WEBVTT 00:00.000 --> 00:13.000 Hello everyone, good evening. Our next speaker is Rex Ha, he is a research scientist at 00:13.000 --> 00:19.000 Homeroom and he is going to talk and let us know more about building local AI with full 00:19.000 --> 00:26.000 tech approach, give a warm round of applause. 00:26.000 --> 00:36.000 Okay, hello, first of them, I'm Rex, a researcher at Homeroom. We are by Zincapar. 00:36.000 --> 00:42.000 Today, I will take you on a journey through an anatomy of thinking machine, exploring how 00:42.000 --> 00:49.000 local interact with us in both hardware and software levels. 00:49.000 --> 00:55.000 A little bit about us, we are building complete environment for open source, local AI. 00:55.000 --> 01:04.000 Our focus is on running AI locally from serving our MS on servers to laptops to even 01:04.000 --> 01:11.000 edge device like mobile robots. We also train a few integrations of a real-time 01:11.000 --> 01:18.000 multi-moder models. We have learnt a few products like 01:18.000 --> 01:26.000 John, Chachibity, Affinity, that's the application that lets you run over any open source 01:26.000 --> 01:33.000 model on your machines. If you prefer to host it yourself, we have Cortex, a local 01:33.000 --> 01:43.000 AI platform. We also release a new model, a family of speech-language models that 01:43.000 --> 01:51.000 can respond to you in a real-time. Additionally, we are also experimenting with robotics simulations 01:51.000 --> 02:01.000 as well. We are basically everything for AI enthusiasts, small businesses and enterprise 02:01.000 --> 02:08.000 need to run a local LM. This brings to the full control for you to easily modify 02:08.000 --> 02:17.000 model knowledge, to adapt with your need and eliminate concern about data privacy. 02:17.000 --> 02:23.000 After focusing on local inference, we wanted to create our own model. 02:24.000 --> 02:30.000 When we look at the foundation of the model, it's too expensive. We can not 02:30.000 --> 02:37.000 complete with mixaxe like meta-open AI or Google, so we think of a different way 02:37.000 --> 02:46.000 that we utilize. How about you train your LM? I saw a few hands, but it's good. 02:46.000 --> 02:53.000 How about you in the multi-model, LM? Oh, I see no. Oh, I see over there are some people. 02:53.000 --> 02:59.000 Well, you guys are like really deep into the field. But then come to the hardware. 02:59.000 --> 03:07.000 Like, have you built your own desktop with Chibiu? Oh, a lot. 03:07.000 --> 03:12.000 And like, you build a Chibiu cluster? Oh, 03:13.000 --> 03:18.000 actually, I haven't built one. Like, like, my other college built it for me to train 03:18.000 --> 03:26.000 model. I know it really know about it. And last one, have you debug X100 cluster for inference 03:26.000 --> 03:34.000 and FI-2 name? Wow, you're rich. You're rich. 03:34.000 --> 03:40.000 Well, I have, I like, that one is like my pen pie of every night. I have to like be 03:40.000 --> 03:47.000 in stuff. Yeah. Okay. So, great. Some of you have, like, I saw a lot. So, you have, like, knowledge, 03:47.000 --> 03:53.000 technically one knowledge, but some dozen so I, I may have everyone to cast up. 03:53.000 --> 03:59.000 So, I will do a quick recap on LM sense of with, like, really high level of restrictions. 03:59.000 --> 04:08.000 So, yeah. Okay. So, what is LM? But attention, guys, researchers out there, please cover your eyes 04:08.000 --> 04:14.000 because I will, like, do a really simple, seriously over simplification for this. 04:14.000 --> 04:20.000 So, please cover your eyes. You already know it. Okay. First, we have a neural network 04:20.000 --> 04:27.000 layers, just for my layers. You see, like, me making up the way you digitalize our version 04:27.000 --> 04:34.000 of our brain. And at a perfect, not really, they're not perfect. They're not really perfect 04:35.000 --> 04:41.000 of, like, me making our brain. But do they work? Yes, probably. But can they handle a lot 04:41.000 --> 04:48.000 of data? Oh, hell, yeah. They, like, they have all of the internet of us. Right. 