WEBVTT 00:00.000 --> 00:07.000 And here we go. 00:07.000 --> 00:08.000 How fun. 00:08.000 --> 00:09.000 Thank you. 00:09.000 --> 00:10.000 Everybody. 00:10.000 --> 00:11.000 A bunch of work. 00:11.000 --> 00:13.000 Can you hear me? 00:13.000 --> 00:14.000 Yeah. 00:14.000 --> 00:15.000 All right. 00:15.000 --> 00:16.000 Let's do this. 00:16.000 --> 00:17.000 Hey. 00:17.000 --> 00:18.000 I'm Stephen Hood. 00:18.000 --> 00:22.000 I lead open source AI projects for Mozilla. 00:22.000 --> 00:23.000 That's my email. 00:23.000 --> 00:25.000 That's hooded Mozilla.com. 00:26.000 --> 00:30.000 That's me receiving instructions for my 8-bit masters. 00:30.000 --> 00:35.000 I'm going to talk to you about Mozilla Builders today, which is a recent program we launched in the last year. 00:35.000 --> 00:39.000 And it is one of the ways that we sponsor open source projects. 00:39.000 --> 00:41.000 There are other ways, but this is one of them. 00:41.000 --> 00:43.000 I'll tell you what the project is. 00:43.000 --> 00:45.000 Why we're doing it. 00:45.000 --> 00:48.000 I'll tell you about some of the open source projects we funded to date. 00:48.000 --> 00:51.000 And the kind of things we're looking to fund in the coming year. 00:51.000 --> 00:53.000 And then talk about how we can finish the work together. 00:53.000 --> 00:58.000 If you're working on a project that fits into the mold here. 00:58.000 --> 01:02.000 Word of warning, this is going to be a lot of QR codes. 01:02.000 --> 01:04.000 You all love QR codes, right? 01:04.000 --> 01:05.000 They're beautiful, right? 01:05.000 --> 01:06.000 I hit them. 01:06.000 --> 01:09.000 But there's so many URLs and projects I want to share with you. 01:09.000 --> 01:12.000 There's no easier way to get that information to you. 01:12.000 --> 01:16.000 So that's off your phone because this is cool projects I want you to check out. 01:16.000 --> 01:19.000 Maybe become a contributor on. 01:19.000 --> 01:20.000 So what is Builders? 01:20.000 --> 01:26.000 Builders is a program by which we sponsor or code develop open source AI projects. 01:26.000 --> 01:33.000 That we think will help open source compete with closed source providers like OpenAI and 01:33.000 --> 01:35.000 Thropic Google and others. 01:35.000 --> 01:40.000 Now, before I move any further, you've probably noted the AI word in there. 01:40.000 --> 01:42.000 You're probably wondering why the AI part. 01:42.000 --> 01:44.000 Why is that in there? 01:44.000 --> 01:47.000 Why isn't just open source projects? 01:47.000 --> 01:50.000 And so like I said, this isn't the only way that we sponsor open source. 01:50.000 --> 01:53.000 But this project is right now about AI. 01:53.000 --> 01:55.000 And here's the reason why. 01:55.000 --> 02:02.000 As a company, when we look at tech right now, we have a sense of deja vu about what's going on right now. 02:02.000 --> 02:07.000 So looking around, I'm scared to say how few of you probably remember 02:07.000 --> 02:08.000 the NetScape days. 02:08.000 --> 02:10.000 But some of you do. 02:10.000 --> 02:11.000 Yeah. 02:11.000 --> 02:12.000 Yeah. 02:12.000 --> 02:13.000 There's my people. 02:13.000 --> 02:14.000 All right. 02:14.000 --> 02:18.000 If you remember NetScape, then you remember how close we came to Microsoft, 02:18.000 --> 02:22.000 basically making the web a feature of Windows. 02:22.000 --> 02:23.000 Right? 02:23.000 --> 02:25.000 It came very, very close to happening. 02:25.000 --> 02:28.000 And open source is a big reason why it didn't happen. 02:28.000 --> 02:30.000 And we kind of feel like it's playing out again with AI now. 02:30.000 --> 02:32.000 And this is like a fundamental need technology. 02:32.000 --> 02:35.000 And whether or not we individually like it or not. 02:35.000 --> 02:36.000 It's not going away. 02:36.000 --> 02:38.000 I hate to break it to you. 02:38.000 --> 02:39.000 It's not going to go away. 02:39.000 --> 02:41.000 And it's already affecting the web pretty profoundly. 