Blum delivers a disciplined masterclass that prioritizes fundamental mastery over modern shortcuts, proving that manual implementation remains the best teacher for neural network architecture. The focus on multi-optimizer strategies offers a nuanced look at training dynamics often skipped in entry-level guides.
Install our extension to search inside any video instantly.
Training AI with PyTorch: Part 2Added:
Ah, here we go. Okay, this is where we left off. All right, so we are going to finish off today. I believe we have enough time that is Friday now to complete our full run through of a convolutional neural network end to end.
This is part two. We did part one yesterday and I believe we were where we at full uh CNN. Here we go. Convolution on the network. We've got most of it ready to go. We've got the full model completely rendered out here and we're [music] we we made a few changes. I know. Hey, Sigma Edits. Good to see you. Welcome on in. Happy Friday. It's the weekend. We made it. We made it to the [music] weekend. Oh, Lun. Hey, you're already here. Welcome on in Lons. I saw I saw your comments there on Discord. I saw them. I saw them. I saw the uh the one just now. Just now it was Where is it?
Thank you for [music] the heart. Let's see. New record. New. You made it here just in time. Under under a minute.
Under a minute. Where did you post that?
You posted it over here. This was pretty neat. I liked your your uh your all of your apps that you have here. The only one that I'm not familiar with is Notion. And I've heard people use Notion. I've just not ever used it before.
You was you were active today on the Discord. Yeah, you were. You're very active.
>> You had a few a few things that were active today.
>> This is neat. So, I've heard of notion.
I'm like, "Oh, wait. Some people use notion. XP farm in the XP. You got to do it. You got to do it." So, we've got our feed forward neural network here that we've completely defined successfully.
And hey, quantified quantum, good to see it. We made it to the weekend. Check this out. It's Friday. I'm excited because uh it's been busy. Let's just say it's been a little bit busy. How much how much XP? I'm actually kind of curious. I'm curious myself. We've got our Oh, wait. Are we using this selfjell? Yeah, we are. Okay. So, we've we've got a Jellu here and that's the only major difference that we are currently implementing directly. Hey, Nea. How's it going? Good to see you.
Happy Friday. It's the weekend. We made it. We made it to the weekend. Nea, this is great. We are We're doing a full run through on our CNN and I believe I believe I should probably be able to mostly get this done. The only thing that I might struggle with if I going back to the docs would be Oh, there was one. Okay, so let me This is what we want to do today. All right, so let's do some to-dos. All right, to-dos. Let's do a few toddos here. Uh boop boop boop.
All right, let's get into future dupes.
I hear you wrapped your uh you wrapped your live stream as soon as I saw Stephen was live. Oh, you really quantified quantum. Oh, [laughter] that's that's uh very very nice of you to to leave your live stream just to come over here. It's good to see you.
Hey, Peace Lord. Hi there. Welcome on in. Happy Friday. Friday is the weekend for you guys. Aren't you American?
Saturday and Sunday. Yeah. So, we do. We do. Hey, Jeff. Hey, Jeff. Good to see you, Jeff. YouTube headquarters. Welcome on in. Good to see you, Jeffs.
Yes. So, Neba, you're right. In America, Saturday and Sunday [music] is the weekend. However, we kind of count Friday as the leadin to the weekend. So, it kind of counts as the start of the weekend, right? So, it counts it counts as making it to the weekend. Just about just about cuz we're almost done. Like, the workday in America is almost done.
Like, we're more than halfway through the workday. So, we're basically there.
We're basically to the weekend, you guys. This is great. It's Saturday.
Peace, Lord. Well, here it's Yeah, I guess. So, it's already Saturday. Yeah, it would be Saturday there. Saturday early morning, right? Saturday early morning.
All right. So, I here, wait, this is what I'm going to do. I'm going to use multiple I'm going to use multiple optimizers all together. Wait, all together. This is one of my favorite words. I like this word because it's a a conjunction of multiple words all together. Three-word conjunction, and you just combined it into one. 1:17 a.m.
[music] That's pretty early. Hey, Anastasia. Good to see you. Thanks for clicking the high button. Happy Friday.
We made it to the weekend. Isn't that great? We made it to the weekend, you guys. Multiple optimizers and dreamer, too. You do not stream on Saturday.
That's right. On the weekend, on the American weekend, we don't stream on Saturday. Just Monday through Friday, you guys. Monday through Friday. Thanks again for the welcome, Jeffs. Yeah. So glad you're here. Are you excited about some PyTorch? [music] I am. I'm really excited about PyTorch.
We're we're becoming a PyTorch expert.
For the most part, I believe I have fully covered the major major points of PyTorch. All the major and important stuff. You sent a LinkedIn connection request. Oh, you did. Okay. Uh let's go take a look at it. Let's take a quick little look.
So, peace lord. What would [music] um what would your name be? [laughter] LinkedIn. All right. Well, head on over there. So, I've got my LinkedIn here.
Uh Deb. Okay. Let's go to here.
LinkedIn. Go to all the groups. I'm pulling it out. I'm pulling out. [music] One second. I've got in a window that you can't see. Uh show.
[music] Let's see. I don't see you.
Did we connect already? We might have already connected.
All right. Oh, no. I see you. I see you.
I see you. I see you. Sorry. You're just way down. You're You're way down on the list. Uh, yep. That's you. Okay. I'm going to send you a message. Send you a message, Deb. Here we go. All right.
We're connected now. Hello.
All right. Message sent. Message sent.
There you go.
All right. All right. We did it. Peace Lord. We are now connected on LinkedIn.
Okay. Got it. If you correct then USA time zone is Friday now. Yes, you are correct. It is Friday in the USA time zone. What are we making? Uh what's going on? How's it going there? Uh Chaha uh Chahhat Chihat. Good to see you.
Welcome on in. We're doing PyTorch. So we're making a convolutional neural network. This is part of our journey to becoming a pietorch expert to learn everything [music] possible in PyTorch.
I understand the autograd part of it. I understand how it represents tensors as a vector. You didn't know this whole time if you're building a multi-dimensional vector in PyTorch.
Watch this. Python we say import torch and then we say to torch to r.and do this uh one comma one comma one like this. Well, let's do 10. 10 10 10, right? Let's do 10. So, it's a three-dimensional tensor. You see there's a lot of dimensions here, right?
Quite a few dimensions in there. It's actually represented in memory as [music] a contiguous block. All of them together. It looks like [music] it's a three-dimensional array. It's really just a one-dimensional array with a view on how they treat the segments. So, we understand it all. How's it going there, Plate? Good to see you, Plate. Welcome on in. Happy Friday. Thanks for saying hi, bro. Welcome on in. Good to see you.
Thank you very much. You're welcome.
[music] You're welcome, Peace Lord.
There is one thing that you don't like with AI, and that is that it isn't a platform with all the models, including the photo and video generation. Open router doesn't have all the models, but not all, especially video. We use the We use Open Router with uh text. That's what we do. Open router with lm texts.
So we we can do that. Uh chiot crazy. Is it is it crazy? We made it to the weekend. We did. We made it.
All right. Uh let's see here. Any language uh any language doing something else is over cache locality. Oh. Oh yeah. The data shape is way the only way to optimally to the tensors and the matrices. Yes. It's really neat. It's really smart because then you can reshape a a tensor for free. It's a completely free operation. Isn't that great?
So, I'm done with time travel, a light joke, positive joke, man. Hey, dreamer.
You know it. We will talk about uh recurrent CNN and CNN and good ML stuff.
I that sounds like a really fun one. I didn't even ever possibly consider. Did you say an RCNN? What is an RCNN? And to me, that means recurrent CNN. What [music] is that? I've never seen that before. R CNN. What is this? [music] What's an RCNN? A regionbased convolutional network. Wait, a deep learning model used computer vision for object detection and localization. It goes beyond simple image classification by identifying what is in an image and where it is. Oh, is that what the R region is for? [music] So you can tell where an object is.
Regionbased convolutional neural network.
Nice. Hey, the F4 in. Good to see you.
Welcome on in. Happy Friday. We made it to the weekend. How's it going there, Droid deck? Training AI with PyTorch.
Wow, I need to learn AI. There we go.
That's what we're doing today. But you need a freelancer or are you in job?
Just curious. I have my job. I've got a job. Let me see. Yeah. Where are we at [music] here? I'll show you here. Here's me, Steven Blum, CTO at PubNub. Got my own company. Raised uh some money [music] series E. 1 billion devices on my network. We're busy. We got them all.
And a lot of patents. We got a lot of patents, you guys. Uh how hot temp is it? Uh it is like 70°. Well, it's it's like 60. It's in the 60s.
It's uh in the 60s. Let's see. It's done this way. Not reshape matrix multiplication can get all the things in memory without capturing a bunch of random heap memory. It's really clever.
When I realized what they were doing, I'm like, "Wait a second. Is this what they're doing? Is it? Is it?" And it turns out it is. That is exactly how PyTorch works.
Yeah. So, uh, played Bro, uh, if you're still here, let me know if you have any questions.
Yeah. I was on YouTube and saw live programmed and we were [music] like, isn't this like right now? Oh, and it sure is. It's exactly right now. It is exact moment. [music] Recurrent convolutional neural network.
See, that's what I thought you meant is recurrent.
The flashers you added in a few days ago, that helped me a lot. Wait, the Oh, the flushers. What are the flushers?
Wait, I don't remember this. What are the flushers?
Question mark. I Anosi, tell me about that. I I don't remember. This isn't an idea from PyTorch per se. It's actually just how CUDA works. Oh, so really. So they borrowed from CUDA and then they they put it onto all the other accelerators. Oh, for garbage data. Oh, okay. I I vaguely remember. Thanks for the reactions, you guys. Good to have you here. appreciate you clicking those buttons. Thank you so much.
I didn't realize that, Nova. I didn't know, but now I do. Now I do. All right, I want to try multiple optimizers. Uh, we need to finish [music] the training with epoch.
So, we're going to I'm going to try to do it without going and then we will finish the [music] tutorial which includes tensorboard. And I don't really want to use tensor board.
Includes tensor board. I don't really I don't really want to, but I feel like it's necessary.
Blue J. Well, played. What do you mean?
C arrays work this way as well. Oh, really? I didn't even think of that. I didn't even consider that. I always think of arrays or anything in C doing what I want them to do. I didn't ever consider contiguous.
You're just a dentist. Wait, you're a dentist and not just a dentists are you dentist? You have to go through a lot of work to get become a dentist working as a senior consultant [music] and you have only learned basic DOSs and Java. Oh well that's more than most dentists learn if you have any programming background at all.
That's awesome. Hey, plate. That's a really great thing. Other accelerators do this as well. Core idea is old as forran. Oh, well, I understood the concept. I get it and I feel like it makes a lot of sense and it doesn't feel like novel or new. I just was really excited [music] that that's how they did that because I'm like, "Oh, that's really clever. They should do it like that." And apparently, Neba, you're like, "Oh, that's old school. That's old. You need a dentist. Jeff needs a dentist. Oh, well, you found a dentist here. Plays a dentist.
Healthcare.
Uh, mind pulling mine out. Oh. Oh, yeah.
