Blum provides a pragmatic look at the iterative nature of model tuning, emphasizing how real-time visualization informs critical engineering choices. His hands-on exploration of modern optimizers like Muon offers valuable, grounded insights for practitioners moving beyond basic tutorials.
Install our extension to search inside any video instantly.
Pytorch Tutorials: TensorBoard and CNN Part 5Added:
Hey everyone, how's it going? Happy Tuesday. We're going to be continuing with our convolutional neural network.
We're going to make sure that our model works pretty well. We're going to train it over some animal pictures and then we're going to put all that output into TensorBoard, which is the whole point of the tutorial. It took us quite a while to get here, though. I think we're going to finish today. We've said that before, so maybe you'll never know. You'll never know. I think we're going to find out.
Oh, I forgot to click the send button here. Oh, one second. In clicking the send button. There we go. Wait, send.
Yep, there we go. Perfect. Okay, so this is our tutorial and we're going to walk through it right here. We've got our optimizer and then we're going to do back propagation. I believe we've already achieved this here.
So, we've got optimizer and back propagation. I do want to try a few I want to try this real quick. Uh I want to change the the uh the criterion.
Wait, wait, wait, wait. Hold on. Uh what what did I have it at before? I don't remember. I'm going to try this though.
Okay, let's see here. Uh starting epoch. Okay.
And I've also added where? See? Oh, negative. Whoa.
Whoops. That's a problem. Hey, Tova.
How's it going? Good to see you. Happy Tuesday. We just happen to have a very big number problem.
It's like the the AI is is learning in the reverse.
Got a negative very big number. E plus 21. That's huge. Hey, Proud by Cory.
Good to see you. Welcome on back. We also have uh Haramana. Haramana, welcome on in. Thank you for clicking the high button there. Jensen 8534, welcome on in. It's unlearning it. Exactly.
Exactly. It's like it's like I'm showing it the right answer and it's like nope.
Like you know this is the answer. Nope.
Right here that learn this. It's like nopem not going to learn. And it just kept doing that until it got to a very big negative very big number. E plus 2100. It's a big number.
Uh all right. So I guess we can't use that criterion. Let's use let's use uh let's use this other criterion here. All right. This one I think works. I think this one works.
>> There we go. That's a little bit better.
And it's learning. There you can. Very slowly.
Very slowly. We're going to have to change our model up a little bit here.
Let's change our model up. Okay. I think we need to modify how this is working. I think this needs to change right here. I think some of this is going to happen.
Uh, we got down to 3.63, which is not a lot.
It needs to be closer to one. And what is our learning rate right now? Learning rate is 0.01. Okay. What if we reduce the learning rate? I'm going to try that just to mess around with the robot just to see. Oh.
Oh. Wo. That was way better. Okay. Our learning rate was just a little too big.
You're working on your PS1 emulator written in C. Are you serious? What? No way. You got to tell us about that prod by core. Tell us more. Tell us the details.
I'm kind of curious about that. Let's see here. Is that going okay? Good.
How's it going there, Kyle? Good to see you. Welcome on in. Happy Tuesday. Hope you're being productive and happy and satisfied. It's always great when you're being productive and happy and satisfied, isn't it? I think that's great. Thank you for the hearts, you guys. Appreciate it. So, we got down to 3.2 that time. What if we even did let's see here what if we did this and see all right zero what if we reduce the learning rate even further let's try it let's go even further will that make it better or worse it just doesn't do anything all right so this is our target learning rate and are we using a JLU or relu I think we're using a relu oh no using jellu okay what happens if we do a gel. Let's see. Let's let's let's change up to a rail here.
Rel.
Is that all the gels? Oh, no. There's more than one. Okay. I'm going to see if this makes it any any different cuz I like I like the gel you better. Also, I want to try the Muan. So, we got the Muan right here. I want to try the Muan optimizer cuz right now we're using AdamW. So, that's that's a couple of things. All right. So, let's see here.
Uh, let me get my to-dos going here.
Where are we at here? All right. So, let's get a few to-dos here. I'm going to use muon.
I want to uh you see if we can get the negative like log likelihood loss working.
See uh oss loons. Hey loons. Good to see you.