02:41.000 --> 02:45.000 If you've been online the last year or two, you probably noticed this happening with AI 02:45.000 --> 02:48.000 Slop, more and more bots than ever. 02:48.000 --> 02:52.000 Like we have to pay attention to this technology because it is affecting the web. 02:52.000 --> 02:54.000 And that's why Mozilla cares about it. 02:54.000 --> 02:58.000 We think of the irresponsible, not to push for this to be open source, 02:58.000 --> 03:01.000 just like the browser itself has been. 03:01.000 --> 03:03.000 And things are already happening. 03:03.000 --> 03:07.000 So if you look at the world of open source AI, there's a lot of stuff you can do today 03:07.000 --> 03:09.000 that you wouldn't be able to do otherwise. 03:09.000 --> 03:14.000 The fact that there were open models that existed all that you can run on your own computer 03:14.000 --> 03:15.000 that's because of open source. 03:15.000 --> 03:20.000 The fact that some of them don't require you to write a huge check to Jensen 03:20.000 --> 03:23.000 and fund his next yacht. 03:23.000 --> 03:25.000 You can just use the CPU already have. 03:25.000 --> 03:26.000 That's a big deal. 03:26.000 --> 03:27.000 That's because of open source. 03:27.000 --> 03:31.000 The fact that you can build apps that don't have to call out to a cloud provider, 03:31.000 --> 03:37.000 they don't have to share your personal data with some server out there in someone's farm. 03:37.000 --> 03:39.000 That's because of open source as well. 03:39.000 --> 03:44.000 Not to mention all the collaboration that we can all do together to improve these projects. 03:44.000 --> 03:46.000 So you've all probably heard about deep seek. 03:46.000 --> 03:48.000 It came out of a week or so ago. 03:48.000 --> 03:49.000 That's an example. 03:49.000 --> 03:51.000 All the other complexities aside. 03:51.000 --> 03:53.000 It's an example of open source doing what it's supposed to do, 03:53.000 --> 03:57.000 which is people building on each other's creations and innovations 03:57.000 --> 03:59.000 and then sharing those back out. 03:59.000 --> 04:04.000 But even so, things need to get a little simpler, I think. 04:04.000 --> 04:08.000 Anyone who's built, I need to open this for say I in the last year or so, 04:08.000 --> 04:14.000 you have experienced a neat grinder of choices you have to go through. 04:14.000 --> 04:15.000 What model may you use? 04:15.000 --> 04:17.000 What inference provider? 04:17.000 --> 04:19.000 What vector store? 04:19.000 --> 04:22.000 What orchestration tool? 04:22.000 --> 04:23.000 If any am I going to use? 04:23.000 --> 04:25.000 And on and on and on. 04:25.000 --> 04:27.000 Meanwhile, that's hilarious. 04:27.000 --> 04:28.000 That image did not load. 04:28.000 --> 04:31.000 That is the picture of a Japanese bullet train. 04:31.000 --> 04:33.000 You ought to say about word for it. 04:33.000 --> 04:38.000 So we've got open AI just like barreling through the country side at 400 kilometers an hour. 04:38.000 --> 04:40.000 Well, we're all like stuck in a traffic jam. 04:40.000 --> 04:44.000 Because they've got this clean, well-defined, easy to use API. 04:44.000 --> 04:47.000 That a lot of people just reach for by default. 04:47.000 --> 04:48.000 It's just easier. 04:48.000 --> 04:54.000 So we need to get open source to a place where that's actually the easiest choice to start with. 04:54.000 --> 04:56.000 And we're just not there yet. 04:56.000 --> 04:58.000 And of course, we can't do this ourselves. 04:58.000 --> 05:00.000 Just like we couldn't do the browser ourselves. 05:00.000 --> 05:04.000 So our theme when we launched this program last year was local AI. 05:04.000 --> 05:09.000 And that means projects that make it easier to build AI powered apps that run locally. 05:09.000 --> 05:12.000 Where the intelligence is actually running on your device. 