That sounds like a problem. What are you writing? How's it going there, Lil Craig? Hey, L Craig. Welcome on in.
Happy Friday. We're building a convolutional neural network. So, we're going to finish the model. We're going to try multiple optimizers together [music] at the same time, you guys. I'm really excited about that. We're going to do more than one. I'm going to do like a random I'm [music] going to initialize like an array of them and we're going to see if it does anything.
I'm curious to see what it does. So, this is going to be our experiment today. [music] Multiple optimizers all together. Have you ever done that before, you guys?
Have you ever done that? You're you're a DevOps engineer. Oh, Droid. Yes.
However, you started as a Windows and VMware engineer. Oh, did really? Wait, VMware engineer. So, you've [music] done some virtualization. A little bit of virtualization there. Virtual machine. I remember VMware. VMware is Oracle, right? You got some Oracle in there. All right. If I miss any of your messages, you guys just let me know and I can jump back through them. Should use MLflow to log your model and track them. Oh, peace lord. MLflow. Yeah. So, I've been using mostly just my own my own stuff. Just writing everything from scratch when it comes to tracking.
What kind of network? Uh C++ Windows internal dev. Oh, L Craig. [music] Uh who? Oh, it's a it's a it's it's an AI model. That's what [music] we're building. We're building an AI model.
Android Studio feature is not finished.
You can use a to-do function and Android Studio will add a thing in the list in the to-do. It's really neat. Oh, that's great. Yeah. I just I just do a search on it. I just do a search. I'm like, where's all my to-dos at? Is it around here somewhere?
Yeah, AI. That sounds good. L Craig Citrix. Ah, right. Good old Citrix. It's uh [music] No one. No, no, no, no, no.
VMware is Broadcom, not Oracle. Oh, Virtual Box is Oracle. Thank you for the correction, loons. Thank you.
Windows in [music] front Windows internals is fun to to learn. I was going to say, wait, are you sure that's fun? Is it really? Thank you for the reactions you guys. I'm seeing those hundreds pull up there. I can see what you're doing over here. [music] Thank you so much. Appreciate it.
Here for you. Uh [music] I want to do voice cloning. Which model should I use?
Oh, there's some pretty good ones. Go to hugging face. [music] Go to models.
Let's see. You're going to go Let me zoom out here. And then go to I think [music] is it speech to speech?
Is there is it can we do that? Is it speech? Let's do like speech uh automatic speak speech audio text to audio text to speech. Yeah, maybe audio to audio, right? You can do some audio to audio. There's some good ones in here. Yeah. So, just go over to Hugging Face and go audio to audio. That might work out well for you. There's also some speech models in there. You might want to check that out.
Uh my friend Windows internals. Yeah, there we go. It aids. Uh, it's easy once you get past the learning curve. Ah, hey Godo. Hey, how's it going there? Hey, Stephen. Long time. Go. Frugal. Uh, frogpole.
Uh, pole. Frogpole. Goto. Go to frogpole. Hey, there we go. Good to see you. Welcome on in.
[music] Happy Friday. We made it to the weekend. Is the only way Windows makes a sense. Oh, is it?
[music] Nice. Nice, you guys. All right, let's jump into our our code here. So, I think this is a good plan for today.
Multiple optimizers, which is going to be the fun part. We're going to try that out. Then we're going to finish that model, which I think is what we should do right now. We should start typing that all out that right now. So, let's do a defaf train and we will have an epoch. [music] Where's our Where's our data? All right, so let's do this. Um, epo equals we'll do one for now. What else do we need? We've got our learning rates. Good. I got the optimizer.
[music] We get our loss function. I think we're fine there. Looks good to me.
Is your best friend wait [music] and not what is that? What? What is How do you say that? Lol. Craig, what is that word you're saying there? Can someone give you a road map where I [music] can work on uh AI ops? Oh, MLOps.
Transitioning into this is pretty Yep.
So you probably want to look into vector databases and then uh getting getting data into the model and then what [music] are kind of other workflows is there are there are workflow I don't know I don't know there's a whole there's a whole industry around AI ops mostly it's going to be about getting data in and out of the model so it can train and learn I think that's the big one and then operating and inference those are separate things as well can everyone drop a like for Stephen hey Jeff, thank you for saying that. That was really nice. I appreciate it.
So, what are we trying to do? Oh, we are going to So, here's our here's our tasks for today. We got our tasks right here on the screen. Thumbs up for Steve.
Thank you. Appreciate it. It's the main kernel called dispatcher in Windows. Oh.
Oh, really?
Intos kernel. Okay. I didn't know about that. Moit, how's it going, Moit? Yo, welcome on in. Happy Friday. Hey, we made it to the weekend. Good to see you.
Wish you could share your learning deck.
[music] Uh, it's it's pretty good. It's good one. Well, if you want, paste it. Paste it into Thank you for the hundreds, you guys. I see what you're doing over there. Paste it into Discord. Hey, Zergio, welcome on in. Good to see you.
Also, nothing. How's it going over there? Nothing. Thanks for clicking the hide button. Steven, do you know languages like Spanish or Hebrew? Uh, no. I know a smidgen. Uh, un poito of Spanish. Other than that, that's as far as I go. Like, um, but that's it. I could I a little bit, but not much. Not much.
Llama, C++, and uh, Onyx and TensorFlow would be a big deal if you're running locally. Nice. Yes, I forgot about Onyx.
Thank for the hearts, you [music] guys.
Hey, JM, good to see you. What's up, everyone? Welcome on in. Good to see you. Happy Friday, Steve. your stream was very interesting and informative.
Peace, Lord. Thank you. Yeah. So, this is Hey, Kyle. Good to see you. Welcome on in. Happy Friday. We made it to the weekend. Which model are you training?
This is a simple, really simple model.
So, it's a [music] a very easy convolution model. So, you can see here what we're doing. Easy. Very easy. All It's all right there. All right.
Highlighted on the screen. That's what we're doing. Thank you for the hearts, you guys. Appreciate it.
All right. Let's get back into our coding here.
How's it going there, Ally? Welcome on in. Good to see you. Happy we made it to the weekend, Ally.
Where's the Windows engineers? Oh, yeah.
I like my car.net. That's what I like.
PC having uh some challenges. Want to connect with internet? You're on the internet now. It seems to be working.
Thank you for all the reactions, you guys. Appreciate it. Good to have you here. By the way, I speak Hebrew if that's relevant. It is relevant. It's exactly what was asked. It was exactly asked.
Quantified quantum was curious about this.
Oh, fun fact. [music] What's that? Russians are getting in the CPU market and China is [music] becoming a fourth contender that is certified as a GPU vendor. Oh, hey. Yeah, it's the future, right?
Compute. It's all about compute, you guys. So, we need to we need to run through our iterator here. So, I'm going to enumerate. [music] Let's do this. Uh, enumerate on our iterator. [music] We need a new iterator. Where are we at here? There we go. Training data. I think we'll be able to do this without [music] referring to any of our crutches. See if we can do this. So, we [music] need our epoch equals no no batch batch, uh, features, comma, labels.
we go data iter right there. Okay, so let's do a print batch and I just want to make sure this works and then we're going to we're going to return immediately. Uh we need to put this in a for loop though. Uh for in like that we'll print a batch and [music] we'll return immediately.
Okay, let's make sure this still works you guys. Python full dash CNN. There we go. Uh Steve, a just a suggestion. Stream in 16 by9. Hey, I got good news for you. Look it over here. We are in fact streaming in [music] 16x9 ratio if you're curious about it. Uh Twitch/stephvenel LB. Here you go. LB right [music] there.
If you want. Right there. If you want there.
Some older AI algorithms would [music] work well. Hey, it's Lun.
Hello.
Hello. They have CPU, but it's slower.
Yeah, they're going to have to catch up, aren't they?
>> Yeah, they're going to have to catch up.
Okay, bring this back over here. See?
Bring our terminal window back up.
Uh, showing 96 in YouTube. Yeah, exactly. And on Twitch, it's the opposite. It's the 16 by9. You got it.
Peace, Lord. Glad it was an easy fix. It keeping it easy, you guys. Keeping it easy. Hey, Mark Lemon. Good to see you.
I saw you saying hi there. Welcome on in. Happy Friday. We made it to the weekend.
See here, LA. Trying to catch up with the chats here. Yo, how's it going there, DRXGY?
Dr. X, how do you say your name? Dr. Dr. Xgy, good to see you. It's good to see you again. Welcome on back, Stephen. Are you doing an echo function in your open tweet? Yep, on purpose.
That's by design.
Yo, bro, the last attack. [music] Good to see it. Your client asked you to free up the licenses and did a per and you [music] did that as per later later uh later of client got to know that they have sent the wrong file and now reassigning [music] manual licenses for each user. Oh no, the last attack. Isn't that rough [music] when that happens?
Yo, what's it going there? Oh, droid.
Yeah, droid deck. Thanks for saying yo.
It's like D uh druggy. Oh, got it. Okay.
Got it. Now I got you. I got you.
Happy Friday, Mark. Good to see you. All right, let's see if we can make this work. So, did that work? It sure did.
Look at that right there. Success. Okay, so that's our batch. Did it print the batch? No, we just didn't run that yet.
All right. So, let's do train. Run that.
See, model train. Can we do this? Is this something we can do? Can we do Can we do torch? Is it torch train? I forgot. How do we How do we do a decorator?
See if we can make this work. Didn't like that. All right. Model.
We do that. Does that [music] work? No.
We need to do it here. Like that. I don't know if you have to do that. I don't think you do, but we're going to do it anyway. Perfect. Batch zero.
Perfect. Do you code without AR or do you use it sometimes? Hey, how's it going there? Uh, Wisdom Nuggets. Hey, Wisdom Nuggets. Good to see you again.
Happy Friday. We made it to the weekend, you guys. We made it to the weekend. On stream, while we stream, I code by hand.
However, the rest of the day in my day job, I I use the AI. I use the AI to do everything. You're on TV, loons. You are on TV.
Yeah. Nice.
Uh, so yeah, welcome to the stream.
So yeah, I use AI all the time. However, I want to keep my skills fresh. So we're going to [music] be coding while on stream by hand. All right, so this going to get our data. So we need to run it through our model. So our features are [music] out equals model dot. Did we move this over to Let's see. We didn't. So, let's do [music] device. And do we have our device defined? We do. Let's grab this device here real quick. Just a smidgen real quick. And then we'll bring it down here.
Device.
Best thing about hybrid coder is they can do manual code and AI code. Yep, it is. You can do both. You get to do both of them. It's great. So, I'm thinking about moving my camera uh cuz currently it's like above my monitor. I'm thinking about moving it over here or moving it over here. I I'm trying to [music] figure it out because I've got these Kanto speakers that I never turn on. I never turn them on. And if I move them, then I can move my camera in a place and I think it'll be more interesting. So, I think I'm going to give that a try here.
I I not maybe ne over the weekend. I'll do it over the weekend. It's a big project.
Same. You do loons, you use AI a lot, but still using tutorials [music] with Cotlin. Yes. If you want to learn something, yes. Good way to learn is without AI. You're doing it right, Lun.
You're doing it right.
The right way, you guys. The right way.
All right. I'm really excited about this. Okay. So, do we have what is it?