Welcome on in. You tried to be the earliest possible. You you did good. You did good. You're like we only started a few minutes ago. Hey Jins, you discovered your ch Oh, you discovered the channel like 3 weeks ago. Nice. and it helped you a lot. That's great to hear for your national AI Olympiads last week. No way. Thank you for your video.
I improved a lot and I may get to to the International Olympians. No way. Really?
That's great to hear. Jens, good job.
Good job, Jens. You did it. You did it.
Lun, you think's your record? Well, you Yeah. So, that's 6 minutes. You made it in at 6 minutes. So, that's a pretty good one. GG, Jensen. Good job. Good game. Good game, gent. All right, so I want to do Muan LG negative log likelihood loss. And then let's see what was else. Oh, we have to we have to we'll try Reu. We're going to try Reu.
We're going to do that right now. We're going to do that right now. Reu activation. And what else? I wanted to uh we need to update the model. Update model the uh convolutions and you know maybe a different different kernel size, you know.
See kernel size I think the best one is three is the answer right uh size of three reduce images let's see reduce images and then I also wanted to add in the tensor board so let's add in the tensor board so tensorboard output loss graph which is the whole point of the tutorial in 1 to 2 minutes it's pretty good You made it. You did pretty good. You're making your new personal best.
Okay. So, how do we do this? We can add a scaler. I don't log the running loss average per batch. See, now I've never done this right here. So, I don't know what this means.
So, we're going to try it. I've never done this before. And I've got our epoch, which is going to give us our And then we're going to have a different Oh, interesting. Training and validation.
Oh, and it knows what it means.
Okay, so this might be a graph with two lines on it. We're going to find out.
We're going to find out. You guys, you're planning on health routine this morning. Like, uh, what kind of health routine? Trying to eat more vegetables and getting fit. There you go. Eat.
That's what I did. I ate vegetables.
Yeah. cauliflower, broccoli, we had green beans, uh peas, uh some corn, onions, and we had garbanzo and kidney beans, and then some a mix, a grain mix. I had what is it? Quinoa, millet, grouts, oat grouts, brown rice. Like you did you know you can mix grains together? You don't have to use all the same one at the same time. You can mix them up if you knew that. How's it going there, Kraken? Kraken, welcome on in.
Shake some lettuce. Exactly. Right. Get the breakfast in. You'll thank me later.
Shake the lettuce. Well, I've I've been doing some I like I like kale and what do you call it? Arugula or rocket. I like those though. I kind of got bored of them and so now I'm doing just broccoli. And you get everything you need from broccoli. You do. You don't need anything extra. You don't need anything beyond broccoli. You don't need to do extra leafy greens. Just eat broccoli and you're fine. Also, I really like broccoli. I like a lot I really What I miss is kale. I know that sounds weird. You're like, "No, no, wait. I stopped eating kale because it just took so long to prepare cuz you really have to like wash it good." And some people like to par like boil it for a minute or two just to, you know, sterilize it and it's just too much work. I'm just going to steam broccoli. That's what Brian Johnson does. Was it? Wait. Uh Brian Johnson. Yeah, I think so. Is that right? Is that the you know the the guy that measures everything? Okay.
Uh data iterator images add graph net and images flush. What is this? Add graph will trace the sample input through your model. No way. Has it rendered as a graph this whole time? All right. I like this right here. This is really easy. All right. What What about images though?
It's only going to get one.
That doesn't make any sense.
Why would you only want add one graph?
We're going to try it and see what it does. All right, we'll try it and see what it does.
Kill goes too hard, does it? Lettuce is delicious. Uh, some lemon and you're good. Squeeze a bit of lemon and you're good. That does sound actually really good. Yeah, I'm with you on that. Yeah, iceberg lettuce. Very tasty. It is pretty good. I like it. Pretty good. and then we'll go through this. Okay, so we've got a lot of tasks here. I don't even know what this section is. Access to healthy food shouldn't be this hard in the US. It shouldn't be. Well, it's because all the aisles are filled with the really easy calories that are not good for you. The processed stuff your your body is not ready for. And all this like buildup can happen in your body.