05:12.000 --> 05:21.000 Because we figured that those sorts of projects would naturally be resistant to being captured by platforms. 05:21.000 --> 05:22.000 Right? 05:22.000 --> 05:24.000 And so these are seven areas. 05:24.000 --> 05:26.000 I'm going to do them enough time. 05:26.000 --> 05:28.000 You can take a picture and ask me about it later. 05:28.000 --> 05:33.000 These are seven areas where we thought open source was kind of lagging behind people like Open AI. 05:33.000 --> 05:35.000 And so we invested in projects in this area. 05:35.000 --> 05:36.000 We did it in two ways. 05:36.000 --> 05:42.000 We did it with an accelerator program where we invested in 14 different open source projects in these areas. 05:42.000 --> 05:48.000 It started in September and it ended in December with a in-person demo day in San Francisco. 05:48.000 --> 05:55.000 And it's attended by press, investors, technologists, academics, and peers. 05:56.000 --> 06:00.000 And all 14 of these teams had the chance to pitch and demo what they're working on. 06:00.000 --> 06:02.000 And they all left with some progress. 06:02.000 --> 06:05.000 Maybe more funding, more users or contributors. 06:05.000 --> 06:06.000 It was pretty cool. 06:06.000 --> 06:09.000 And you'll notice down there by the way in this little corner. 06:09.000 --> 06:11.000 You recognize that computer? 06:11.000 --> 06:12.000 Yeah. 06:12.000 --> 06:14.000 It's the same one. 06:14.000 --> 06:15.000 Same one. 06:15.000 --> 06:16.000 I brought it with me. 06:16.000 --> 06:17.000 And we used as a dumb terminal. 06:17.000 --> 06:20.000 Like a VT100 terminal that talked to an LLM. 06:20.000 --> 06:22.000 Like it's connected to an A's bulletin board system. 06:22.000 --> 06:24.000 So like everything is interconnected. 06:24.000 --> 06:26.000 And that's the level we're operating. 06:26.000 --> 06:28.000 So. 06:28.000 --> 06:29.000 Yeah. 06:29.000 --> 06:30.000 All right. 06:30.000 --> 06:32.000 So let me tell you about three of these projects just really quick. 06:32.000 --> 06:34.000 Just to give you a flavor. 06:34.000 --> 06:35.000 Again, there were 14 of these. 06:35.000 --> 06:36.000 So transformer lab. 06:36.000 --> 06:38.000 If you want to train your own model. 06:38.000 --> 06:39.000 You want to fine-tune it. 06:39.000 --> 06:41.000 You want to evaluate it. 06:41.000 --> 06:45.000 That requires you to set up an understanding bunch of tools and techniques. 06:45.000 --> 06:51.000 Transformer lab turns those all into a installable, gooey application. 06:51.000 --> 06:52.000 It's pretty cool. 06:52.000 --> 06:54.000 So you can go to this QR code. 06:54.000 --> 06:56.000 Go there GitHub install it. 06:56.000 --> 07:00.000 And you can be tuning your own model learning how it works. 07:00.000 --> 07:01.000 Pretty amazing. 07:01.000 --> 07:03.000 Another one is, please. 07:03.000 --> 07:06.000 This is actually, I believe this team is based in Europe. 07:06.000 --> 07:10.000 And what they're trying to do is create truly open data sets and models. 07:10.000 --> 07:12.000 And I don't mean open source. 07:12.000 --> 07:14.000 The way our friends at meta define it. 07:14.000 --> 07:17.000 I mean actually open source. 07:17.000 --> 07:20.000 So everything is disclosed and documented and reproducible. 07:20.000 --> 07:24.000 And for the data sets, it's all truly above board. 07:24.000 --> 07:29.000 So there isn't like undisclosed copyrighted data stolen from somewhere. 07:29.000 --> 07:34.000 It's all public domain, creative commons, things like that. 07:34.000 --> 07:39.000 During the program, they released a data set called Common Corpus, which notably is trained 07:39.000 --> 07:44.000 on a lot of non-English content and is fully open and their own first model. 07:44.000 --> 07:46.000 Based on that data set. 07:46.000 --> 07:48.000 And then Thia is the third example. 