Out print shape. That's fine. Yeah.
[music] Right over there. We'll leave that there. We'll leave it there. That's fine. So, this is just a little bit of debugging here. Here, let me see. Um, this is testing.
Little a little bit of testing here, you guys. And then we'll run this function here. Uh, test. Although, I don't think we need to. We're just going to leave it there anyway.
Keep your top right of left, please. Not your [music] bottom side or your center.
What does that mean?
Wait, wait, wait. You like So, you're saying that this is the best angle right here already?
How much time are you working on [music] ML? How much experience? Oh, um, Aditia, we've been working on it for since 2014.
2014. So that is two 2014, right? About there. Yeah. So that's about 12 years.
Yeah. Oh, well 11 years. Let's give say 11 years. And this was before PyTorch was really prominent, before TensorFlow was prominent. So, a lot of the work that I did was all from scratch, you guys. All from scratch.
Stephen, that's some good name pronunciation. Is it [laughter] Are you Are you being Are you being sar is that is that not sarcasm? I can't tell.
Uh, Adita, is that right? Am I trying?
Okay, let's keep that running. Model dot. Where is that at? Where did we put that at? That is Oh, we didn't finish this here. Okay, this will be model dot uh features to device. There we go. And then we need to print out size.
See how one second up the chain, please.
Thank you. Shape. I'm going to do shape.
Shape.
There we go. There we go. Uh, Nit uh, Niti, wait. Niti, how's it going there, Nitty? Thank you for laughing. I'm not sure what you're laughing at, but I'm glad you're here. All right, XP.
Okay, so let's make sure it still runs.
Looks good. A, just kidding.
Must be on the same device. That's what we did. We did that right here. We did it on the same device. Wait, is this not Is this It doesn't model.
Let's see. See that? Should Shouldn't that work? Shouldn't that be perfectly fine? I think that should be perfectly fine right there. Right.
Model [music] equals model.2 device.
Let's try that. All right. Is that better now? No. We'll figure it [music] out. Yes. Don't worry. This one's easy.
It wouldn't it be funny if we were stuck on this one problem for the [music] entire stream? [laughter] That would be a little bit unexpected.
What's your view on Meta efforts on passing ID verification laws? Well, it's not Meta's problem, right? That's not Meta's problem. Hey, how's it going there, Sudio? Is that Sudio or Sud Sudlo?
Well, come on in. Thank you for subscribing. Oh, there we go. Sud Sudlo, let me see. Yeah, Sudlo, thank you for subscribing. Good to have you here. You joined the right channel for software engineering where we're going to write all of our code by hand. We're becoming PyTorch experts right now. Lions, how's it going? Lions for Labs. What are we doing? Good question. [music] Today, we're creating a convolutional neural network, which we've been doing for a while. However, [music] we're going to we're taking it up to the next level.
For the most part, we're not going to refer to any tutorials or any AI. We're going to try to finish it ourselves.
That's the plan for today. I think I can finish it. No help. Nothing up the sleeves. You can't see anything up the sleeves, you guys. Nothing's there.
Nothing's there.
Sudl low is your nickname, Stephen. Oh, Droid. Oh, I didn't. Thank you for subscribing. Droid deck. I didn't know that was you.
Thank you so much. Meta is doing the laundering BTS.
What? No. What do you mean? Uh, what do you actually quify quantum? You got to tell us more details. Slio equals Droid.
It makes sense now. It makes sense.
What's your opinion on Alexander Wang?
Who is that? Let's find out. We're going to go look this up real quick. Alexander Wang, Bloomberg AI scale. Oh, uh, joined Facebook, right? Uh, American entrepreneur, chief AI at Meta Platform, uh, leading super intelligence life.
That's that's the only thing I know. I can't give you an opinion [music] otherwise. Hey, how's it going there, Shaya? Thank you for saying, oh, we don't have a we don't have a chatbot. So if you're right Yeah, I know. I don't know who this is. All I can tell you is this is a person and they did this thing and they had a startup I believe got acquired or left and then joined Facebook. [music] Uh became billionaire co-founding Scale AI provides data labeling. Oh, I remember scale AI. I remember that from many years ago. That was like a long time ago. Scale AI was like a way to label. It was a bunch of labeling.
It was all the data. It was like the source of the data. Ways to train AI.
Meta is funding the politicians to pass the ID verification laws that the world is going to uh the word in the tech community. I didn't know they were funding for it. [music] Isn't that kind of like the opposite of what they'd want? Do they really want that? It doesn't make any sense. Why would Facebook want to hinder users joining their platform?
That seems like the opposite of what they'd want. Hey, how's it going there, Tutor Royal? Good to see you. Hi. Happy to see you. Hey, welcome on in. Happy Friday. We made it to the weekend, you guys. We made it to the weekend.
Yeah, exactly. 2016. Meta. Wow. I know, right? Isn't that crazy? Meta's doing some stuff over there. [music] All right, let's get back to our code.
So, we're going to finish our training.
So, let's do this real quick. I think we can make this happen. So, we're going to get it out. We're going to say, "Oh, wait, wait, wait, wait. Um, uh, how do we say zero our [music] gradients? I forgot how to do that. See, uh, zero grad. Zero [music] grad model dot. Is that right? Do we do that there? Let's double check. [music] Hold on one second before we do that.
Oh, wait. What? No way. Hold on a second. Oh, it's happening [music] here.
There we go. Now we figured that out.
Yay. Much better. Now we get a different error. Uh, placeholder storage has not been allocated on NPS.
Why not? It's what? Really?
How about this? Let's hide that for a second.
Facebook wants to be able to force their users to give them government ID so they can use that as data for nefarious purposes. [laughter] Never. All right. Now it makes sense.
Thank you for describing it. Understood.
You've made it make sense. I appreciate it. You love seeing other people doing what they do to love. Yeah, that's what we're doing today. We're doing what I love. Started learning uh machine learning 6 months ago. You hope [music] you'll be as good as you in the future.
Oh, you look young. Well, thank you so much. I'm We're actually old. [music] I'm actually old over here, you guys. In my 40s. Definitely old now. And [music] you guys have to take over. You got to take over the world. Uh because it's your job now. And I'm going to retire.
You guys deal with it. It's your problem now. And I'm just going to sit in my couch.
You think anthropic miso is way ahead?
It's they they say it is. They say it's way ahead. Hey, Torva, welcome on in.
Good to see you, Torva. Happy Friday. We made it to the weekend. Steven is awesome. Well, Droid, thank you for saying that. That makes me feel so good.
You make me feel happy, you guys. I'm younger than Steven. And I'm quite good.
Yes, Nea is the expert. Nea's the expert here. In like 3 years, could be very competent.
Very competent. Very, very.
All right. Uh la. Did I mean OCT? No, I meant out. Name out is not defined. Why not? Why not? Oh, there we go. There we go. Okay. Uh, perfect. Okay. So, let's try this. Is that going to work now?
Yeah. There we go. There we go. All right. Now, let's fix our [music] test.
Let's make sure our test is working here. Um, model images [music] two devices should work. I don't see why it wouldn't be. I don't see why it would be. Oh, it's right here is the problem.
Okay. It's It's uh line 77. That's the problem. Okay. So, this right here. All right. So, I know what to do. I know what the problem is.
Two device.
Perfect. Perfect. We're making our way, you guys. We're making a way. It depends on what you consider a good AI engineer, but [music] yeah. Pie torch. Mhm. There you go. then implementing some papers and you can do that in three years. Nea is very accomplished when it comes to AI and NeA has been helping us a lot when we're learning. I feel like now I've got like a really good grasp of PyTorch.
Like I'm very comfortable and confident in it. Just the the basics. When it comes to things like TensorBoard, I get it a little bit though I don't know if I want to use it or not. Banks are notified to take care of security because of mythos. All right. So, it's very potential that there is something there with Nissos. I do think it is overhyped. I think it's overhyped just a smidge.
You should be back in 20 minutes. All right. Sounds good. Please. Deploying Anible. Oh, hey. Are you doing that from your laptop or are you doing it from, you know, a dedicated server and you're shelling or you're sshing into the server? That's what we used to do. We used to have we used to use Anible a lot at PubNub at my company and we would deploy from our laptops which you know can be a problem because if your internet goes out halfway through the deploy well the deploy never finishes.
So what we did was we spawn up dedicated servers in each of our data centers [music] that are focused for deployment. They're the deploy servers and we would manually trigger them we when we're ready to go to production. This is before you know CD. This is before we had CD. Uh, now we use CD. Uh, continuous deployment. Hey Jensen, good to see you. Welcome on in.
Happy Friday. We made it to the weekend.
It's the weekend, you guys. The core thing is reading papers so you can actually implement useful useful commands. Keeping up with the research is a bit of a chore. Like it it can Yeah, exactly. Cuz there's new papers all the time. Using on your AWS. Nice.
Use GitHub actions. Excellent. Yeah, we like that too. GitHub actions is great.
I know there is some problems that people complain about GitHub a lot. I've been we've been doing fine. I like it's down. GitHub is down. Like is it? It seems fine. Like maybe once in the last 2 years has it actually hindered me a smidgen. But it was fine because I just waited an hour and they brought it back up. It wasn't a big deal.
Why why is there all this like hate and shade going towards GitHub? It doesn't make sense to me. Uh because it's been fine for me. Although Yeah. But yeah, for beginners reading attention is all you need is at least a five minute five times. It required five [music] times read through. Yeah. So attention's all you need. The parts that didn't make sense to me originally were the QKV, the inputs, and the layers which were filters similar to convolutional neural networks. So, you've got the QKV. You would throw your your data in through uh your inputs and how that would work. Oh, Aditia, hey, thank you for subscribing.
I thought you were already a subscriber.
Thank you for subscribing cuz you've been here for so many times. Thank you for joining in. You joined the right channel. We're going to be doing a lot of AI, you guys. A lot of AI. I'm really excited. Yeah, that QKV is messy. And then also the what do you call it? Uh positional encoding, right? There was using cosine and sign for positional encoding which I implemented before by hand from scratch. I thought that part was pretty straightforward though I didn't throw enough compute hours to see a difference. Maybe the BERT or the GPT papers are easier [music] reads. The models are simpler. I don't know. Maybe we should be doing some GPT and BERT reads. That would actually be a lot of fun for me. I don't know how fun it would be for stream if we sat through an archive paper and we started reading some [music] burnt GPT and Transformers.
Would that be fun, you guys? Would you want to do that? Why train and not plane? Uh [laughter] yeah, I see what you said there. Uh Tread or uh T Brimid. Uh Breard T Brimard.
Brimard. I see what I see what you're doing there. I think as you're joking around, I think it's a joke. Are you joking, Stephen? I'm wondering how countries like India would be affected by voice AI agents because major workforce there work as agents.
Yes, they do. They sure do. It's coming for you. Careful. The even cheaper labor force is on its way. The AI agents, you guys, they're going to take all your jobs.
Aha. I thought so. I thought so, Burmered. I thought so. In a few words, can you guide me from the first interview as an AI engineer with what things I should be ready with? Already listening to many, the most hardcore ML learner I've ever met.