04:48.000 --> 04:54.000 And then, like, there's an encoding layers, which is, like, a component that can help 04:54.000 --> 05:02.000 to get the model to translate human language into a machine language. And then, we have, like, 05:02.000 --> 05:08.000 the decoder layers, which is translate back the after calculations and stuff. It 05:08.000 --> 05:14.000 will translate the machine language back to the human language. Yeah. But the main focus 05:14.000 --> 05:20.000 of this talk is on how this AI can run on our machine. So, the key thing you will need to know 05:20.000 --> 05:28.000 that these layers, just the stack of components, not let the machine can think. Okay. 05:28.000 --> 05:35.000 So, another piece of knowledge that you may need to know is that what exactly inside of the 05:35.000 --> 05:42.000 model. So, let's say, when you were, say, you were running a mixture of llama, what 05:42.000 --> 05:48.000 does that really mean? Like, there are different versions of mixture. You can see, like, this 05:48.000 --> 05:55.000 one is, like, 4 gigabytes of storage. And at one, it's, like, 24 gigabytes. So, what is the 05:55.000 --> 06:02.000 value difference here? The answer lies in the models parameters. The more parameters 06:02.000 --> 06:09.000 the means, like, the is larger and also is smarter. So, it comes with, like, we need a lot 06:09.000 --> 06:17.000 of, we need a lot more hardware. Another factor is that quantization model is not linearly. 06:17.000 --> 06:24.000 You can see, it's like a logarithmic row. So, scaling up the models and just sustain 06:24.000 --> 06:33.000 approach. But let the smaller models, smarter, is the way. So, and I, and believe me, this graph 06:33.000 --> 06:39.000 which was through in the last year. But this year, smaller models, like, we saw that 06:39.000 --> 06:47.000 two when, seven, be llama, three, they are all really good. And we can do use it for the 06:48.000 --> 06:56.000 basis. So, as expected, with a previous point, bigger models, require more memory. So, 06:56.000 --> 07:03.000 if we want to serve, let's say, eight users in the row. The V-ramlet, it increase even more. 07:03.000 --> 07:09.000 And if we want to run it with, like, flow 16 for the better performance, the GPU needed 07:09.000 --> 07:15.000 is going to be insane, right? Yeah. And this is why, and we are making so much money this 07:15.000 --> 07:21.000 way. And, yeah. And also, that's why, like, multiple companies right now, like Amazon, 07:21.000 --> 07:28.000 Google, or even some startups. They are trying to develop their own hardware to, supposedly, 07:28.000 --> 07:37.000 to influence the language model. So, to make it more practical, you want the most intelligent, 07:37.000 --> 07:44.000 I, I, to serve you at home. So, how many GPUs do you need to run 80, 70 million 07:44.000 --> 07:54.000 model on a consumer cart like, 30, 60 is, like, it cannot. But for, like, even, like, we 07:54.000 --> 08:02.000 100, you need, like, 16, but for, 100, you need, like, a row for, but those are not, like, 08:02.000 --> 08:08.000 simple machine. And you can see on the right here is a picture of my colleagues house 08:08.000 --> 08:14.000 before we move everything into the data center. If you can see, like, when we run 08:14.000 --> 08:21.000 ding a 70, be model, into our internal use, it could hit, like, his house up for a whole 08:21.000 --> 08:28.000 winter. Yeah. And, like, if you, an Apple fan, you might think that you can 08:28.000 --> 08:39.000 start getting max to run the models. Okay. You are stuck in to run the model. Yeah. 08:39.000 --> 08:44.000 Actually, this is, like, you can see, they actually using it to run the model. But, 08:44.000 --> 08:49.000 not just getting, please, no, don't, you do not run your house like that, like, please, 08:49.000 --> 08:56.000 don't, please, don't do this at home. Okay. But, but this view is lush enough, that company 08:56.000 --> 09:02.000 is, like, I said, before, like, start