WEBVTT 00:00.000 --> 00:05.000 Hi, everybody. 00:05.000 --> 00:07.000 Hi, everybody. 00:07.000 --> 00:08.000 Is that on? 00:08.000 --> 00:09.000 Hi, everybody. 00:09.000 --> 00:12.000 And... 00:12.000 --> 00:15.000 OK, today is very, very exciting. 00:15.000 --> 00:19.000 And then we're doing a WLC about AI. 00:19.000 --> 00:22.000 And even more crazy is on this fight. 00:22.000 --> 00:26.000 So, and I'm getting the lowest, OK? 00:26.000 --> 00:28.000 And so, let's get cracking. 00:28.000 --> 00:32.000 So, I'm willing, I'm a deep... 00:32.000 --> 00:33.000 OK. 00:33.000 --> 00:35.000 So, I'm a software guy. 00:35.000 --> 00:38.000 So, and I'm a founder with deep computing. 00:38.000 --> 00:40.000 If I make some of this fight hardware, 00:40.000 --> 00:41.000 do it in the code. 00:41.000 --> 00:42.000 I found it. 00:42.000 --> 00:43.000 And then... 00:43.000 --> 00:45.000 And then, I'm a series about software. 00:45.000 --> 00:47.000 So, I start making all my money in software, 00:47.000 --> 00:49.000 but I lose it all in the hardware. 00:49.000 --> 00:52.000 So, it's just similar to today, JB. 00:52.000 --> 00:53.000 OK? 00:53.000 --> 00:56.000 So, now, we have the king here. 00:56.000 --> 00:59.000 And now, President of the WLC. 00:59.000 --> 01:04.000 And one of the most weird guy that I've met in the world, 01:04.000 --> 01:06.000 who has money. 01:06.000 --> 01:07.000 OK? 01:07.000 --> 01:10.000 And now, 223, we made an A-code, 01:10.000 --> 01:13.000 the most expensive RISFI laptop, $5,000. 01:13.000 --> 01:14.000 OK? 01:14.000 --> 01:17.000 And, and to the other way, I make another version, 01:17.000 --> 01:19.000 and much more cheaper, more affordable. 01:19.000 --> 01:21.000 But I should still. 01:21.000 --> 01:22.000 OK? 01:22.000 --> 01:23.000 The chapter is, I went to the U.N. 01:23.000 --> 01:25.000 And then, and then after that, 01:25.000 --> 01:27.000 I parallel with Framework. 01:27.000 --> 01:28.000 And then, start doing motherboard, 01:28.000 --> 01:30.000 and not going to do the whole laptop, 01:30.000 --> 01:33.000 because losing too much money. 01:33.000 --> 01:34.000 OK? 01:34.000 --> 01:36.000 So, and Framework is good. 01:36.000 --> 01:37.000 Good for you, the pin. 01:37.000 --> 01:39.000 I created it, modularized it. 01:39.000 --> 01:41.000 And we played the motherboard. 01:41.000 --> 01:43.000 And, you know? 01:43.000 --> 01:46.000 And you can do whatever you want to do with it. 01:46.000 --> 01:49.000 And you can even print the whole shell case yourself. 01:49.000 --> 01:51.000 I just do anything you like. 01:51.000 --> 01:52.000 OK? 01:52.000 --> 01:53.000 Very innovative. 01:53.000 --> 01:54.000 So, that is Framework. 01:54.000 --> 01:55.000 Modular Framework. 01:55.000 --> 01:58.000 And then, now, I start making the motherboard for them. 01:58.000 --> 02:00.000 The first one is for core. 02:00.000 --> 02:02.000 In 2020, 25. 02:02.000 --> 02:04.000 And then, the problem now, finding out, 02:04.000 --> 02:06.000 RISFI, SOC, they are coolers. 02:06.000 --> 02:08.000 People make the SOC, 02:08.000 --> 02:10.000 and they are not knowing where to sell to. 02:10.000 --> 02:11.000 What to do with it? 02:11.000 --> 02:13.000 So, and, and they don't, now, all the start, 02:13.000 --> 02:15.000 with the money, now, time, and money. 