Are you eating animal products? You're going to have gallbladder problems. Are you eating uh too much uh calcium oxide things? You'll have like g uh kidney stones which is a problem. And are you eating um you could clog your arteries if you're eating too much animal food, animal products or if you're eating those, you know, processed foods like all this all the problems will it's what you call it um it's an illness. It's what do you call it? illness. Uh, voluntary illness is what you're choosing based on what you're eating. You get to volunteer for your own illness. You are choosing it. I know it. All these foods are really tasty and addictive. They are the candy, right? Which going to give you diabetes or um you know the animal products. They're really tasty, right?
Animal products. They're going to give you a whole bunch of other problems.
It's a voluntary illness, you guys.
Why you a Why U G E? What's that? What's the G E? Why you are Wait, why you age?
Maybe. Adrian, Adrien, hurry, are you saying why you age? Your body has a built-in DNA mechanism change that for evolution purposes require you to have a maximum life. In the early age of life, there were animals that did not have that attribute. And what would happen is they'd multiply until they exhausted all the resources in the area and then the entire population died. So only the rem this is my hypothesis. I'm making it up.
I think this is why we age.
You're going on Twitch. All right, Lun sounds. Oh, hey, you're already on Twitch. Hey, look at that. Am I mod here? Can you make me mod here, please?
Yes. Okay. Uh, how do I do that? What is this? Click to reply. Uh, yes. We need to do that. All right. One second, you guys. How do you do this? Click uh gift.
I don't know how to do that. Uh do you guys know? What is this? Oh, wait. I think I just did it. Here we go.
All right. You're good now. Now you're Now there you go, Lun. All right.
Success.
I think I think you did it. You're good to go. Wait. Are you quitting Python?
Yes, I am. Zero gaped. Zero gaped. We are absolutely doing Python. He's a software engineer. Zero gap. Yeah. How do you get subtitles on a live? Oh, I've got that ready for you. So, we've got subtitles right here. This is software I wrote 8 years ago. You can see right there. A little bit of 8 years ago there. Some 8 years. And it allows you to add subtitles onto your live on Zoom, on Microsoft Meet or Teams, Google Meet, whatever you want to do, you can add them. Hello everyone. Hey JM, good to see you. Welcome on in. Happy Tuesday.
Good to have you guys here. Yeah. And so you can have these subtitles here. So check this out. I'll I'll paste this over here on the link share game cuz it's very often link share. Here we go.
Ooh, we got an open source here. What is this? Uh the open source AI lie. Wait a washing. Broken definitions. Who benefits? No major AI models meet the open source definition. Who's faking it?
They are. They're all faking it. They're faking it. All right, this is for the stream here. I'm going to remove that here real quick. We go. Okay, that's sick, bro. Isn't it? Zero gap. Yeah, you can. It's right here on Discord. So, if you want it, you can. It's It's open source and free forever. How's it go, Code Meitsu? Good to see you. What's up, broskies? Welcome on in. All right, let's get back to our tutorial here. All right, we've got a lot of tasks here. We got a lot of tasks. And then we needs to uh finish finish the tensor the tensor board tutorial which we should be able to do today. I think so.
So there's a lot of things. These things are going to be really fast right here.
We can do those really easily. Good to see you, bro. Good to see you too, Mitsu. How's it going there? Uh Neggo, hey Neg Lego, welcome on in. Sigma Edits, you're here, too. Hey, welcome on in you guys. You're all all jumping on back. Good to have you here.
All right, so let's get back to our code here. Good to have you here. Thanks, man. Thank you. Thanks for stopping on in and saying hello. We're going to be some coding convolutional neural networks on our journey to become a PyTorch expert.
You're going to watch as I work through all the tutorials and mess with the code while I learn how to become a PyTorch expert. We're pretty good now. We know most of our stuff. So, I want to mess around with this really quick. Do we need to pull everything? Do we always need a pool? And then we're also going to do we need this many comp nets. So there's a lot of stuff that we have to do here. Every time we make a change, we have to rebalance our inputs and outputs between each of the layers. Hey, donut do LFG. That's great. Are what are you giving away? Oh, giving away knowledge.
More AI hype. Yeah, exactly. That's what we're doing. Join the Discord, you guys.
Thank you, Kyle, for mentioning it.
Thanks, Mod Tools. Here are 10 times better. Nice.