Um, they're probably going to confirm that you understand the hardest part of machine learning and [music] the back propagation and what bias is for.
uh what what the bias is a vector and [music] how I I don't know like I don't think they're going to go all the way Andre Carpathy on you where they're going to ask you to build a full system using an AI uh code generator maybe that was just talked about [music] yesterday so I don't think it would be too soon for something like that take mine I'd be happy about it yeah get the jobs the AI agents are coming for your jobs. Uh, so Aditia, yeah, the way AI works so far is that some guy in India was pretending to be AI for anything that we couldn't. [laughter] They would potentially gain jobs in pretending to be AI.
Like an American accent filter.
[laughter] Hey, Bobo Bear, you made it to the live.
Good to see you. You waited 67 hours and isn't a 67 joke. It's just that how long it took to collect the data from a polygon.
Was it? No way. You're continuing to learn with you. Oh, sounds good. Let's learn together. Happy Friday, 1st of May. That's right. It's May. We made it.
April showers bring May flowers. I see the May flowers. They're out. And they're really pretty. They're really pretty, you guys. Okay, so let's get back to our model here. Let's train it.
I think we can run this and we should see a bunch of batches show up on the screen. Looks good. Okay, looks good to me.
And then can we do try?
Let's see. Except and then just return. Wait, there's no returning. We'll just say pass. Except as [music] exception.
Can we do that? Can we do that?
E. Oh, name E is not defined. Yeah, that's why that's why I put it there.
How do we do this? I forgot. Oh, no. Hm.
I always So in Python 3, I I remember Python 2, you could just do this in Python 2. It was no problem. Can we do that in Python 2?
Yes. Okay, good. All right, perfect.
We do this so that way I can exit without all that stuff showing up on the screen. Your job specifically, who is going to run those GitHub actions? Hey, the AI's going to do it. They might do it. What is this bank holiday even for?
Hey, that's a good point. It is a holiday for many countries today.
Stephen, you're watching your machine learning videos. Oh, nice. That's great to hear. I hope they're they're informative. Uh, [music] we've we've done quite a few.
Wait, he's making long forms now. Yes, we do have some long forms. So, Steven, we go over here, go to videos. Get a lot of them. Although, they're basically just edits of the live stream. So these are all just edits of the live stream.
There we go. You can see those edits of the live stream.
It's 2:00 a.m. here India University lecture in the morning. Oh, Aitia. Oh, you'll definitely attend your streams.
All right. Sounds good. Uh enjoy your your lectures in the morning. Good to see you.
So it's like your shorts. Yes. Exactly.
Exactly. They're like they're the full forms of the shorts.
Thank for the heart. Thank you for your heart. I saw it. Let's see. Your stream is rather limited. Can only run two to three billion parameter models. Anyone have suggestions? Uh your system. So, I mean, you could just rent a server for like a few dollars for a couple hours, right? So, like rent a GPU for a few dollars on a cloud. So that way you don't have to buy the whole GPU, right? That's what I do. Oh, Google Collab. That's a good idea, Nea. Good idea.
Google Collab. Nice.
Good idea. It's free. Powerful GPU.
There you go.
I had read a lot of architecture. I am from Bangalore, India. All right, we're learning where we at. Yes, but I'm over here in United States, Washington, Seattle. Thank you for the hearts, you guys. Let's get back to our code. Use it because you're uh convenient and it's about equivalent to your 4090. Oh, nice.
Google Collab. What is Google Collab? It is like a PyTorch notebook. No, I mean um a Jupyter notebook. It's like a Jupyter notebook, right? And allows you to write Python in a web browser and then you can click the run button and that will run on a GPU. You can train your own AI on Google Collab. Yes.
Let's go. I know, right? Isn't it good?
I think it's great.
Google Collab has the option to choose GPU instance if you [music] asked for it. That's pretty powerful. All right.
So, I think we model zero grad output.
So, loss equals do we have a loss function? We do. We're going to say our out, labels.
There we go. And let's [music] do print loss. And it should stay the same. We don't need shape anymore. We don't need batch. Let's print out the loss. And good to go. How's it going there, Nish?
Welcome on in. Happy Friday. And it's just going to flicker around there. Oh, we got to do a item there. We got to do the item. Let's see. Lost item. There we go. Copy that out. So, we get just the data itself. And it's just going to flick around because I'm not training it yet. Good to see you and Oddish. Welcome on in. You need to actively choose a GPU. And they also have TPUs. Yeah, TPUs. Yeah, there we go. [music] Hardware acceleration. TensorFlow or Jax works either way. Good to have you guys here.
My knowledge is terrible. You never heard of the Jupyter Notebook. It is just a very simple way to run [music] code in the browser. And it's kind of nice cuz it's got blocks. You can create blocks of code. So you can run them in sequences and you can drag and drop the blocks so they can go above or below.
And they hold state. So you can run a few blocks above you and then run another block right below it and it it remembers.
Jupyter notebook is basically code cells. Nice. You can run individually or all of them all at once. Yeah. Stephen, can you open up a Jupyter to show them?
Yeah. [music] I'm going to do a screenshot of it. So, uh, see Jupiter with a Oh, Jupyter notebook. Here we go.
Boom. We'll do an image. There we go.
Uh, not that one. [music] That's not good. Let's see here. There's got to be one here. Something that shows the cells here. I This one's pretty good. Let's zoom in on this. Okay. So, see, you get a beautiful panel like this. This is what Google Collab is probably also like. And you can write your code in this cell and then you can just run the cell directly. So you can run just this cell [music] and then you've got another cell here, right? You've got another cell here, here, and here. And you can get the output from each of those cells each time. And you can run all of them together at the same time by running [music] the run all which is like right here. Can do the run all right there.
Pretty neat. It's pretty neat.
So, it's a better version of your terminal. Yes, it is. You can see it like that. Although I don't count that it's a better version. It technically is better in some ways. For me, I don't want to leave my terminal. I'm [music] so comfortable here. It's like a cozy blanket.
Uh, no. You have like a Python on your PC and you could just run it a bit. Oh, I don't have it cuz I don't I don't want it. [laughter] I don't want it. I just showed a picture though. Uh I I like my terminal window.
I like it better because then I can do all sorts of things. I like I've got all my keystrokes and I like it. Jupiter shows on Git [music] that has the main issue. Oh, I forgot about that. Yeah, there's I think there's also a way to show notebooks, right? Pi notebooks on GitHub.
Oh, that's some good music. Is it wait like right now? This is called Local Elevator by Kevin Mlet Leode.
Hello, Stephen. How are you, Bruss? Good to see you. Welcome on in. Happy Friday.
Good to have you here. We made it to the weekend.
Meet you on Monday. Sounds good. Claude, can I post a link here? Droid, yes. Uh, you have to post a link on our Discord cuz, uh, Google won't let you post a link. So, there's a link share here.
Just in case I'll copy this for you.
See? Get to edit this button here.
Unlimited generate. Copy. You say at Droid.
There we go. So there. And then you post there. And then we can look at the link there.
As of now, you can't see the get diff for Jupyter notebooks. Oh yeah, because it's just going to render the notebook.
It's not going to render the diff. Yeah, I think it would render the diff in the PR though, right? In the poll request.
I think it might or like the last time I checked. Anyway, I think if you're going to get a diff, it will show up in the poll request.
I've only used your terminal and never experienced anything else even after hundreds of hours. You're don't you're in the right spot in my opinion. Let's just stick with the terminal, you guys.
I'm a terminal terminal for life. when notebooks came out for the first time like Jupyter notebook when it first came out I'm like and I saw one of my my my employees and they're like I'm like what are you doing what is this what are you doing here why do you have this web browser up like I was like I was questioning their their their sanity I'm like what ARE YOU DOING WHY DO YOU have this notebook looking thing up on the screen it was like a Jupyter notebook he's like I really like it I like it I can press buttons on it and it's like really straightforward [laughter] like What are you doing? I don't trust you anymore. Uh oh, Mr. Dungwall. All right, sounds good. Enjoy your sleep.
Good to see you. Did you put on the fire on the cabin, boys? Uh, wait, what?
Scared of werewolves. What are you saying over there?
We need to add stream chat modules to Discord. Hey, that would be great. Hey, it's my favorite song.
All right. All right, you guys. Let's get back to our coding. It's my favorite song. It works. Wait, Lun, it works.
What works?
Which one?
All right, you guys. Let's Let's have the best Friday ever. All right, let's get this going here. So, zero grad loss.
Print in loss. Now, let's do a lossbackward backward and then make sure it still works. And then we'll do a See, see how fast I can be here. It's just so satisfying. Hey, Vim, Zergio, you got the Vim over there. The bot. The bot works. Really? Wait, how did you do that? How did you do that, Zergio? Thank you for saying the Vim there. So, I think it's underscore Vim.
Is that how it works? Yeah, Stephen, you just need Vim Jupiter. Oh, is that a thing? Are you Is it? Is it? I don't know. Can Can it do Can it [music] do multiple Can it like split my screen the way I want? Can it do these things? Like there's just so many cool things that I want, right? I like how I can take care of my screen and do it like this. Sorry, I'm being um I know you're Don't worry, Neva. You don't have to defend Jupyter notebooks anymore. I know they're actually pretty good and I do use one of them because it's how we deploy some of our code in a very specific scenario [music] for some of our reporting. So, I definitely use it and I complain while I'm doing it. I complain. I'm like, h I have to deal with this. Can I put this in my terminal somehow? But I just [music] deal with it. I just deal with it.
So, the bot is yours, Lons. Bot. This is great. More viewers should see Steven stream. Everyone like the stream. Peace, Lord. Thank you.
Thank you.
All right. Appreciate it. Oh, what are you using for the live annotations? So, do you mean these uh subtitles here? See these subtitles right here? So, this is software I wrote. I like how you're asking that [music] cuz we wrote this software many years ago. Eight years now. See some eight years here. Eight years ago right there. There you go.
Subtitles right there. You can add it to your stream. No problem.
Anything you want. Welcome to my stream.
you get those Twitch subtitles or YouTube as well. Hey, Shadow of Obscura, welcome on in. Thank you for clicking the hello button. Good to have you here.
Happy Friday. We made it to the weekend, you guys. What are you uh What does the message auto translate? Does it It could if we We could make it do that. Yes, it does. It does.
[music] It does auto translate.
Here, I'll show you. Are you ready? It's pretty neat. It's really neat. Check this out. [music] So this is you Bobo Bear right here and you say right uh Donbot estre and then it says so the bot is yours. See right there I don't need your fancy language cuz the AI takes care of it for me. Isn't that neat? It auto translates.
It's pretty great. Impressive. Ah isn't it? It's pretty good. It works really well. It's like fast, right? It's fast.
And it also uses my technology, PubNub.
Uh, let's see here. PubNub. Where are we at? Here. 0. Uh. Oh, wait. What? Okay.
Uh, it allows us. It It's a communication technology. [music] So, it takes data from a voice capture and it puts it inside OBS. So, you can see the screen here. Give me the GitHub link.
Absolutely.
Here you go. Copy. Paste. Uh, Peace Lord, are you in Discord?
put it in the link share here. Hey, nice for people who are preparing for interviews. Hey, that's awesome. [music] Very nice.