07:48.000 --> 07:53.000 This is an open source IDE that has AI assistance in it. 07:53.000 --> 07:55.000 And there's a lot of people doing this right now. 07:55.000 --> 07:57.000 I'm sure you've probably heard a lot of them. 07:57.000 --> 08:03.000 Their particular approach to Thia is that they are working with open models, running locally. 08:03.000 --> 08:05.000 And they keep the human in the loop. 08:05.000 --> 08:07.000 So it's not trying to just do the job before you. 08:07.000 --> 08:10.000 It's trying to help you use the developer to be more productive. 08:10.000 --> 08:15.000 And so during the builder accelerator, they launched their first version of this capability. 08:15.000 --> 08:19.000 And we're going to integrate it with MamaFile, which I'm branded for your convenience. 08:19.000 --> 08:21.000 We'll talk about that more in a minute. 08:21.000 --> 08:22.000 So that's the accelerator. 08:22.000 --> 08:25.000 The other thing we did is what we call collaborations. 08:25.000 --> 08:27.000 So collaborations are a little more organic. 08:27.000 --> 08:29.000 It's not like a program you apply to. 08:29.000 --> 08:32.000 They are what? 08:32.000 --> 08:33.000 That's really subtle. 08:33.000 --> 08:34.000 Thanks. 08:34.000 --> 08:35.000 All right. 08:35.000 --> 08:36.000 I'm going to speed up. 08:36.000 --> 08:38.000 There's so much to talk about. 08:38.000 --> 08:43.000 So these are one-on-one relationships we have with people who are building cool projects. 08:43.000 --> 08:45.000 And give some examples. 08:45.000 --> 08:46.000 Lomfile. 08:46.000 --> 08:47.000 So Lomfile. 08:47.000 --> 08:49.000 I love this project. 08:49.000 --> 08:50.000 I'm the project lead for it. 08:50.000 --> 08:54.000 And Justin Toney is the brilliant independent developer who did all the work. 08:54.000 --> 09:01.000 And what this does is it democratizes access to open models by making them ridiculously easy to use. 09:01.000 --> 09:03.000 Because it turns them into a single file executable. 09:03.000 --> 09:07.000 And you can download it and run it in any computer with no installation. 09:07.000 --> 09:09.000 Any hardware, GPU or not. 09:09.000 --> 09:12.000 And it will let you use that model right out of the box. 09:12.000 --> 09:15.000 And we did this by combining two projects we love. 09:15.000 --> 09:17.000 One of them is Lomfile CPP. 09:17.000 --> 09:20.000 Which if you're in this room, there's a good chance you've heard of. 09:20.000 --> 09:22.000 And love as much as we do. 09:22.000 --> 09:25.000 Also Cosmopolitan, which is Justin's project. 09:25.000 --> 09:30.000 And this adds the sort of single file multi-platform executable magic. 09:30.000 --> 09:33.000 And when you run a long file, you get an open model. 09:33.000 --> 09:35.000 But you also get everything else in the box. 09:35.000 --> 09:38.000 So you get a UI for it, you get an API server, 09:38.000 --> 09:40.000 and you get a scriptable component. 09:40.000 --> 09:42.000 Like it works from the command line. 09:42.000 --> 09:47.000 So you can bash and pipe to your hards content. 09:47.000 --> 09:49.000 This is a quick video. 09:49.000 --> 09:52.000 Just an example of the performance enhancements we're able to make. 09:52.000 --> 09:55.000 So these are both different copies of Lomfile. 09:55.000 --> 09:59.000 Doing the same thing, which is reading a scientific paper and generating a summary of it. 09:59.000 --> 10:01.000 On the left is an old version. 10:01.000 --> 10:05.000 On the right is the latest version that has all of the CPU performance enhancements. 10:05.000 --> 10:06.000 Justin did. 10:06.000 --> 10:10.000 And you can see it's almost done. 10:10.000 --> 10:11.000 Yeah, it's done. 10:11.000 --> 10:14.000 But on the left, the old version is still reading the paper. 10:14.000 --> 10:18.000 So it's not even done ingesting the context window that we've fed it. 10:18.000 --> 10:20.000 And again, this is running just on a CPU. 10:20.000 --> 10:22.000 There's no GPU in this machine. 