up congrock, they introduced new, new thing, new 09:02.000 --> 09:07.000 hardware, cone LPU, which is the hardware specifically to design for the last 09:07.000 --> 09:14.000 wish model inference. So, this makes the prediction progress insanely fast. So, all 09:14.000 --> 09:20.000 of those just to support a beefy model, that's now even more talkative. Recently, we have 09:21.000 --> 09:28.000 heard a lot about, like, open AI 01. Yesterday is, like, O3 and the tsunami of AI 09:28.000 --> 09:35.000 advancement in the west cone, beefsick. They are own, like, the success, like, lay on 09:35.000 --> 09:41.000 the technique called test time computes. So, which mean, essentially, is, like, when 09:41.000 --> 09:47.000 you let the model model thing longer, or let it, like, talk longer. And, hopefully, it 09:47.000 --> 09:57.000 becomes smarter. But, well, yeah, it's not a way the case. Okay. Okay. But, the 09:57.000 --> 10:03.000 mentality away here is that, we, we all know that techniques, these techniques will 10:03.000 --> 10:11.000 be more viral, and also it consume even more GPU time. So, this brings to another aspect. 10:12.000 --> 10:18.000 Okay. So, GPUs now would, like, not only need more size, like, bigger size, 10:18.000 --> 10:24.000 they also need to calculate things faster. So, you have probably heard of 10:24.000 --> 10:29.000 such a flop on the news when every day, or modes there and GPU, right? 10:29.000 --> 10:36.000 These measure how many floats by operations are up a machine can perform. So, 10:36.000 --> 10:43.000 more importantly, we found that, when we testing on our, when we develop the app, 10:43.000 --> 10:49.000 we found that the memory bandwidth is crucial for faster inference. Inside 10:49.000 --> 10:55.000 of the GPU, a lot of data change flows happen between SRAM and HBM. So, increasing 10:55.000 --> 11:03.000 of memory bandwidth boost data. And, yeah, we have another option, if we don't want to 11:03.000 --> 11:09.000 choose NVIDIA GPU, you don't want NVIDIA to get rich. So, you can switch to 11:09.000 --> 11:17.000 Snapdragon chips, which extend the AMD include NBUs. Okay. So, next up, 11:17.000 --> 11:25.000 people want to move to the desktop workstation. And, the car is more 11:26.000 --> 11:32.000 about, so, which is communicate with GPUs and all the components. Here, 11:32.000 --> 11:37.000 you can put a multiple things into onto the motherboard and how it doesn't 11:37.000 --> 11:43.000 break. These components are typically less and provide more computing 11:43.000 --> 11:50.000 power. And, the next level, we scale up to servers. They make a better 11:50.000 --> 11:57.000 components offering, we can run it 24 hours per day. And, each of the 11:57.000 --> 12:04.000 unit is designed to be modular. So, we have multiple CPU and multiple GPUs 12:04.000 --> 12:11.000 connected together. This could give you, okay. So, at the right side, 12:11.000 --> 12:17.000 could give you a roughly ideal of what needs the server. This is a really old 12:17.000 --> 12:24.000 architecture, up to this day. But, yeah, you can look at it as a reference. 12:24.000 --> 12:29.000 So, when you have a lot of money, you don't want, you don't have to use it. 12:29.000 --> 12:35.000 You can buy multiple server and connect it into a data center 12:35.000 --> 12:44.000 racks. This will connect it through infinity and buildings. Okay. So, those are all 12:44.000 --> 12:52.000 of the intuition I want to give about hardware. So, let's come to the main of 12:52.000 --> 12:57.000 the talk. So, this is the talk is about like the Antonomy of machine 12:57.000 --> 13:03.000 doing inference. Okay. So, I just want to say something. We have learned 13:03.000 --> 13:08.000 over the past 20 months of starting the project. As of this week, 13:08.000 --> 13:15.000 we have been a lot of it over 2,000,000 times. And, well, we have gotten at least 13:15.000 --> 13:22.000 86,000 cross reports. And, this is only count because we enable the crossing 13:22.000 --> 13:27.000 part over the past three months. So, over like, what are the half year, we have 13:27.000 --> 13:33.000 a lot of users across like crazy out there. So, yeah, you can see here, 13:33.000 --> 13:39.000 AI is still in this infancy. We are solely in the game. If you hear, 13:39.000 --> 13:45.000 I'm talking about AI, I'm inferring things today. It's still really early. 