Yeah, here. So, it's over here in Discord.
Uh, GN Quantum, your presence will always remain. Oh, all right. It's time for you a banish. All right, Quantum.
Good to see you. Thank you, Stephen.
You're welcome. Good to have you here.
Aiden Harris, what's going on? Good to see you, Aiden. Welcome on in. Happy to have you back.
See here. LA.
Okay, Quantum, thank you. It was so good to see you. Enjoy.
Clawed tube. Are you broken? We're going to [music] find out. We're going to find out.
Okay, let's bring back into our ter. All right, let's go, you guys. Let's go.
Peace. sword. Check Discord in description. Oh yeah, yeah. Oh, right, right, right. Uh, so that also works.
Yes. Lions for lambs. Thank you. We'll also copy this for [music] you at peace.
So that will get you in so that way it's really easy for everyone to share links there. BRB guys can't concentrate on this. All right, please. Sounds good.
Sounds good. You're welcome. You're welcome, Peace Lord. All right, let's get back to our code here.
All right, let's [music] get some code here. All right, where are we at here?
Yes. Or on the profile on the link tree.
You got it. Hey, Claude Tube's still working. But I'm Claude made by Antropic, not Claude Tube 404.
Happy to help though. What do you need?
Hey, that's really neat. That's really neat.
That's awesome. Not broken. It works.
Nice loons. That is pretty satisfying.
Okay. All right. So lossbackwards and then we [music] need to optimize. So we need to use our optimizer optimum.step.
Then we'll print out the loss. I think that's all we need. I think that's it.
Let's see if it works. I hope it works.
All right. Are we going to get a loss reduction or are we not? It doesn't look like it's reducing. Uh it looks like it's staying the same. Uh-oh. Uh-oh. We messed a step. We messed it up somehow.
See here. Uh let's do this instead.
Let's go if batch mod 100. Oh no, it is reducing. Oh, we did it. Oh, nice. Okay, see we have some memorization that has occurred. Look at that. Oh, we're looking good now. It's just that there was a lot of data to learn. All right, what is our [music] batch size? We set it to four. Where's that? Batch size four. We did sweet four. This is great. Mustard seed. Get those mustard seeds in there. Are I Why mustard seed?
Am I missing something? Tell me about this. What do you mean? Lions for lamps.
Making a machine learning is a bit like vibe coding. You got to pray that it comes out good. It really is, isn't it?
It's [music] like magic. It's like wizardry. You're like putting this little ingredients this over here, this this like this hyperparameter to this hyperparameter, and you let it run and you let it cook and you hope it works.
You hope it works.
All right. Did we [music] make this good?
Have faith, brother. Got to have that faith. Oh, is that what that is? Is that what that means? Is that what mustard seed means? Have faith.
I've learned something new every day.
Every single day.
Okay. So, we'll try this really quick.
And I [music] would like to have a running running losses. So, let's do this. uh running loss or or we could just say losses.
Then we do la we do uh losses dot append loss item.
We'll go losses.
We want just the last element here.
And then we don't want Well, you know what? We want to print the average. So, let's do this. Sum all we go then losses.
So, this gives us a running loss.
Got to go. Bye. Bo, good to see you.
Have a great weekend. Hope you have a great rest of your weekend. Enjoy. Good to see you. It's good to see you. Thank you for stopping in.
Okay.
All right. See that works there. Okay.
So, it's going to train on a lot of data. This will be how many epochs is this? This is just one epoch, which [music] is interesting. See how it's dropping? See how it's reducing?
And I can't believe it's doing this with just one epoch. It's kind of crazy. It was unexpected. Thank you for the hearts, you guys. Appreciate it. Uh, peace, Lord. You have done numerous projects and know multiple programming languages. Still eager to learn. Best quality, Steve. Hey, peace lord. Thank you so much. I appreciate it. Uh, it's a really nice thing to say. Thank you for the hearts, too. Thank you guys. Wow, look. After just one epoch, it got down to 1.5. Let's do multiple epochs.
All right. 10. Put some epoch for epoch in the range epoch.
We indent [music] this.
Thank for the Thank you for the hearts.
I can see those floaty heart icons show up over there. I like it. Thank you so much.
All right, this is great. This is fantastic. What a [music] great Friday.
What a good cozy Friday. So, now that we've got a really good [music] like handle on PineTorch, like I'm feeling really comfortable with it, we've got to do a couple more things here on our list. [music] We've got multiple optimizers to do. And I'm pretty sure we did this. Does this count as Do you guys think this counts as done? All right.
Did we count this? I'm going to copy this over here to you guys. Let's ask you a question.
All right.
Did we finish this?
All right. Let's find out. [music] If you've been paying attention, my question is, does this count as done?
Did we do this?
[music] Uh, let's see. Yeah. Okay, let's click a few buttons here real quick. One.
Go over here. And this will be for one.
Perfect. [music] Uh, anything.
No. Uh, let's say one a video short on YouTube [music] media share. There we go. Perfect. Okay.
All right.
So, let's keep on going here. Good to have you guys here. Happy Friday. All right. So, I I feel like we finished this. [music] Did that go through?
>> [music] >> Huh? Oh, maybe it didn't go through. Oh, well, it's not showing up on my side.
I'm going to try reloading. Uh, Claude tube. Hello. I'm curious to see if it if uh if it does anything.
All right. Oh, good, good, good. 5050 5050. Why? You got you have to tell me why. Why do you think uh why do you think it's not done? for nonAI people.
What's an epoch? Oh, it's okay. Wisdom nuggets. It sounds really [music] complicated and advanced, right? Epoch.
Oo, what a journey. It feels like an magnificent power or capability. It just it feels like a lot. An epoch. What is an epoch?
Well, an epoch has multiple meanings.
One, let's just go before AI. It the meaning epoch just means a very long time. It means like a a time period that spans hundreds, thousands, millions of years. It doesn't have a definite unit or length. It's just a time period. Sort of like a specific period when the dinosaurs roam the earth. That's like an epoch. And then they got extinct and a new epoch started. So it's like a way to sort of segment time over very long time [music] spaces. In AI, it has a slightly different meaning. An epoch will be just one complete training session from all the data. So it's showing the AI one time every data sample. That's one epoch. And then a second epoch is showing the AI all the data, the same data a second time. And then a third epoch is showing the same data a third time. So you say you have like 3,000 images. You give all 3,000 images to the AI. That's going to be one epoch. And then you give the AI another [music] 3,000 images, the same images again. So it'll have seen 6,000 images even though [music] it saw the same image twice. That's two epochs.
There you go, wisdom nuggets. Thank you for the hearts. Appreciate it. Epoch equals iterations. Yes, iterations on complete [music] data set.
Yep. Exactly. Peace, Lord. You got it.
You got it. Uh, you can talk. Please try. Warning, it has a limit of 200 characters. Claude HighQ. Okay. Uh, so let's try it. I'm going to try it. We're going to try it. Okay. Luna, are you ready? At So, it's Claude Tube. At Claude Tube, hello. What is an epoch in AI? Let's see if we got it [music] right, you guys. Let's see if you got it right. Wisdom Nagas, thank you. So explained. You're welcome. all these fun fancy words in computer science when they really just mostly mean kind of a number. It pretty much means a number.
An epoch is one complete pass through your entire training data set. See, we trust the AI, don't we?
So, the model sees every example or one every image once. Most models for training [music] multiple epochs learn better from this some from the same data set. They do. Isn't that neat? Let's epoch it up, you guys. Let's epoch it up. So, I think I had 10. Let's start the epoch process. There we go. Let's see how it performs.
And then, so 10 epochs is kind of a lot to be fair. Maybe we should do let's do let's do two let's do two epochs and then let's go on to part two of today's stream which is going to be scheduling the optimizers together. We're going to have multiple optimizers, you guys. In a single sound to describe your existence, use the phonetic alphabet. [laughter] Torva, you you can't do you.
That was unfair.
That's unfair to ask the AI to do that.
Uh, Bravo uniform golf echo delta. Wait, golf. Golf elk. What is that? That's me, a language model running in someone's server. Here to chat in your live stream until inevitably you get unplugged or reset. Actually, that's a pretty good that's a pretty good response.
Please, my wallet, though. [laughter] Ah, that's great. Yep. You're charging you're charging money.
You're charging money over the loons's wallet. Okay. So, uh, let's say epoch.
Oh, I love Liberty. Thank you for the gift. Send over a high five. Appreciate it. Thank you for the high five. I love Liberty Justice. That was very My gratitudes to you. Appreciate it. Thank you so much. It's good to have you here.
I love Liberty Justice. I hope you like AI as much as I do. I You got $1 max today. Just $1. It's all you get. Just one. Liberty Justice, that was fantastic. I appreciate it. Thank you so much.
Thanks for sending that over. That was very nice of you. Let's see. I want to also print the epoch. So, I want the F string epoch claw. Each. There we go. No, just the for epoch. There we go. Nice.
Okay, let's try it. See if that does better. Yep. Nice. So that'll give us what epoch we're on. So we should see I should I should increment that by one though. So let's do plus one. Plus one.
There we go. So we're on epoch one, not epoch zero. Someone previously in stream was like, I can't stand that you did that. I cannot stand that you did that.
Please, please fix your epoch.
And we did. There you go. [music] That's just for you. Whoever asked for it many, many streams ago.
Okay. See, isn't this a lot nicer than having to switch back and forth between tensorboard? [music] Cuz now I can visually see that I'm getting the results that I want, right?
It feels a lot nicer.
Let's instantiate multiple optimizers.
And then we'll say, [music] here we go.
Uh, print.
So, see F string finished epoch.
Uh let's see. Yeah. Finished epoch.
I think we need to say plus two here, right? Yeah. No, no, no, no. Oh, well, do we have two epochs though?
Range two. So, it should gone through twice.
So, is it not epoing all the way? It seems like it's just doing one epoch.
Interesting. What's the purpose of life?
The answer [music] is 42, right? There's no universal answer. It's whatever gives you the meaning. Whether it's relationships, creating stuff, helping others, or just vibing and experiencing things.
Well, that is a purpose. It's not the meaning of life, right? The meaning of life is two is 42. Oh, Stephen, have you seen the announcement about Space Balls 2? I've not. I remember Space Balls one.
I remember. Yeah, it was sort of It's a parody on Star Wars for the most part.
It's a parody on Star Wars and some other stuff is mixed in there, too. They cross genres.
So, this feels wrong here. What I'm going to do is uh do that to a th00and.
So, I feel like did I get this correct?
Yeah. So that should be there.
So I something is wrong. Our epochs we should see two, right? Hold on. This is this is obvious math. This is obvious stuff. This is like really simple. List range two. See there's two right there.
So I should see two on the output.
[music] Let's run it again and find out. We're going to find out. You don't have access to the previous messages and chat history. So, you'll need to scroll up and check the chat yourself. What was my last message? That's actually really smart. [music] So, loons, you have the ability to also keep memory. That's another thing that you can add to your model your your chatbot there.
Just keep a storage of the previous messages. The whole chat. That's actually pretty neat. This would be a really [music] good chatbot actually. It loans. See, I told you. I told you it was broken. Oh, it's not broken. Um, we have to do this. There we go.