Oh, nice star. All right, that's actually the best method. Can't learn without proactive trail tests. You got it, CodeMu. We got to mess around with this. All right, let's do Oh, so we did.
Did we try Reu yet? Let's try it. Let's see what we get with Reu. Uh, Jellu did a little bit better to be honest. I think our jelly look reel is having a problem. It can't even barely. We needed Jellu. We needed it. Let's try negative log likelihood loss real quick. Let's see if we can do anything here. Yeah, Jellu is way better. Unless we need to increase the learning rate.
We might need to do that. See? Nope. All right. IT DID NOT LIKE THAT. It does not like that. As you can see here right now, what you're looking on the screen is a problem.
You won't be able to stay long. Oh, but you're excited. Oh, thank you, co uh donut dude. Glad that you got to say hello and and visit for a bit. Just talking something we can use forever.
Oh, okay. So, we could we are having a hackathon later. Not not today, not this week, maybe not next week, but maybe the week after. I will put when I when I know the date, I'll share it and there'll be there'll be like prizes and stuff and it's something you can use forever.
You can use it forever.
So, negative log likelihood loss is not good. However, with ReLU, do we need a larger learning rate? Uh, here we go.
Little bit of errors there.
All right. Yeah, Railu likes bigger learning rate. It does. I'm think I'm going to go back to Jellu, though.
Let's go back to Juu. All right. So, la.
We'll do this.
Reu with jellu. There we go.
Okay.
Uh, all right. N lll. We need to not have that one because obviously it's a problem. There we go.
HS, good to see you. Welcome on in.
Happy to t Oh, yeah. It's Tuesday. It's Tuesday. All right. So, it likes that.
Oh, it likes that. Well, okay. Let's see if we can mess around with more of this up here. Uh, I think we need to So, this is our comet one, and we're pooling here. Do we want to just pull straight out like this? What if we just did this?
Uh, wait, wait, wait, wait. We're going to need to do input.
All right, let's mess around with that.
Is that going to I think that should work without crashing. No. Okay. So, we do have to modify it. Uh because we're pooling a lot more. Okay. Here. You know, let's just try it. Let's just try this. All right. Let's mess around with this. All right. We need expected input.
Got 2633 which is up. Oh, right. Okay. One second. Uh, let's do let's do one and let's do three. Can we do that? No. No, we can't do that. We can't do that.
Okay. Uh, because of how everything's aligned up here. Uh, I guess we'll mess around with it. Fine. Fine. We'll mess around with it. We'll find Oh, hey loons. Thank you for the pen. I didn't know you could do that. That was pretty neat.
Hey, Quantum. Fight Quantum. Welcome on in. Good to see you. Happy Tuesday.
We're continuing with our CNN convolutional learner network. Okay. I want to also grab this here. We need to change this up here cuz we're going from one to two. Let's go to one to two. Let's just do one to two. One to two. And then what we got to do is Is this your goal? You just did it, Kyle. Woo. Check that out.
>> This is your goal.
>> Thank you. You just did. I was not expecting that to happen at all. Kyle, thank you. My gratitude. Thank you, Kyle. Shout out to Kyle. Flip with Kyle, you guys. Thank you. Explanation point.
And then some party poppers. It's all always, you know, always got to always got to do the Thank you. Thank you very much, Kyle. So, what what you basically obviously we do this anyway for you, Kyle. For you, Kyle. However, if you've got a link to share and it's classroom safe, we can take a look at it. We could take a look. Check this out.
Explanation.
Exclamation.
See, it's got some weird little syntax things. I need to fix that. That's got That's a bug. We found an exclamation bug. Oh, yeah. So, Kyle, yeah, if you have any links to share, if you have anything to share the links.
All right. Set another goal. La. Okay.
Uh, see another one. Okay.
Anything you got? Extra curly brace. I know. Code me too, right? I was like, why is it there? I have no idea.
Loons. Thank you. Thank you for the links. All right. So, let's see here, you guys. So, I want to We've got our inputs. We're doing We're just going to reduce this down. We're eliminating these two lines here. So, we got that line and then we got this other line over here, right? So, we got rid of those. And then we're only having two layers here. When we do this though, it's going to ask us for a specific size, which apparently it didn't. Oh, is it learning a lot better now? Yeah.