13:45.000 --> 13:51.000 And, I met that 30 years from now. AI will be much more seamlessly integrated 13:51.000 --> 13:58.000 into our machine. And, your laptop will feel more human light than ever. So, yeah. 13:58.000 --> 14:05.000 So, let's talk about how machine with single CPU inference requests. 14:05.000 --> 14:10.000 When you ask a model to think, what happened in the hardware's field? 14:10.000 --> 14:18.000 Okay. Let's pretend that I'm a machine. And, I want you to ask me, 14:18.000 --> 14:24.000 what is Jan? Then, if I want to generate the tokens like Jan, 14:24.000 --> 14:30.000 he says something something. The first thing first, I asked a CPU. 14:30.000 --> 14:35.000 Let's make the function code to retrieve the model weight from this, 14:35.000 --> 14:40.000 which is like HDD or RSD. I have to pull another model. 14:40.000 --> 14:44.000 And, let's say, it's like foggy by a storage size of the model. 14:44.000 --> 14:48.000 So, I have pulled it into the CPU. 14:48.000 --> 14:54.000 Yeah. On the other side, I want my simple step diagram. 14:54.000 --> 14:59.000 So, I draw it to simplify everything to for people to understand, 14:59.000 --> 15:03.000 people that are inside of those wire and chips. 15:03.000 --> 15:09.000 Okay. Then, after all of the model weight load from this, 15:09.000 --> 15:14.000 let's assume that we will have a latch in our V-Dram GPU. 15:14.000 --> 15:20.000 So, I want every single layer of the model onto the GPU. 15:20.000 --> 15:29.000 And then, when, and then now it's like, your problem right now is coming through over. 15:29.000 --> 15:41.000 Okay. At first, the prompt will be like tokenize our simple way to translate into the machine language. 15:42.000 --> 15:46.000 And then, it will execute through transformer layers. 15:46.000 --> 15:50.000 Usually, it's in parallel in the GPU and then, 15:50.000 --> 15:55.000 generate a possible predictions. 15:55.000 --> 16:01.000 And the final prediction, the next squad, is met by the CPU, 16:01.000 --> 16:04.000 which, as a look-up table to find the best, 16:04.000 --> 16:09.000 then the most suitable word for the given context. 16:09.000 --> 16:13.000 A response is then returned to the user. 16:13.000 --> 16:20.000 So, well, this is a journey of how a simple question from a user to the machine 16:20.000 --> 16:24.000 through everything, it thinks, and it will respond. 16:24.000 --> 16:29.000 For the next token, it will continue the process again and again, 16:29.000 --> 16:33.000 until it reached the end of the answer. 16:34.000 --> 16:37.000 Well, that was an easy one, right? 16:37.000 --> 16:43.000 But what if you have like a multiple GPUs and one to use a bigger model, 16:43.000 --> 16:47.000 which doesn't fit into the single GPU, right? 16:47.000 --> 16:52.000 Same first step, we also need the model to load into the happiest. 16:52.000 --> 16:56.000 From the happiest into the CPU. 16:56.000 --> 17:00.000 And the CPU wishes for the model weight into the GPUs. 17:00.000 --> 17:03.000 And here is the interesting part. 17:03.000 --> 17:08.000 The model will fit up, and yes, and yes, 17:08.000 --> 17:11.000 it will distribute the layers into GPUs. 17:11.000 --> 17:16.000 So, like the CPU right now is like GPU 1. 17:16.000 --> 17:18.000 Do you got this one? 17:18.000 --> 17:20.000 Do you got this layers, GPU 2? 17:20.000 --> 17:22.000 You got this layer. 17:22.000 --> 17:25.000 Depends on the algorithm, do you want to use these 17:25.000 --> 17:28.000 are like the two main common algorithm? 17:29.000 --> 17:36.000 Which we have like pipeline parallelism, which is like we split the model horizontally 17:36.000 --> 17:38.000 between GPU 1 and GPU 2. 