02:15.000 --> 02:18.000 So, and, don't need to think that. 02:18.000 --> 02:20.000 Next, push it about here. 02:20.000 --> 02:21.000 OK? 02:21.000 --> 02:24.000 So, and, and now, we've, uh, 2020, 02:24.000 --> 02:28.000 we have a chip-back version with a very powerful 50-toffs. 02:28.000 --> 02:29.000 50-toffs. 02:29.000 --> 02:30.000 What is that? 02:30.000 --> 02:31.000 Is that chip-back? 02:31.000 --> 02:32.000 I mean, it's cool. 02:32.000 --> 02:34.000 Is that, uh, it's a core. 02:34.000 --> 02:35.000 OK? 02:35.000 --> 02:37.000 So, and it's 2-dye. 02:37.000 --> 02:38.000 Doom. 02:38.000 --> 02:39.000 All right, so. 02:39.000 --> 02:40.000 Where is the AI class? 02:40.000 --> 02:41.000 The AI-550-toffs. 02:41.000 --> 02:43.000 So, you have an NPU. 02:43.000 --> 02:44.000 Yes. 02:44.000 --> 02:45.000 N CPU. 02:45.000 --> 02:46.000 Yes. 02:46.000 --> 02:47.000 N GPU. 02:47.000 --> 02:48.000 Yes. 02:48.000 --> 02:49.000 Yes. 02:49.000 --> 02:50.000 No. 02:50.000 --> 02:52.000 Well, well, OK. 02:52.000 --> 02:53.000 OK. 02:53.000 --> 02:55.000 So, and this is what it looks like. 02:55.000 --> 02:58.000 And come over to our wrist-by-bove, and take a look at it. 02:58.000 --> 02:59.000 You can take it apart. 02:59.000 --> 03:02.000 And where is it, uh, printed TSMC? 03:02.000 --> 03:04.000 Oh, yeah, yeah, yeah, yeah, yeah. 03:04.000 --> 03:05.000 Curve nanobita TSMC. 03:05.000 --> 03:06.000 Ten nanomins of TSMC. 03:06.000 --> 03:08.000 And you can have it in the homework? 03:08.000 --> 03:09.000 Yes. 03:09.000 --> 03:10.000 OK, good. 03:10.000 --> 03:11.000 How? 03:11.000 --> 03:12.000 How? 03:12.000 --> 03:13.000 It's just a motherboard like that. 03:13.000 --> 03:14.000 Put it in. 03:14.000 --> 03:15.000 Yeah. 03:15.000 --> 03:16.000 No. 03:16.000 --> 03:17.000 No. 03:17.000 --> 03:18.000 Yes. 03:18.000 --> 03:19.000 And the open source portet. 03:19.000 --> 03:20.000 And you use our ball. 03:20.000 --> 03:22.000 And figure out how to use it. 03:22.000 --> 03:24.000 And on the express or the NVLC. 03:24.000 --> 03:25.000 How waste for free. 03:25.000 --> 03:26.000 OK. 03:26.000 --> 03:27.000 So, so. 03:27.000 --> 03:28.000 So, so I can take my framework. 03:28.000 --> 03:29.000 He moves the machine. 03:29.000 --> 03:30.000 Yeah. 03:30.000 --> 03:31.000 Yeah. 03:31.000 --> 03:32.000 Yes. 03:32.000 --> 03:33.000 That's it. 03:33.000 --> 03:34.000 Yeah. 03:34.000 --> 03:35.000 West, whatever you want to work. 03:35.000 --> 03:37.000 Everything open source. 03:37.000 --> 03:38.000 OK. 03:38.000 --> 03:39.000 Now, even more crazy. 03:39.000 --> 03:40.000 OK. 03:40.000 --> 03:42.000 Twenty, twenty, six. 03:42.000 --> 03:43.000 Some crazy guy. 03:43.000 --> 03:45.000 Made a crazy SOC. 03:45.000 --> 03:49.000 With a customized instruction on the CPU. 03:49.000 --> 03:50.000 What does it mean? 03:50.000 --> 03:53.000 That means that you have, usually, a instruction. 03:53.000 --> 03:54.000 Wispy instruction. 03:54.000 --> 03:55.000 Low and store. 03:55.000 --> 03:56.000 Or add. 03:56.000 --> 03:57.000 A multiple and add. 03:57.000 --> 03:58.000 Whatever. 03:58.000 --> 04:01.000 Now, they have a weird, weird matrix. 04:01.000 --> 04:02.000 Multiply. 04:02.000 --> 04:03.000 Instructions. 04:03.000 --> 04:04.000 So. 04:04.000 --> 04:05.000 But is it still respite? 04:05.000 --> 04:06.000 Yes. 04:06.000 --> 04:07.000 But custom. 04:07.000 --> 04:08.000 Yes. 04:08.000 --> 04:09.000 Custom. 04:09.000 --> 04:10.000 Wispy. 04:10.000 --> 04:11.000 And now you took custom. 