000 --> 06:11.000 and you can get recommendations from one timeline to the other, 06:11.000 --> 06:12.000 if you want. 06:12.000 --> 06:15.000 And how can you use this stuff? 06:15.000 --> 06:17.000 Apart from looking at these things, 06:17.000 --> 06:18.000 and saying for instance, 06:18.000 --> 06:20.000 this algorithm is a bit weird. 06:20.000 --> 06:22.000 Family and parenting embeddings close, 06:22.000 --> 06:25.000 are close to not say for work stuff. 06:25.000 --> 06:29.000 Just the heads up, the red ones are the public timeline. 06:29.000 --> 06:30.000 It's not my personal time. 06:31.000 --> 06:33.000 You can also do semantics search. 06:33.000 --> 06:34.000 You can take a sentence, 06:34.000 --> 06:37.000 such as I'm a fan of free software and everything open source, 06:37.000 --> 06:40.000 and get posts which are related to that thing, 06:40.000 --> 06:44.000 or you can get the post ID from the previous embedding plotting, 06:44.000 --> 06:47.000 and just give the number and see the post which are close to that. 06:47.000 --> 06:49.000 You can do re-ranking, 06:49.000 --> 06:52.000 that is, you can have your own numerical timeline, 06:52.000 --> 06:54.000 and in terms of performance, 06:54.000 --> 06:56.000 it will run on your laptop. 06:56.000 --> 06:58.000 I tried it on a 2016 MacBook. 06:58.000 --> 07:00.000 It runs in 40 seconds, 07:00.000 --> 07:02.000 and you can embed your stuff. 07:02.000 --> 07:04.000 So it's fully local. 07:04.000 --> 07:06.000 You can embed stuff without connecting anywhere else. 07:06.000 --> 07:08.000 It runs in your browser. 07:08.000 --> 07:10.000 We only think you download these libraries 07:10.000 --> 07:12.000 that Mario wants to use, 07:12.000 --> 07:15.000 and then you only connect to your API 07:15.000 --> 07:18.000 and to the local embedding server. 07:18.000 --> 07:21.000 And this is a call for arms as I told you before. 07:21.000 --> 07:23.000 I would like it to grow as a tool 07:23.000 --> 07:25.000 that people can use to build stuff 07:25.000 --> 07:27.000 and easily share it with people. 07:27.000 --> 07:29.000 And once more, you will see the protocols 07:29.000 --> 07:31.000 not flat first thing. 07:31.000 --> 07:33.000 It would be great if you had something like a proxy 07:33.000 --> 07:36.000 that runs in between the client and any kind of server 07:36.000 --> 07:38.000 that does this thing for you. 07:38.000 --> 07:39.000 So this is it. 07:39.000 --> 07:41.000 Thanks a lot for hearing me. 07:41.000 --> 07:42.000 And thank you. 07:48.000 --> 07:50.000 Alright, thank you David. 07:50.000 --> 07:52.000 That was awesome. 07:52.000 --> 07:54.000 So we've had three short talks really close together. 07:54.000 --> 07:57.000 We haven't had time for questions for those. 07:57.000 --> 07:59.000 The speakers I'm sure will be available somewhere 07:59.000 --> 08:00.000 in the Fediverse. 08:00.000 --> 08:04.000 And we'd love to hear from you if you've got questions for them. 08:04.000 --> 08:07.000 Could folks just try and use a few more of these seats 08:07.000 --> 08:10.000 that are still sitting in the middle here 08:10.000 --> 08:12.000 because we've got lots of people coming and going 08:12.000 --> 08:13.000 which is wonderful. 08:13.000 --> 08:15.000 I want to make sure that everybody gets a chance 08:15.000 --> 08:17.000 to set comfortably. 08:17.000 --> 08:21.000 So for those of you that don't know me at all. 08:22.000 --> 08:27.000 My name is Andy Piper and I work with the MasterDone project 08:27.000 --> 08:30.000 but I also try and contribute across the Fediverse 08:30.000 --> 08:34.000 and I'm really excited to welcome our next speaker.