17:38.000 --> 17:41.000 So, it's your like, it's your own half like 16 of layers, 17:41.000 --> 17:43.000 and other one have 16. 17:43.000 --> 17:47.000 And the other one is like TensorFlow parallelism, 17:47.000 --> 17:50.000 which split the model vertically. 17:50.000 --> 17:55.000 So, layer will be distributed in GPU 1 and 2 accordingly. 17:56.000 --> 18:00.000 So, everything works nearly the same as in the single GPU right. 18:00.000 --> 18:03.000 So, the model is now nearly the same. 18:03.000 --> 18:08.000 The model is now encoded into the machine language with some layers 18:08.000 --> 18:11.000 running through the GPU 1 and some like, 18:11.000 --> 18:14.000 when we run through GPU 2. 18:14.000 --> 18:17.000 And after processing the results are combined. 18:17.000 --> 18:20.000 And the final prediction is met. 18:20.000 --> 18:22.000 It's like nearly 100%. 18:23.000 --> 18:25.000 Like, it's 90% of similar to the, 18:25.000 --> 18:28.000 we got a one single GPU. 18:29.000 --> 18:33.000 And finally, when we stack up the GPU stacks together, 18:33.000 --> 18:36.000 we become a multi-note. 18:40.000 --> 18:41.000 Okay. 18:41.000 --> 18:45.000 So, if you have a multi-note scaling of HPC data center. 18:45.000 --> 18:46.000 So, okay. 18:46.000 --> 18:49.000 To be honest, like, as I said before, 18:50.000 --> 18:54.000 I have not, like, fully understanding about what this is. 18:54.000 --> 18:58.000 Because, like, my hackware colleague will do all of this. 18:58.000 --> 19:04.000 But the main thing we want to check here is that all of the other companies 19:04.000 --> 19:07.000 out there, like, open AI, like Google. 19:07.000 --> 19:09.000 It holds, they hold, 19:09.000 --> 19:12.000 charge it with the audio menu on, like, 19:12.000 --> 19:17.000 for our users on this kind of setting. 19:18.000 --> 19:21.000 So, let's come with me on this. 19:21.000 --> 19:25.000 So, let's start, like, at our former beginning, 19:25.000 --> 19:27.000 we always have to load now. 19:27.000 --> 19:32.000 I will talk about more advanced techniques that, 19:32.000 --> 19:35.000 right now, the companies out there, 19:35.000 --> 19:38.000 we've used to serve their models. 19:38.000 --> 19:41.000 First thing first, we have the KVCatch. 19:41.000 --> 19:45.000 Since the code is usually autoregressive, 19:46.000 --> 19:49.000 which means each of the newly predicted tokens 19:49.000 --> 19:52.000 only depends on the previous token. 19:52.000 --> 19:54.000 So, at each generation step, 19:54.000 --> 19:59.000 we are recanculating every attention for the previous tokens. 19:59.000 --> 20:03.000 Which is not the most efficient way, right? 20:03.000 --> 20:05.000 If we calculate it one, 20:05.000 --> 20:07.000 we can cache all of those, 20:07.000 --> 20:11.000 and then just use those to calculate the next token. 20:12.000 --> 20:18.000 That's a really high level of what is KVCatch working. 20:18.000 --> 20:23.000 This is the common text that anthropic open AI 20:23.000 --> 20:26.000 and also Google used to serve the user, 20:26.000 --> 20:30.000 because this technique can reduce, like, 20:30.000 --> 20:33.000 90% of the inference time, 20:33.000 --> 20:39.000 if we end just easy way to enable this one. 20:39.000 --> 20:42.000 Another technique is that speculative decoding, 20:42.000 --> 20:46.000 where we use a last main model with, 20:46.000 --> 20:48.000 like, the best performance we can have, 20:48.000 --> 20:51.000 alongside with a smaller model. 20:51.000 --> 20:54.000 The reason for a smaller model, 20:54.000 --> 20:56.000 a small model, right here, 20:56.000 --> 21:00.000 is that it can handle simple, repetitive tasks, 21:00.000 --> 21:03.000 and quickly generate token inspiration, 21:03.000 --> 21:08.000 and it could significantly reduce the inference time. 21:10.000 --> 21:14.000 Okay, so this is the comparison from open AI, 21:14.000 --> 21:17.000 the speculating decoding. 