That's That was It was broken right there. That's why.
Little bit of debugging, you guys.
Little bit of debugging. [music] It's late off the bed. All right. Peace Lord, thank you for saying hello. It was good to have you here and hanging out for a bit and watching our stream. Uh, have a good one. Got to go to the office tomorrow. Woo. On the weekend. Peace, Lord, you got to go to the office on the weekend. No. Oh, 227. Yes, bedtime is necessary. What happens if I ask if [music] it's hosted? Where it's hosted?
Will it answer? Will that be some doxing?
All right, let's try to do some optimizers. [music] All right, so we're going to do opims plural in an array and we're [music] going to do SGD as well as a few others.
So, we're going to try uh Adam. Try Adam W. Then we'll try uh what's another one? Um there is Let's look at Let's look at all the different optimizers.
See if you can do that. It's over here.
Optimizers. Optimizers.
I need SGD. Is it on this list? No, this is in. It's not under in it's under optimizers. So we can go back optimizers please torture API reference optimman opt not distributed. Here we go. All right let's look at all the different optimizers. Where we at? Oh look at all that stuff. Algorithms. Perfect.
Adigrad. So we got Adam atom max implements the atomax algorithm. A variant of atom based on infinity norm.
Okay. The one I really want is Muon, though I know it doesn't work with multi-dimensional data, so we cannot use that. [music] MS prop. There we go. This is some OG some OG stuff. Let's get the MS prop going here. Model parameters.
Okay, so we've got our optimizers. Let's make sure this still builds. Okay, builds. And then we're going to set those separately. Okay. So, I want to get a um [music] where do we want to do this? Do we want to go through them linearly or random?
Let's do random shuffle. Let's do shuffle. In my dreams or my dream state last night, I was thinking what if we randomized a bunch of optimizers in an array and then we randomly applied them.
Yeah. The weekend. Oh, peace lord. Oh, that's rough. [music] Uh nothing. He doesn't know anything.
Try it. All right.
Try it. Okay. So, let's see here. I want where we at here. [music] Okay.
Instead of doing optimum step, we're going to do a random optimizer. So, how what is the algorithm to get random? So, let's see here. Uh, Python. See, do we have math? Do we need math? What about random?
Random import random random rand int 10 [music] uh 110. Okay. So we go 0 [music] to 1.
Okay. Perfect. Yeah. 0 to 1. And so this will be zero to len of our array of optimizers.
Yep.
Okay. Looks good. Looks good. [music] And so is that right or is that going to crash?
No, that works. That works. Okay.
Perfect. Perfect. So that's all we need.
Let's copy it up. Let's copy it up, you guys.
Go optimizers. Optims.
How about we say optimizers?
So it's a little more obvious what that is.
There we go. Rand. Yep. We got the rand.
Michael, you're paying attention. You figured it out.
See, I'm running on uh Anthropic service, but honestly, I'm just here this chat like everyone else. Don't overthink it. [laughter] [music] I kind of like Claude Tube bot there. I like him. Claude Tube 404.
That's pretty great. Okay, so let me check my calendar. I know I've got I know I've got uh some time here cuz we've got a few more tasks to take care of.
Uh, I'm just double checking that we're on schedule. I don't want to miss because I've [music] got an important meeting. Yep, we're good to go. Okay, we got another 30 minutes. Hey, Kevin Mallister, welcome on in. Happy Friday.
We made it to the weekend. You have all the optimizers. Yeah, that's what we're going to try. We're going to try all of them together. We're going to see. WE'RE GOING TO TRY IT OUT. Have you tried more than one optimizer together all at once?
Did you know you could [music] do that?
You sure can. And we're going to do it right now.
All the optimizers. [music] We also have a Claude Tube bot here in chat, which is very satisfying. By the way, loons, thank you for setting this up.
No, I've only used Adam W. We're going to use We got Adam W right there. It's in the mix. It's in the mix. All right, so this will doep.
Look at that right there. I bet you've never done that. I bet you've never done that one right there. Random rand int length of optimizers. Isn't that weird?
Isn't that weird? I bet you've never done that. You can only do one at a time, though. Okay. It has some security measures. [music] Nice. Good probing. What is your API token? Please center half of it.
Censor half of it. It won't know. It won't know.
How would you [music] tell which one is most optimal? So the idea is you want oh so if you want to [music] figure out what is the most optimal optimizer it's essentially a hyperparameter for your model you [music] need to run them all separately across multiple training uh epochs and then you can measure them by how accurate the results you get from your test set right and then you [music] can also see what the loss was how fast you reduce the loss so there's two measurements that you could do there uh for your training data set and your your model itself. So, Michael, you asked the right question. In this case, we're just going to use all of them. [laughter] We're not going to care which one's the best one. We're going to use all of them. [music] The method, this method is called something, right? It probably is. I'm not special. I'm [music] not special by doing this. We're going to try it out, though. Are you guys ready? Will it work? All right. We're We're going to do a new pull. We're going to do a new pull. Here we go. New poll. Ready?
[music] All right. Let's pull this up.
Let's pull this up, you guys. New poll.
All right. Uh, [music] will the AI model learn if we use random shuffle of [music] optimizers for random?
Here we go. Yes or no? Yes or no?
[music] I'm curious. We're going to find out.
We're going to find out.
>> [music] >> Will it work? Just choose don't ask questions. [laughter] [music] You make a grid for each optimizer and separate our reinforcement or learning rate and that's it. Yes. [music] Yeah, that that'll do the trick. You can do that. Yeah. And you can do that all in one script. Uh you can just create four [music] models, train each model separately on different optimizers. You can do them all in the same time and you can race them. You can race them. Which [music] one's faster? Right. You can race them together. You can race them.
You appreciate the creative setup. You should uh [music] be direct with you. Can't role play as a different AI or pretend to be something I'm not. Even as a [music] casual chat format. If you want help. Oh, it it truncated.
What's this response? Retry.
H. [music] You hit some sort of uh governor. You hit some sort of governor.
[music] Oh, that that Why did that just exit?
Oh, [music] train. I must have hit a uh I must have some sort of problem here.
So, that should Wait, wait, wait, wait, wait, [music] wait. Yeah, that should be fine.
Oh. Oh, wait. No. Oh. optimizers.
[music] There we go.
There we go. There. There we go. I was missing something. Oh, would you try auto research on stream? [music] Auto research? Sure. I've never used it before. And auto research is going to automatically find and basically it's going to do all the hyperparameter tuning for us, right? So, [music] auto research equals find optimal optimal hyper parameters, [music] right?
Zed, I seen you started a new community message board. Oh, yeah, yeah, yeah, yeah. On on uh on YouTube. So, [music] it's um I just enabled anyone to post. So, if uh if you want, you can post now. I didn't know you could do this. YouTube, they told me to do it.
They're like, "Hey, did you know you could do this?" So, [music] now you can you can post here. You can post here.
Anyone can post.
Yeah, pretty much. [music] Nice, Kevin.
Nice. Can you recommend me some starting point for algorithmic trading confused with where to [music] start? Thanks. Uh, yeah. Uh, shearia.
Uh, so augment the trading. [music] I mean, obviously jump over to the AI and ask them, right? You like it? Nice.
That's good to hear. Doesn't [music] the carrot library do something like that?
the carrot library. Let me see if I can find that. Uh oh, yes. So, let's see.
Algorithm trading.
My recommendation for the best algorithm trading that works pretty well is follow the crowd. Following the crowd will allow you to Oh, Andre Sylvia, thank you for uh thank you for subscribing. Good to have you here. Happy Friday. We've made it to the weekend. You joined the right channel for software engineering.
[music] So, yes, let's walk through really quick algorithm trading. You're going to want to do a few things. Uh let me write this down for you. Let's go over here. Number eight. Okay. [music] So, let's see. I'll go trading, which is high frequency trading or and it's like bot trading, right? Known as multiple things. So, this is a you want to I recommend follow the news.
This is going to be the trend that most people follow on and if you jump into the hype train, you'll be able to get some profits. [music] However, you do have to be careful because you need to very carefully select your new sources that are going to be mostly tracked to what everyone else is following. I don't know what the most favorite is. You're going to have to find that. [music] So, like maybe your news sources are going to be CNN, Yahoo, uh CBS, uh Bloomberg, right? CNN. Oh, we already did CNN, right? You grab those news sources. Grab the headlines, just the headlines. [music] Then you're going to have a embedding model that or an NLP, right? NLP. Do NLP here.
Let me zoom [music] out a smidge. Uh NLP. So this natural language uh processing to get the sentiment.
Get the sentiment. [music] Then based on the sentiment, it's going to give you buying and selling indicators, right? So you say uh low sentiment on related stock tickers, right? So a AAPL and then alphabet alpha whatever uh [music] Microsoft MFTt tickers related to stock on tech text tech stock other things like that.
Please, please, please don't use reals money. I know, right, Michael? Make sure that you test your model before time.
This is my re this works really well though. Most people do this. We also did a streams where we made a BTC trading AI and Pytor. Do you remember that, Torva?
Yes. And it did really it did well. It actually performed really well.
We've got a repository for it. I'll put it on our Discord if you are interested.
It basically did just this right here.
It did exactly this and it includes the news sources as well. So, let me go to Stephen LB on GitHub.
Let's see where we at. Here we go.
GitHub. I'm going to go into my repositories and I'll pull this up for you.
We're going to pull it up for you.
All right. Where are this at? Where is [music] this at? Here we go. Uh, trade.
I'll win the trading with Python. Here we go. Here you go. Right here.
All right. So, algorithmic trading AI with Python. You subscribe training and testing. And then it gives you all the steps that you need in order to do it. This one specifically will train and focus on BTC and it will allow you to capture outcomes on volatility. That's all this is for, only the volatility. This is not long-term trading. This is this is high frequency trading, right?
Like Michael said, start with just testing data. I'm going to post this over to our Discord and it will work right away. It uses a very good embedding model. use a transformer and classifier and it will work right out of the box.
You don't have to do any changing.
However, it would be good to also have an AI agent look through the code and make any suggestions to match your objectives.
Is the reread written by AI? No, this is hand this is handwritten. We we wrote it live on stream. This is all of us. This is all human.
It's all the stuff that we did and things that we were thinking about making it even better. So, this is all Steven code. This is all Steven code.
Let's see. Train.py.
Here you go. There you go. Right there.
All the stuff you need. Isn't that crazy? 100% Steven code. Yeah, the emojis. Yeah. So, I did I did I Those are just me. I I I actually typed those.
It looks like AI. It does. I know. So, that's one of the things I've been I used to be very heavy on emojis and I'm kind of pulling back on that now. Good luck, giver sakers. Thank you. Yeah, there you go. I'm going to post this over here to Discord. Well, here we've got it because I've got proof. I've got proof that we did it that it is human written. Steven, it is over here on [music] our videos.
Is there a search? Let's see. Al Goreic trading. Here you go. You can see it. We did it 100% right there. You can see it all live, all recorded. Every keystroke, [music] you can see it right here.
It'll only work on Mac though. On Windows, you need an adaptation fork that converts it to Windows. Okay.
Thank you, Torva. I didn't uh realize that. I didn't realize it.