04:11.000 --> 04:12.000 My. 04:12.000 --> 04:13.000 So. 04:13.000 --> 04:14.000 Yes. 04:14.000 --> 04:17.000 So basically, if VLC later on, I want to 04:17.000 --> 04:18.000 Instructions. 04:18.000 --> 04:19.000 So. 04:19.000 --> 04:20.000 And. And they be. 04:20.000 --> 04:21.000 Start charging people. 04:21.000 --> 04:22.000 Have one billion dollars. 04:22.000 --> 04:23.000 And we can make a. 04:23.000 --> 04:25.000 Specify VLC chip. 04:25.000 --> 04:26.000 Go VLC. 04:26.000 --> 04:27.000 Chip. 04:27.000 --> 04:28.000 Just 30 million dollars. 04:28.000 --> 04:29.000 And I can have a. 04:29.000 --> 04:30.000 That's in the framework. 04:30.000 --> 04:31.000 Yes. 04:31.000 --> 04:32.000 When. 04:32.000 --> 04:34.000 March. 04:34.000 --> 04:35.000 Is it free? 04:35.000 --> 04:36.000 Yes. 04:36.000 --> 04:37.000 For VLC's free. 04:37.000 --> 04:39.000 Yes. 04:39.000 --> 04:40.000 Yes. 04:40.000 --> 04:41.000 If they. 04:41.000 --> 04:42.000 Only contribute. 04:42.000 --> 04:43.000 I become guilty. 04:43.000 --> 04:45.000 All your commit. 04:45.000 --> 04:47.000 I will take your money first. 04:47.000 --> 04:48.000 Every commit. 04:48.000 --> 04:51.000 I give you $10 back. 04:51.000 --> 04:53.000 Yes. 04:53.000 --> 04:56.000 Okay. 04:56.000 --> 04:58.000 So now, the portal now. 04:58.000 --> 04:59.000 We have a. 04:59.000 --> 05:00.000 Where this. 05:00.000 --> 05:01.000 Eat. 05:01.000 --> 05:03.000 Out of imaginations. 05:03.000 --> 05:04.000 How there. 05:04.000 --> 05:07.000 What the fuck with the software? 05:07.000 --> 05:08.000 Or, natives. 05:08.000 --> 05:10.000 You can find. 05:10.000 --> 05:11.000 Right? 05:11.000 --> 05:12.000 Why? 05:12.000 --> 05:13.000 Because one kiss? 05:13.000 --> 05:14.000 Yeah. 05:14.000 --> 05:16.000 Why no one can get a no idea. 05:16.000 --> 05:18.000 So, let's take a look. 05:18.000 --> 05:19.000 Okay. 05:19.000 --> 05:21.000 Now, and then I just say shit. 05:21.000 --> 05:22.000 I need it. 05:22.000 --> 05:27.000 I watch video all the time and say, I think, I mean, this, what's the best, before, I know 05:27.000 --> 05:29.000 about, I know nothing about JB. 05:29.000 --> 05:31.000 I know nothing about WLC. 05:31.000 --> 05:34.000 So, and I don't, I know Chrome Google, say, pack off. 05:34.000 --> 05:35.000 I don't want to deal with you. 05:35.000 --> 05:36.000 Go away. 05:36.000 --> 05:38.000 And, and office. 05:38.000 --> 05:39.000 Oh, come on. 05:39.000 --> 05:40.000 It's out of the office going to die. 05:40.000 --> 05:42.000 I'm not going to waste my time on this. 05:42.000 --> 05:44.000 So, the model is hard. 05:44.000 --> 05:46.000 So, when we are, I, well, only have a hardware. 05:46.000 --> 05:51.000 First, I'm reporting all the models and quantitize it and analyze it and whatever. 05:51.000 --> 05:53.000 And then the framework is even harder. 05:53.000 --> 05:55.000 You know, PyTorch, go away. 05:55.000 --> 05:56.000 Respire. 05:56.000 --> 05:57.000 No interest. 