10:22.000 --> 10:24.000 So it's still thinking. 10:24.000 --> 10:25.000 You get the impression. 10:25.000 --> 10:26.000 Finally it's done. 10:26.000 --> 10:29.000 It's going to generate its output now. 10:29.000 --> 10:32.000 So Lomfile, there's a lot going on here. 10:32.000 --> 10:34.000 We have a new API server working on. 10:34.000 --> 10:37.000 That should be faster and more stable. 10:37.000 --> 10:41.000 Then the one we have now, which we inherit from Lomfile CPU. 10:41.000 --> 10:48.000 This is pushing a larger effort, which is we want to just enable developers to build local AI applications. 10:48.000 --> 10:50.000 And we're also moving to a new governance model. 10:50.000 --> 10:52.000 We're moving away from sort of the BDFL model. 10:52.000 --> 10:54.000 Justin worked full time on this for a year. 10:54.000 --> 10:56.000 Now she's switched into part time. 10:56.000 --> 11:03.000 So if you are a C or C++ developer, if you're interested in this problem space, there's going to be opportunities to get involved in the project. 11:03.000 --> 11:05.000 Welcome you to reach out. 11:05.000 --> 11:07.000 Real quick, SQLiteVec. 11:07.000 --> 11:09.000 If you've heard of this, this is a cool project also through builders. 11:09.000 --> 11:12.000 Turn SQLite into a vector database. 11:12.000 --> 11:18.000 So you can use it to build rag applications that run on a user's computer locally. 11:18.000 --> 11:19.000 That's Elixir C as project. 11:19.000 --> 11:21.000 It's awesome work. 11:21.000 --> 11:26.000 And Web Applets is an attempt to do kind of like anthrophic artifacts or check 11:26.000 --> 11:28.000 GPT plugins as an open standard. 11:28.000 --> 11:32.000 So that we can make sure that open web standards continue to drive the evolution of 11:32.000 --> 11:36.000 applications online even if AI starts to get involved. 11:36.000 --> 11:38.000 So where are we going next? 11:38.000 --> 11:41.000 The theme for this year is the next user agent. 11:41.000 --> 11:43.000 And what do I mean by that? 11:43.000 --> 11:48.000 Well, what I mean is this that computing is changing whether we like it or not. 11:48.000 --> 11:52.000 So for 75 years, we have been building computers based around 11:52.000 --> 11:56.000 a fact which is that they can't understand us. 11:56.000 --> 12:00.000 So we put all the effort into making it easier for 12:00.000 --> 12:04.000 us to understand the computer to talk to the computer. 12:04.000 --> 12:08.000 And then like just last year that all changed. 12:08.000 --> 12:11.000 To some definition of the word understand, 12:11.000 --> 12:13.000 computers can not understand us. 12:13.000 --> 12:16.000 They can make sense of our world, 12:16.000 --> 12:20.000 our writing, our video and images now and sound, 12:20.000 --> 12:22.000 and they can take action on our behalf. 12:22.000 --> 12:27.000 And so what this means is things could change in some really weird ways. 12:27.000 --> 12:32.000 We built this gengotower of technology. 12:32.000 --> 12:35.000 These are all the things we built for 75 years 12:35.000 --> 12:38.000 that power everything in our computing world. 12:38.000 --> 12:41.000 And so that's kind of said it, I'll say it too. 12:41.000 --> 12:43.000 Like browsers are pretty complex. 12:43.000 --> 12:45.000 They're actually kind of like a computer inside a computer. 12:45.000 --> 12:47.000 You know, you've got like process management, you've got 12:47.000 --> 12:50.000 Wazzin, you've got JavaScript, you've got the rendering engine. 12:50.000 --> 12:52.000 It's like a computer. 12:52.000 --> 12:56.000 So it should be no surprise to us that browsers are going to get effected 12:56.000 --> 12:59.000 by this new technological change. 