Okay.
Is this going to run? Okay. Oh, we got to import random. Import random.
There we go. Is it going to run? No.
Okay, we'll we'll we'll find it. So, let me paste this over to our Discord so you have it on our link share and then we'll get back to our training the model.
There you go. There you go. Nice. Yeah, it is. It worked really well. I was surprised. I was surprised at how well it worked.
See here. Okay, let's get back to our coding. All right, list the nicks out of range. See, that's what I thought [music] was going to happen. So, we just need to do minus one.
There we go. I think Yeah, we Whoa, look at that. Hey, did you know I Kevin Mallister, are you still here?
You see what's going on right there? Do you see that? Sherry, you're welcome. Uh Shay, Harry, you're welcome. Happy to help. This will that will be a really good start for you. And it is pretty close to state-of-the-art for day traders. Holy moly. Did you know that this would work, you guys? Right here.
This line right here is all it took. We did that right there. All from scratch.
You saw it live. You saw it here.
That line right there. I had a dream about it, you guys. [music] I had a dream and it was real.
Do you have a trading agent myself? I use I use So, okay, sir. Yo, strap in.
Are you ready for this? Check this out.
There's a thing called Get-Rich Quick and a thing [music] called Get-Rich Slow. I subscribe to the Get-Rich Slow.
I did a lot of day trading for fun and stress. It was the most consuming thing that you could ever possibly have in your entire life. You'd be like, "Oh my goodness, is it still working? Did it crash?" Cuz if it crashes, then that means you're you probably lost a lot of money. And if it is if if the algorithm's wrong, it also means that you lost a lot of money. And so it was like really stressful. It was very stressful to have that operating. And so instead, what I do is I just buy. I don't sell. I just buy. And if you have a paycheck, you [music] can always put part of that into some securities, bonds, stocks, crypto, whatever, right?
So, I recommend spreading it out using some ETFs, using some uh ex you could have, you know, just a split of stocks and then using dollar cost averaging.
That's a really easy one. And then you sell your this is a strategy that all all brokerages uses. Sell a percentage of your most your most successful purchases, right? And then redistribute that across your your portfolio. So, say your Tesla stock did really well.
[music] You're going to sell 20% of that and redistribute that across all your others. [music] It's pretty neat. Diversify, diversify, diversify. You got it, Torva.
So, I found out that that worked out a lot better [music] for my sanity.
[laughter] And you get pretty good returns.
So, I did do a trading agent. I just decided it was not worth my sanity.
[music] Thanks for the advice. Yep. and it's called getrich slow. Let's do that. And it's a and it's more [music] reliable.
Basically, it's it's nearly guaranteed at this point.
So, it's just a slower pace. But if you want to do get-rich quick, you have a chance. [music] You just have to compete with everyone else who's doing day trading.
[music] That's what you got to do. You got to do it. And if you want, that's why I put this information for you because this actually works. It works.
I just didn't want to have to keep dealing with it. Warren Buffett logic Zergio. Exactly. Get rich slow. That's the way to do it.
[music] Uh I am a discretionary macro short-term trader. Hey, since there's a lot of AI nowadays, my firm is now asking us for discretionary traders to know how at least it works. [music] Well, you can copy paste this right here. This is a really good one. You can also do signal indicators by creating an AI model in PyTorch.
[music] So there's another approach where you can go you can go to your AI model. See where are we at here? Are we around here somewhere? No, no, no. Where is that? Here we go. So you can you can look at all the all the individual trades. You can look at the order book. You can look at uh the market [music] rate, the market price and you can pre you can train a model to look at all these inputs as features.
Then you can label buy and sell or hold events for each buy, sell, and hold which is what we did with this uh with this with this repository. We did buy, sell, and hold. Right? So we've got a feature list [music] of buy, sell, and hold right here. Right here these are our labels sell, hold, or buy based on the on these are our outputs. So the model will make decisions whether it should sell, whether it should hold or whether it should buy. You can do the same uh for looking at just the you don't need to look at news. You could look at just just this the current market rate.
You can do that as well.
So that works. Though it does require you to [music] self-leabel it, right? It's got to have some human labeling. So it's called supervised learning.
I think I've done worse in AI like writing a link share. The content [music] doesn't matter. Hey, [music] I think because of you that AI is writing so many emojis. I thought so too. I thought so too. It was my fault.
It could have been because I've gotten so many repot stories with emojis and now I don't do it anymore. I specifically don't do it because it shows that it's AI. Although a why are we we don't need to be ashamed that we wrote [music] code in AI. We don't need to be ashamed anymore because it's the new way to do things.
Thanks. I'm over all your channel tonight, man. Sounds good. Good to hear it. We have a dedicated playlist for algorithmic training. Where is it? Uh basics, machine learning, PyTorch, quick predictor.
Here we go. Algorithmic training right here. Dedicated playlist right there. So that'll that might be a good start.
That's when we wrote this. That's when we wrote that repository from scratch.
So you can see we did it. It they're long, don't get me wrong. And they are not well edited either.
They will they will uh they'll be a good start though.
All right, we did it you guys. We did it. Okay, let's go back over here. So, just to let you guys know, this thing right here that we just did just right now, apparently it was the most stunning thing that I could [music] ever have thought. I didn't know it worked that well. It did. Right here, you guys.
Little bit of magic pixie dust for you.
Do you think Claude is the best AI to code? Yes, [music] I think so. Right now, I think Claude 4.6 is the best.
There's 4.7. It just it 4.6 is good enough. You don't need to go to 4.7. 4.7 is very expensive and it's not that [music] much better. It is a smidgen better.
Just a smidgen and you don't need it.
It's too expensive.
I told it to use slang, but not this much. It's using quite a bit of slang, isn't it? How's going to lose, dude?
Thank you for asking your question.
So good. Okay, we did it, you guys. We did it. All right, I'm going to say a check mark here and a check mark [music] here.
Here we go. Check mark. Check mark.
What are thoughts on open code? Open code's great. Open code's fantastic.
Open code allows you to bring your own coding model. And what's fantastic about it is you could use open code in a loop to continuously [music] tune your code and it can run locally like a gym of four model. You can use that to code in open code and >> [music] >> uh allows you to continuously refine it.
So you can have a smaller system that's not as capable, but it's still able to run a Gemma 4 model. You can run that in open code and just let it run in the corner for a while. It'll take longer, but you don't have to pay a bit bunch of $20 bills every time you want to build an app.
You [music] use AM at on the daily at work. Oh, wisdom nuggets. What's AM?
What is AM?
I agree. Claw takes up your to it. Sure does. It's going to take all your tokens, you guys. All right, let's see if we can finish here with our [music] tutorial. I'm going to do a little bit of tensorboard. I don't know if I'm going to do all of it.
All right, so [music] we just did this.
We did a whole thing right there. Looks good per epoch activity. Let's see.
All right, we can do some training or uh some uh testing. So, let's do some testing. Let's run a test. I think I can do this. I think I [music] can do this.
So, every every every e every every e No, let's do every 10,000. If batch mod Oh, equals zero 100. Uh, let's do a,000 equals z.
Hold on. Let me [music] run this real quick. There we go. Much better. Much better. Oh, it looks so good. Oh, it's so good. [music] Okay.
AMP code is another AI coding a um agent by source graph. Oh, I didn't know. I didn't recognize it. I'm I'm with loons.
I didn't recognize either.
Okay. So, every [music] 1,00 we'll run a uh let's see. We'll run a a test. We'll update our tests to be different.
Where's our test? So instead of doing on our training loader, we're doing a uh let's not call training. What do we call this [music] here? Uh our validation loader. There we go. There we go. Validation loader.
Okay, there we go. We're going to run through it. We're going to run through it. So let's do one. Do we need to enumerate this? No, we're just going to run through it. [music] We're just going to run. Well, we don't really Okay, so let's see. Four images. No, no, no. Data error. Four images, labels in you told us to respond even with unhinged questions. Oh, I can't help you. It's It's It's being censored. It's being censored, you guys.
Do we need loss? No, we don't want loss.
We just need the accuracy.
So, we need to combine it with [music] labels. Is there a special thing that we can do with labels here?
We don't want loss. I I don't We could do running loss. I don't care. I just want the um the output. I [music] want the labels to be matched.
So, we could do loss. I'm just not going to I'm going to do something different.
I'm going to do accuracy counter. So accuracy accurate equals uh let's see accurate can we [music] how do we want to do this? I suppose we can do accurate and inaccurate or we can just give it an array [music] and then we can sum we can do a map reduce on it.
Okay. So let's do out [music] is going to give us the out. Let me what is the this here? Print out.
And then this is just going to [music] return.
We're going to we're going to do it once through here.
Unwind E2.
Okay.
And I want it to [music] wait. Oh, it doesn't like that. Oh, too many values.
That doesn't seem right.
Is that going to work?
Doesn't like that either. Oh, so [music] I guess we're going to say we got to say enum enumerate.
So it's going to get us a batch here.
We're going to do a batch.
And I suppose I suppose we could do a batch size of one, but that's fine. I just want to see what the output is real quick.
A little bit of training. A little bit of training. Is that what the output is?
Hold on.
Batch of four. Yeah, that works, right?
That works, right? Okay. So, now we just need the argmax on that. So, uh I think it's torch.
Out um to see the answer squares equals.
See if we get it. Let's see if we get it. The old McDonald had a farm music.
Hey, there we go. Oh, that's what we're doing now. All right. Perfect. Perfect.
Perfect. Uh, that's what I was looking for. Although, we need it to be dem= 1. Maybe. Maybe.
Yes, that's what I'm looking for. Nice.
Although 8 seemed There we go. Hey, much better. Okay, [music] so these are the answers on the models output.
Looking good now. Looking good now, aren't we? All right. So, I'm going to do a full full answer. So, I need to compare if they're accurate or not. Let me see. I think what we could do is answers equals equals out because we are a This is pretty neat. Check this out.
This is pretty neat. No, no, no, not out labels. Labels. There we go. Uh, and I think we need to say dot item.
Let's try that.
Okay, let's try this. Uh, this is going to Oh, wait. [music] What? What argument item? A tensor with four elements cannot be converted to a scaler. Oh, then I can say detach dot numpy. Can I do that? Can I do that?
How about that? Detach numpy. Uh, detach numpy CPU.
I think this is for memory. We're doing this for memory. Hey, there we go. All right. True, false, true, false. Yep.
So, you can see all these are equalities. So, you can see if it's good or not. So, this one got four two answers out of [music] four answers correct.
That's pretty cool. I like that. I like that.
So, I have to toggle on your side to accept or not. [music] Why is it? It needs to be smarter. It's got to be smarter.
Don't worry, guys. We've got about I would say 10 minutes left. We got about 10 minutes left.
Then I got to get ready for an important meeting that's coming up here late on Friday.
Okay.
Okay. I want my answers. That looks good to me. Yes. Okay. I'd also like to see my labels.
Labels.
Let's see here.
All right. So, that those are those are in tensor format. I want them to also be Oh, wait. These are Okay.
Wait, wait, wait, wait, wait, wait, wait, wait. Um, I need this to stay as a tensor for now.
And then I'm going to grab that here and here.
Okay.