05:57.000 --> 05:58.000 Go away. 05:58.000 --> 05:59.000 Don't buckle my main nine. 05:59.000 --> 06:00.000 You know, I don't want to maintain it. 06:00.000 --> 06:05.000 And then, now we have all the CPP. 06:05.000 --> 06:08.000 All the, what, Nama, all Nama. 06:08.000 --> 06:10.000 But, I think I'll be all knows about it. 06:10.000 --> 06:12.000 We, we're hacking around the AI branch, you know. 06:12.000 --> 06:14.000 And, and we know all this pain. 06:14.000 --> 06:16.000 And now, probably, we have MSWL. 06:16.000 --> 06:21.000 So, we, this fight guy, an MPU guy is really miserable, you know. 06:21.000 --> 06:24.000 It's just putting and putting, I'm very similar to, 06:24.000 --> 06:30.000 to, we have seen, put into different RVV and code that. 06:30.000 --> 06:31.000 That's nice. 06:31.000 --> 06:32.000 This is the whole life. 06:32.000 --> 06:33.000 We have the same, okay? 06:33.000 --> 06:34.000 So, now, on VLC. 06:34.000 --> 06:35.000 All right? 06:35.000 --> 06:36.000 So, pass it on to... 06:51.000 --> 06:52.000 That's why I'm mute. 06:52.000 --> 06:53.000 Yeah. 06:53.000 --> 06:55.000 So, now that we have... 06:55.000 --> 06:56.000 Okay. 06:56.000 --> 06:58.000 We need a picture, come on. 06:58.000 --> 06:59.000 Of course, you know. 06:59.000 --> 07:03.000 So, now that we have, almost, open source laptop free 07:03.000 --> 07:06.000 and that you use them with five is how do we actually move 07:06.000 --> 07:07.000 everything on research? 07:07.000 --> 07:08.000 If you listen to my talk before, 07:08.000 --> 07:11.000 a ton of things that we're going on AV1 and on the x64 07:11.000 --> 07:14.000 and so on is to stop rewriting a research five, SIMD. 07:14.000 --> 07:18.000 But, yeah, one of the reasons to work on, 07:18.000 --> 07:20.000 you all know that, right? 07:20.000 --> 07:22.000 No, I'm gonna lose that. 07:22.000 --> 07:27.000 Unfortunately, a lot of money is invested in innovation 07:27.000 --> 07:29.000 today and in technology about AI, right? 07:29.000 --> 07:32.000 And, um, one of the things that buzzed me quite a bit 07:32.000 --> 07:38.160 bit is that most of those AI stuff are leading to a cloud service that you need pay and 07:38.160 --> 07:44.800 it's completely your non-free and you're more than non-free. You cannot know what's happening 07:44.800 --> 07:52.560 and they can cut you. So the question we heard is, can we actually run some AI things for 07:52.560 --> 07:59.680 something that is useful on device? So it's a bit more dawn today but it wasn't too much 07:59.680 --> 08:06.640 year and a half ago. So last year we did a huge demo at CS where we had like automatic 08:06.640 --> 08:12.720 translation, such a speech to text, right? So whisper which is okay, right? You've all known 08:12.720 --> 08:18.400 that on YouTube, it's CT on YouTube, we did better of course but mostly it was a translation 08:18.400 --> 08:24.960 from all those languages to a ton of languages. So most of you are either French Spanish, 08:24.960 --> 08:29.760 you English speaking Italian, right? And so you don't care too much because everything is dubbed 08:29.760 --> 08:36.560 anyway. But if you talk to a lot of people from Eastern Europe or people in India, South 08:36.560 --> 08:41.520 East Asia, they have a ton of issue because the ton of content is not translated, right? So the 08:41.520 --> 08:46.320 idea was to do, hey can we do all sorts of translation, right? And all that on your machine. So 08:46.320 --> 08:52.560 it's completely decorrelated, it's open source AI which is of course not open source, right? It's 08:52.640 --> 08:57.680 open weight model, right? It's open news or whatever you would call that but still it's better 08:57.680 --> 09:04.320 and we integrated what I didn't but some people more clever than I did and we did that demo and 09:04.320 --> 09:09.920 it was like a huge success. So now the question was