12:59.000 --> 13:01.000 And if we care about browsers and care about the web, 13:01.000 --> 13:03.000 we've got to just be thinking about it. 13:03.000 --> 13:07.000 We can't just leave that to the big for-profit companies. 13:07.000 --> 13:11.000 Because we can wake up one day and find that everything's changed around this. 13:11.000 --> 13:12.000 So this is what we're doing this year. 13:12.000 --> 13:15.000 We're funding projects in these areas. 13:15.000 --> 13:17.000 Take a picture of this if you want. 13:17.000 --> 13:18.000 And I'll time to go through them all. 13:18.000 --> 13:21.000 But the unifying theme is, as we develop this tech, 13:21.000 --> 13:24.000 we want to make sure that it leads to a healthier web. 13:24.000 --> 13:26.000 Because by default, that's not assured. 13:26.000 --> 13:29.000 We want to make sure the web continues to be a healthy place, 13:29.000 --> 13:31.000 even as the technology changes. 13:31.000 --> 13:33.000 All right, that's it. 13:33.000 --> 13:35.000 That's my email. 13:35.000 --> 13:37.000 Reach out to me if you're interested. 13:37.000 --> 13:38.000 This is our website. 13:38.000 --> 13:40.000 Lots to read about lots of good projects there. 13:40.000 --> 13:41.000 Thanks a lot. 13:41.000 --> 13:43.000 And so if it's not finished, 13:43.000 --> 13:45.000 it's because I wanted to have time for questions. 13:45.000 --> 13:47.000 We have five minutes for questions. 13:47.000 --> 13:49.000 If you want to leave the room, please do so. 13:49.000 --> 13:50.000 Downstairs. 13:51.000 --> 13:54.000 Remember, we have swagon on the table. 13:54.000 --> 13:56.000 The two tables downstairs in table here. 13:56.000 --> 13:57.000 Questions? 14:05.000 --> 14:06.000 Hi. Thanks for this presentation. 14:06.000 --> 14:10.000 I'm really interested in LLMs and what it's employers and how 14:10.000 --> 14:12.000 mushy-like and leverage this. 14:12.000 --> 14:15.000 And recently, I discovered that the web is like an experiment 14:15.000 --> 14:19.000 in Firefox to do like AI chatbot and some integrations. 14:19.000 --> 14:20.000 And I wanted to try it. 14:20.000 --> 14:23.000 And I could only connect it to proprietary cloud platforms, 14:23.000 --> 14:27.000 not to LML file, not to LML-CCP-CPP. 14:27.000 --> 14:30.000 And I managed and or I could connect it, 14:30.000 --> 14:32.000 but then it wouldn't work with LML-CPP 14:32.000 --> 14:34.000 local after modifying about config. 14:34.000 --> 14:36.000 And I'm just wondering what's going on, 14:36.000 --> 14:40.000 because I would expect to be able to use it local first 14:40.000 --> 14:42.000 with LML-CPP from Mozilla, 14:42.000 --> 14:44.000 and I don't understand what's going on with you. 14:44.000 --> 14:45.000 You can. 14:45.000 --> 14:47.000 So if you're having troubles that might have been a bug 14:47.000 --> 14:48.000 in nightly or something like that. 14:48.000 --> 14:51.000 So there's a talk later about the AI framework 14:51.000 --> 14:53.000 that we're building into Firefox. 14:53.000 --> 14:55.000 So that speaker can maybe address it more, 14:55.000 --> 14:57.000 but I would say just tread the leaves nightly. 14:57.000 --> 14:59.000 There's a developer flag you have to throw that 14:59.000 --> 15:01.000 makes local host an option. 15:01.000 --> 15:03.000 And then if you use LML file or LML-CPP, 15:03.000 --> 15:04.000 it should work. 15:04.000 --> 15:07.000 They have the same endpoints in the same API signature. 15:07.000 --> 15:10.000 So it's definitely our intention that you can use it locally. 15:10.000 --> 15:11.000 I do all the time. 15:11.000 --> 15:12.000 Yeah. 15:12.000 --> 15:13.000 Talk to me afterwards. 15:13.000 --> 15:14.000 That team. 15:14.000 --> 15:17.000 So if you have any questions, you can take a look at it. 15:17.000 --> 15:19.000 Thanks, Art, and more questions. 15:19.000 --> 15:20.000 I don't see any hint. 15:20.000 --> 15:22.000 Thank you very much. 15:22.000 --> 15:26.000 Remember, if you want to leave the room with bell stairs.