21:17.000 --> 21:20.000 The right size, you can say, is really fast. 21:20.000 --> 21:22.000 But, like, the left side, 21:22.000 --> 21:25.000 if you only use the big model, 21:25.000 --> 21:31.000 it takes ages to just generate the answer. 21:31.000 --> 21:34.000 Okay, it's finished, thank you. 21:35.000 --> 21:38.000 Okay, so this is, you can see, 21:38.000 --> 21:40.000 like, given Android Compatty, 21:40.000 --> 21:42.000 he's, like, a top lid of, 21:42.000 --> 21:46.000 X, the X, open AI lid. 21:46.000 --> 21:50.000 So, he took about a beauty of this technique. 21:50.000 --> 21:53.000 Okay, so, after, like, 21:53.000 --> 21:56.000 looking down all of the steps, 21:56.000 --> 21:59.000 like, how the normal text base, 21:59.000 --> 22:02.000 generate if AI work within our machine. 22:02.000 --> 22:05.000 Now, we, if we want to add one more step, 22:05.000 --> 22:07.000 like a sci-fi movies, 22:07.000 --> 22:10.000 where we can command the hour machine, 22:10.000 --> 22:12.000 then it will understand with our voice, 22:12.000 --> 22:15.000 and it will respond back. 22:15.000 --> 22:17.000 What would it look like? 22:17.000 --> 22:21.000 So, without in my, we started project isigo. 22:25.000 --> 22:29.000 Oh, I don't think the audio go through. 22:29.000 --> 22:39.000 Uh, can we somehow get the audio hidden? 22:39.000 --> 22:42.000 A little bit of technical help. 22:51.000 --> 22:53.000 Maybe I will just, like, 23:00.000 --> 23:04.000 hey, this is itigo. 23:04.000 --> 23:06.000 The local real-time voice AI, 23:06.000 --> 23:09.000 it can understand human speech and talk back. 23:09.000 --> 23:11.000 And with the latest checkpoint, 23:11.000 --> 23:13.000 you can run this little guy on your device, 23:13.000 --> 23:15.000 and it's fully open soft. 23:15.000 --> 23:17.000 But your scene here is the demo version of it, 23:17.000 --> 23:19.000 and it's running on just one 3090. 23:19.000 --> 23:22.000 So, let's give it a try. 23:22.000 --> 23:27.000 Hey, it'sigo, what's the meaning of strawberry in Japanese? 23:27.000 --> 23:30.000 Strawberry in Japanese is called itigo. 23:30.000 --> 23:34.000 It's a sweet and juiciness that's very popular in Japan. 23:34.000 --> 23:36.000 Here we go. 23:36.000 --> 23:41.000 Okay, so all of this is just like running on a single 3090 machine, 23:41.000 --> 23:43.000 and it reads, respond in real-time. 23:43.000 --> 23:46.000 And like, this is the only, like, a project of three months or less, 23:46.000 --> 23:51.000 is actually thinking around and trying new stuff. 23:51.000 --> 23:55.000 Okay, I will go quickly, go through the intuition of how it works 23:55.000 --> 23:57.000 to speed up the top-congeneration. 23:57.000 --> 24:02.000 So, that's why we can see this general thing is really fast. 24:02.000 --> 24:06.000 And this is an interesting fact, 24:06.000 --> 24:10.000 because like, we train our isigo entirely on text data, 24:10.000 --> 24:12.000 not a single audio file. 24:12.000 --> 24:14.000 Yes, you, you hold it right. 24:14.000 --> 24:17.000 We're training a speech understanding model 24:17.000 --> 24:20.000 without any audio files. 24:20.000 --> 24:22.000 How, how it goes. 24:23.000 --> 24:29.000 It actually, like, we, we did it with the simple model, 24:29.000 --> 24:32.000 like, transform model, transform model, 24:32.000 --> 24:36.000 which ended with, like, the input essay, text English, 24:36.000 --> 24:38.000 Germany, our French. 24:38.000 --> 24:42.000 We input it both, all of them into the model, 24:42.000 --> 24:45.000 and let it predict the speech tokens. 24:45.000 --> 24:46.000 Yeah. 24:46.000 --> 24:49.000 It's just like the beauty of the transform model, 24:49.000 --> 24:53.000 because it's working as a translator between human language 24:53.000 --> 24:56.000 to the speech language. 24:56.000 --> 24:59.000 So, with the spirit of open source, 24:59.000 --> 25:01.000 we met everything available, 25:01.000 --> 25:05.000 even our training code, the data set, the model ways, 25:05.000 --> 25:08.000 and even our discussions, 25:08.000 --> 25:11.000 even some fail attempts and papers. 