can we, but the problem is in order to do that, 09:09.920 --> 09:14.080 you still need to have very expensive graphic cards which needs a lot of harm which is 09:15.040 --> 09:20.880 expensive except if you have a kidney to sell lately and so like if you have some very small 09:20.880 --> 09:27.520 risk 5 chips which were $400, right? So I'm saying that before the DDR race price. Now they're 09:27.520 --> 09:34.880 killing me. And one of the things we did was that we used a model for translation which was 09:34.880 --> 09:41.680 which took around 35 languages and outputted 130 languages out and of course like we had no idea 09:41.680 --> 09:47.680 if the output was great so we outsource the testing at CES. We took random people and say hey what's 09:47.760 --> 09:52.560 your language and we try with this now which actually worked very well. It's not amazing but 09:52.560 --> 10:01.040 compared to nothing it's a ton better. The problem is that yeah in order to do that you need to 10:01.040 --> 10:07.840 compile the whole DLC and you need to run it on all the AI frameworks and the all the AI frameworks 10:07.840 --> 10:12.880 are a nightmare right? Because everyone in AI is basically working now on something where you 10:12.960 --> 10:20.560 have a docker which is using 25 gigabytes of dependencies and on that works on one graphic card 10:20.560 --> 10:27.520 inside Azure or GCP or scaleware of course but anyway but that's exactly the problem right? 10:27.520 --> 10:33.280 Here you need to run something on all type of machines with people who don't update their drivers 10:33.280 --> 10:40.400 or people who don't use non-open source drivers or different type of cards right? So and now like 10:40.400 --> 10:46.560 there is this crazy guy which is even crazy out of me who write and say well I'm going to run that 10:46.560 --> 10:52.560 on riskfight on just like a DLC is not on riskfight yet and so that's basically what we did right? 10:54.240 --> 10:58.800 We had to clean all the AI branches because of course this was a demo and between demo 10:58.800 --> 11:06.000 and production there is quite a bit of work. Whisper is as you've seen is those all the dot CPP case 11:06.080 --> 11:13.200 is basically what we call GGML which is done by a guy who's probably Gregory G something like 11:13.200 --> 11:18.320 that's why he gets called GGML who's basically running all the inference on CPU and have 11:18.320 --> 11:22.800 acceleration for all the blasts right? So cube blasts open blasts and all those stuff right? All those 11:22.800 --> 11:28.000 configurations so you're faster and so we had to modify whisper to be able to compile with 11:28.000 --> 11:33.680 it's V5 and also modify also because one of the things that they don't tell you is that most of the 11:33.840 --> 11:39.520 time you don't have one model for AI right? You have one model per chip right? And so that's why 11:39.520 --> 11:43.760 you don't see that right because everyone is using Q there right? But the fact of like once you have 11:43.760 --> 11:48.400 the model you need to retranslate them right? It's a very annoying and even like with a whisper 11:48.400 --> 11:53.360 to CPP if you're running on villcan or on Q light it's not exactly same so you need to change it 11:53.360 --> 12:01.200 to run and to run on local models and then you need to accelerate everything using using the NPU so 12:01.280 --> 12:08.400 basically what we did was using whisper GGML tiny and DPS67 compile everything for with the SDK 12:08.400 --> 12:16.000 of ESWIN and test and yeah come to our booth we sell you how much the machine is and yeah and 12:16.000 --> 12:23.360 we have around five talking per second which is which is okay it's a bit too small and too slow 12:23.360 --> 12:27.520 we should have something like two or three times but the next generation we'll have it right? 