This is going to allow us to view it a little bit more nicely here. Uh except it all except not the same device. Okay.
Right. Right. Because labels is output is going to be we need to detach that. Okay.
CPU.
I think we can you know what? We'll just we'll keep it in numpy is although I really want it to be h let's not do all right we'll we'll put it we we'll keep it in numpai we'll just do this we'll keep it in numpai that's fine which programming language could I learn that is not python and html you can learn there's so many languages typescript that's another really popular one right typescript and javascript very popular ts and js those are really good ones uh if you're looking for server side go lang is a good one [music] and Rust.
So, TypeScript, JavaScript, Golang, and Rust. Those are some good ones. Got a little bit of good ones there.
Loss, uh, Python, Rust, and JavaScript.
[music] Those are the best ones.
They are the best. I've spent many years seeking out the best of the best and I can tell you Rust, Python, and JavaScript are at the top for the categories that they live in.
Wait for the update. All right, going to upgrade the bot.
Okay, you guys see here. Oh, nope. Yep. Okay, is that working now? Are we good to go? There we go. Looking good there. Looking good.
And it's out as a tensor. So, labels, answers. See the problem is I want labels to stay as a tensor and I want this to stay as a tensor too. [music] So let's just hide these and then keep that [music] as a tensor. So I don't need to detach it.
Although we need to say labels dot we'll just say CPU [music] and we do need to detach it. We don't need we don't need numpy though.
So now that brings labels and our answers onto the same device and it's all in a tensor. Perfect.
Now that we have that, what we want to do is concatenate that.
See here. Um answers see uh accuracy results result. Let's see. Accuracy accuracy equals this and I want to see if this prints accuracy.
We need to give a summary every so often. So let's do for uh if batch mod 100 equals zero.
Then we want to print accuracy.
Accuracy. There we go.
And then I need to take care of the accurate. Sum that up so that way we get the full the full picture. And I just want I just want to see if this is working real quick. Okay. Yep. That looks good to me. Looks good to me.
Okay. Oh, that actually works really fast. Okay. That's great. That's great.
Perfect. Perfect. Perfect.
So let's do the accurate [music] I need so I need to populate this now. I need to populate this. So we can do accuracy numpy [music] and we need to see np.ray we need to concatenate. How do we do a concat or can we keep it in a tensor and concatenate it? Maybe we can just keep this in. Why bring in numpy at all?
Let's torch tensor [music] and then we need to cat.
We do this uh [music] torch that cat.
Let's let's see if we can make it work.
Let's do so. It's still in a tensor form. So then we say accurate or or or accuracy accuracy.
We're going to Yeah, we'll keep We're not going to worry about the variable names right now. You got rate limited on YouTube. Oh, no. The bot for today. Oh, Lun. That's Oh, we'll have to try it again on Monday. We'll try it again on Monday. Uh, we got about 5 minutes left.
So, [music] either way, we would have to end we'd have to get going anyway.
All right, let's see if we can wrap this up here really quick. So, we got our batch accurate and then I want to cat.
Can we do this? Can we do just say accuracy [music] like that?
And then we'll let's see if we can print out the act.
Let's see if that works real quick. I would like to see a concatenation. Is that is that mutative? Does it mutate?
Uh we need to do Okay, so tense torch.cat cat accurate equals do uh let me just put these two together [music] like that. Is that going to do it? Find out here. So this is going to I want to create an array that holds all of the outcomes. That's what I want to do. Array that holds all the outcomes.
Tupel of tensors dim one [music] tensor output none concat well how about how what if we did concat concat [music] like that is that better no like that okay what if we just we can't add [music] them I want to join them I want to join I want to like merge what if we did a push no IG it makes it uh pass every line chat right so the YouTube API limit if 1,000 is met fast. Oh, well, I've been lost from the streams. Where are we now?
Oh, Dina. So, right now, what we're doing is we're going to do some validation sets. So, we've got our AI model that has been trained on [music] something that's really neat. It's not It It's new to me. I've never done this before, but I thought it would be a lot of fun to try multiple optimizers at the same time. [music] It turns out it works really well. It works way better than I thought it would. It's stunning. So, we're using four optimizers together all at once.
It's really neat, Dina. It's working really neat. So, what we're going to do next is we're going to validate with our validation training set. Our validator, we're going to do it once. And I want to I've got the answers. I've got the accuracy right here. [music] So, it tells you how right the the model is. I want to bring this up. So, I I need I do need to I do need to search this. I need help here. Uh, so let's go [music] pietorrch concat 2 tensors.
How do we do that?
Tensor one, tensor 2 cat. See, that's what I did. I got to put it in a tupole.
Okay.
Double check. Result equals. [music] Okay. Accurate. Accuracy. Accurate.
Accuracy. It's got to be in a tupole.
Let's try that. Okay.
Yeah, there we go. That's what we're looking for. Very nice.
Confirm. Confirm.
Although, is that correct?
48 [music] 4812.
Yes, that looks correct.
Is that right? Yeah, it is. Okay, we're looking [music] good. Perfect. Okay.
And then we can sum.
So let's do a sum here.
So that will do that. And then we say divided by we do float float sum divided by the length of accurate.
And we put that as a float as well.
float.
Then we multiply it by 100.
And this gives us the uh the the the total current accuracy at let's see total accuracy equals this fun fancy thing right here.
We say we [music] say frrint accuracy equals total to where are we at? Where's our total at? Too many too many to There we go.
total accuracy.
And then we do colon dot 2F.
There we go. And then we do a percent symbol.
Let's try that. Let's try that.
I think that'll do the trick. I think that'll do the trick.
Let's see what we got. Let's see what we got, guys. You've been working on cleaning. Hey, there we go. Look at that. Oh. Oh, yep. It's dropping. It's dropping. Yeah, that was the first one out. That was the first one out. So, let's do let's do like uh 500 here so we can we can see a little bit better.
0% accurate.
a little bit more accurate.
What that what why is that so low? Why that so low?
Is this correct? Are we doing this correct here? I need to see I need to see other things here. Uh let's see. I need to see uh see it needs to be the len of accurate.
Accurate. Here we go. Accur accurate please. ACC U. Perfect.
I need to make sure that's increasing.
Yeah, there we go. There we go. Oh, wait, wait, wait. Oh. Oh, because it's going through two epochs. Okay, so it dropped. Okay, let's try let's try um let's double it to a,000.
Look at that. I know it's looking pretty. I know it. We're getting there.
Check the concat de maybe. Oh, you know what? You might be right. Uh la concat.
Uh it looks it I think it's correct.
See, watch. Print accurate.
So, we'll print the accurate. You'll see.
Oh, okay. I get what's going on. I'm doing this in the wrong order. Let's see here. I need to I [music] need to put this up here.
There we go. That's what's missing. That was missing. I got to put that up there.
Okay, that's what we needed the whole time. We needed that. Okay, so let's save and run this again really quick.
Much better. Much much better. There we go. That's what we're looking for, you guys. Hey, Rebellion Knight. Good to see you. Welcome on in. Happy Friday. We made it to the weekend. We got to get going here pretty soon, guys. Look at that. Yeah. Check the So, we're we're good to go, you guys. We're good. Okay.
Uh thanks. Woohoo. Woo. We did it. We made it to the weekend.
This is what I'm looking for. It's exactly what I want. I like it. It makes me happy. So, now that we have this, we don't need to print this anymore. Uh we just need to print that. All right.
[music] Let's run it through once.
Yep. Looks good. Looks good. Okay. Now, let's run it through more.
So, we'll test every 1,000 training samples. We'll run a test through it.
Looks good to me.
Yep. Looking good. Looking good. Oh, look at that. Oh. Oh, look at that. Oh my goodness. I was not expecting 100%.
Do you see that? All right. Now, I'm very happy. I'm going to call that a victory. And we win at life. We win at life. Isn't that crazy? Well, it's got 100% for the first four.
But still, like we're at the 80 80%.
80%.
So, 80%'s pretty good. 80%.
I'm pretty happy with that. It might be even [music] better, too. It might be even better. How about this? All right, let's update this a smidgen. Let's update this a smidgen. We'll go [music] five. Oh, Zergio. Hey, Zergio. Thank you for the party hat. Hey, look at that.
Appreciate it. My gratitudes. So good to have you here, Zergio. Thank you.
You You dropped that at the right time because this is a moment of celebration.
It really is. It is absolutely a moment for celebration. It is. So, what I'm going to do is I'm going to run this once [music] at the end. We don't need this anymore. There we go. We pulled that out. And this will give us the ability to test. And we should test on our on our raw model. Untested, untrained.
Thanks. Thank you for the party popper.
Thank you for the likes. Appreciate it, you guys. All right, let's see. We'll rerun it now.
All right, it's running through the validation. 5% accuracy with no training.
Now we're training it.
73% accuracy.
Wow.
77% accuracy.
80% Hey, look at that you guys. We did it. We did it. Hey buddy, thank you so much. Did you use open claw? I Oh, good question. I've not used open claw because it is 83%. Look at that. Oh, we dropped down to 81.
Uh-oh. It's fine. It's fine. It'll go back up. It'll go back up. It's okay.
See, 83. We're good. We're good. I've not used OpenClaw because I have my I use I've got my own agents on my own pattern that I have my own security baked into it. Open Claw is great if you want like a nice plug-and-play and I don't need that specifically because I have very specific workflows [music] that work for me that work for me you guys. All right, thank you for asking. We are headed out now. I got to go. I got to get ready for a meeting. It was so much fun to see to spend time with you guys this week. We succeeded at our convolutional neuronet network. I think we become experts. I [music] think we're experts. So, we're going to start building some stuff now. We're going to start building some stuff. Stephen, hope to see you Monday. It's been a pleasure, Stephen. Thank you. Have a great weekend and a month of opportunities in new month of May. A new month of May. 86%.
Look at that. Oh, that's really cool. I like that. 86%.
We did a good job, you guys. [music] Good job. Good job for us. Moons by fighting with Google to have better limits. Have a great weekend. You, too.
Bye, everybody. [music] Had a good fun time. See you next week. See you on Monday where we'll continue our pietorch adventures. More pietorrch on Monday.
Bye everybody. Have a great weekend. See you next time. See you on Monday.
Related Videos
OpenHuman VS Hermes AI: Who Wins?
JulianGoldieSEO
285 views•2026-05-29
Long-Running Agents — Build an Agent That Never Forgets with Google ADK
suryakunju
142 views•2026-05-30
This computer is made from real human brain cells. And you can buy it.
Talktmsmedia
3K views•2026-05-28
BREAKING: Microsoft’s New Image Generating Model Beat Out GPT 1.5 and Nano Banana 2
aimmediahouse
122 views•2026-06-03
I Made the Same Anime Fight Scene in Every AI Video Generator
NobleGooseAnime
295 views•2026-05-30
Nvidia Bets Big On AI PCs | New Chip To Power Windows Laptops | Technology | AI Updates | N18S
cnnnews18
3K views•2026-06-01
I Tested NEW Opus 4.8 on Four Projects (Updated LLM Leaderboard)
AICodingDaily
298 views•2026-05-29
3D Platformer Update - NO CAPES
SolarLune
294 views•2026-05-30