25:11.000 --> 25:15.000 And so, if you want to see us fail in real time, 25:15.000 --> 25:18.000 you can, like, look into our discussion, 25:18.000 --> 25:20.000 and watch the live stream. 25:20.000 --> 25:22.000 I will be there, and, like, 25:22.000 --> 25:26.000 sometimes I will say something bad over there. 25:26.000 --> 25:31.000 Okay, so, okay, the last thing I want to discuss today 25:31.000 --> 25:35.000 is, like, how to bring the machine into a real world. 25:35.000 --> 25:37.000 We have talked about the live cycle, 25:37.000 --> 25:41.000 AI within a machine, but what if we could put the AI 25:41.000 --> 25:45.000 to interact with a real real environment? 25:45.000 --> 25:47.000 So, before it can do that, 25:47.000 --> 25:50.000 we need to train it on a simulated environment. 25:50.000 --> 25:53.000 So, unlike human baby, 25:53.000 --> 25:57.000 we would take, like, 10 or 12 months to start working. 25:57.000 --> 25:59.000 We can simulate thousands, 25:59.000 --> 26:02.000 or even 100,000 robots in parallel, 26:02.000 --> 26:05.000 to help them to learn the tasks, 26:05.000 --> 26:10.000 like, working in multiple environments. 26:10.000 --> 26:13.000 Okay, so here's not the machine fan 26:13.000 --> 26:15.000 from head up, and video, 26:15.000 --> 26:18.000 but the department said that the simulation 26:18.000 --> 26:20.000 has to write the robots learning curve 26:20.000 --> 26:23.000 by 10,000 times compared to real life. 26:23.000 --> 26:26.000 Yeah, it's insane, right? 26:26.000 --> 26:30.000 So, this is, hopefully, 26:30.000 --> 26:32.000 in new guys can hear it. 26:32.000 --> 26:35.000 This is a really, like, initial attempt 26:35.000 --> 26:40.000 to control a robot using an easy-go model. 26:41.000 --> 26:43.000 So, this is a voice, 26:43.000 --> 26:46.000 we will put it into the model. 26:48.000 --> 26:51.000 The place is not going to. 26:51.000 --> 26:54.000 So, let me just play the video. 26:54.000 --> 26:58.000 So, right here, we used our voice 26:58.000 --> 27:01.000 to talk to the robots. 27:01.000 --> 27:05.000 We said that it's a left 40 degree. 27:06.000 --> 27:09.000 And then we go straight ahead. 27:13.000 --> 27:15.000 I don't know if you guys can hear the voice from, 27:15.000 --> 27:18.000 but it's like, to left 40 degrees, 27:18.000 --> 27:20.000 and then we'll go straight ahead. 27:20.000 --> 27:22.000 So, in the model, 27:22.000 --> 27:26.000 the robot will execute the action. 27:26.000 --> 27:29.000 So, the behind the scene of all of those is that, 27:29.000 --> 27:32.000 the easy-go is just just like, 27:32.000 --> 27:35.000 it will use the voice, 27:35.000 --> 27:38.000 convert it into the JSON format right here. 27:38.000 --> 27:41.000 And now, which is the represent of, 27:41.000 --> 27:44.000 it's to this represent of the action. 27:44.000 --> 27:47.000 So, every that I then stream into 27:47.000 --> 27:49.000 the robot in simulated environment 27:49.000 --> 27:51.000 and the same process will apply 27:51.000 --> 27:55.000 when it moved to the real world later. 27:55.000 --> 27:57.000 So, what's next? 27:57.000 --> 28:01.000 After the simulation comes with the physical, 28:02.000 --> 28:06.000 the seem to real is the term used to just like, 28:06.000 --> 28:09.000 just for everything the robot has learned 28:09.000 --> 28:12.000 from simulated virtual, real world. 28:12.000 --> 28:14.000 And you can see here, like, 28:14.000 --> 28:16.000 at home real, 28:16.000 --> 28:19.000 we just have just bought a real robot to test it out. 28:19.000 --> 28:21.000 So, everything is still ahead. 28:21.000 --> 28:26.000 And we update our process, 28:26.000 --> 28:28.000 publicly even if I fail, 28:28.000 --> 28:32.000 so, if you guys can see it through our discord. 28:32.000 --> 28:34.000 And yeah, that's it. 28:34.000 --> 28:35.000 Another talk. 28:35.000 --> 28:37.000 Hopefully, it gives you a glimpse of 28:37.000 --> 28:39.000 the current state of AI right now. 28:39.000 --> 28:40.000 Thank you for listening. 28:41.000 --> 28:42.000 Thank you. 28:42.000 --> 28:45.000 Thank you.