12:28.480 --> 12:34.160 No you're going to do it well I only give the hardware okay and they pay you and every time 12:34.160 --> 12:43.200 they do one commit it's 10 dollars off of it he's on record right so you so anyway he is the 12:43.200 --> 12:50.400 guarantee okay anyway you should look at RIS5 for multimedia right it's a future it's 12:50.400 --> 12:56.560 also the only way that we're going to get out of X86 and our patent so you should look at RIS5 12:56.560 --> 13:02.080 and thank you very much with the best talks of FUSDAM! 13:27.440 --> 13:37.440 Yes I don't think it came here but why? Where does he buy it? No you just use TPSR I don't know 13:37.440 --> 13:53.760 the what? Hold it like a gear okay let me do it now okay where do you buy it? Yeah 13:53.760 --> 13:59.920 I couldn't go go and subscribe our open source community sponsorship program and we give it to you 13:59.920 --> 14:08.560 yeah and then you just write your proposal I'm going to go to Goethe's VLC port on on on the just now 14:08.560 --> 14:12.880 that thing you okay the RIS5 CPU you cast in my AI instruction then you would get it for me 14:14.720 --> 14:18.480 any any more questions another question another question another question yes 14:24.240 --> 14:35.440 and I repeat yeah I repeat the questions and and why we we didn't use green or some other 14:35.840 --> 14:41.680 mistrol or whatever model actually we can use whatever model is but you need to help us to put it 14:43.280 --> 14:47.840 but that's part of the fun okay so that's why we have all the sponsorship program and then you 14:47.840 --> 14:54.480 guys predict the best STT model or TDS model or or whatever our model they're pleased that we have 14:57.200 --> 15:03.760 there is also the issue is that so example the first version of VLC on translation was a framework 15:03.760 --> 15:10.720 called seamless M T4 M 42 which is something no one knows but which was the one that has the most 15:10.720 --> 15:16.160 language support right so it's also the problem right is that you need to have 30 languages in and around 15:16.160 --> 15:21.120 wanting it out and that is one of the most difficult part and to be honest like just knowing 15:21.120 --> 15:25.520 which is the right model is probably full time right there is a new model every two weeks right 15:25.520 --> 15:30.080 and it's better than the previous one but it's very difficult to test right because I would need 15:30.080 --> 15:34.880 100 people from wanting the different languages to tell us which one works well or doesn't 15:36.960 --> 15:43.920 another last question one two three you should go down on the VLC boost to get some nice 15:44.000 --> 15:48.160 goodies because it's the best boost ever and you should be on the wrist five boost so that you can 15:48.160 --> 15:53.760 see the demo of this laptop that you can buy to replace your framework CPU and use a wrist five 15:53.760 --> 15:57.440 CPU thank you very much