Yeah, that's learning a lot better.
Look at that. Look at it go, Stephen.
You heard? Uh, maybe we're going to find out in a second that you are having a breakthrough in stopping aging in China.
Wait, they are having a breakthrough in stop. Are you serious? Really?
Did they find the aging genes?
Because here's my theory. There are Well, I I just explained this like 10 minutes ago, though. I'll explain it again. Aging is a builtin mechanism for a species, and it's by design through nature and natural selection.
Because there have been species that didn't have this gene populated so much that they exhausted all the surrounding natural resources and there was nothing left to eat. So they went extinct the entire populace and only those with the gene mutation that allowed them to age and you know maintain a stable population were able to propagate and that's what we're left with today. Hey Dreamer, welcome on in. Good to see you.
We're just talking some health today.
Stopping aging or slowing agings. Uh we know we can slow aging a smidge, right?
And I don't trust the thing in China per part. Yeah, exactly right. I know. They always make these big claims and you're like, I don't know about that. I'm suspicious.
I'm suspicious of what China say. Let's go in there. Uh Saram, hi Ank. Welcome on in. Yep, we are. Ank, you used the proper term. Good job, sir. Good job, sir. or bro, however you want to say it.
Okay, I want to uh apparently I don't know how we did this. I have no idea, but it works a lot better now. Look, we got down to 2.9. We got down to 2.9.
Heard aging keeps aside caption America direct from frozen comment chat. I have no idea. Dreamer, I don't know what you're saying. You're really good at You're really good at saying words a whole bunch. How do you get those subtitles? Oh. Uh, three three of them.
These are in our Discord. So, you can see it's open source. I This is code that I wrote many years ago for your YouTubes and stuff. Oh, wa. Hey, one of Torva. Thank you. Whoa, that's amazing.
Torva, what are you doing over Torva? All right, Torva, thank you. Explanation point. Few party poppers in here. Torvo.
Whoa, that was unexpected.
I appreciate it very much, my gratitude.
And indeed seems to be the goal, Kyle.
It does. Did you reach it?
Let's see. Are we're almost at the goal.
How How do we get the subtitles? Oh, how's it going there? Uh uh Jiao Xyo, the subtitles. Yes. So, they're in our Discord. you can get the subtitles and they work in OBS which then you can also use in you know Zoom, Google Meeting, Microsoft Teams and you can have live subtitles over your over your output.
Isn't that neat? You can do it. Yep. So it's right here in our Discord. Just click this link share link right here.
You click that. It's going to open up a new window and it's going to give you instructions on how to install it yourself. It'll work. Quantified quantum used it and proved it. Stephen, captions like this. You need to use web Google's speech to text then broadcast via websockets. Yes, exactly. Exactly.
Exactly.
And is there a package JSON? And so you can see let's see here only dev dependencies are needed. There is no in JavaScript here. It's called JS. There is a library called subtitles. Is this it right here? Yeah, right here. And that does the the majority of of the coding here that you need. That's where all the answers. This incredible, dude.
What? I know, right? Completely open source, free. You don't have to pay for a single thing. It's absolutely free.
What is your content about? By the way, you're new. All right. We do software engineering here. And my name is Let's see here. Let's see if I can scroll over. Where am I at here? Here we go.
All right. My name is Steven Blum, CTO at PubNub, technology company that is for communication APIs, which is what we're using right now. You see on the screen these subtitles right here, subtitles, they are actually powered through my technology communication that requires you to have an open browser window. So you have a browser window and then you can pipe that data into OBS. We see here's OBS right here.
So that's one of the use cases for my technology. Obviously we've got a lot more use cases, just a few. Are you guys hiring? Yes, we are. So, let's go to PubMed careers. Let's see. Do we have a careers window open? I think we do.
Yeah. Here's my website. Go to my website. Go to our careers.
And we've got a senior backend engineer for AI and large language model systems.
And we have a senior solutions architect. So, we got two open software engineering roles. Stephen, you recently heard about IPv8. There's no way.
There's no way. There's IPv8. you we don't need a 98. We only need six. Six has 433 trillion possible addresses, right? You don't need any more than that. Or is it trillion trillion? I can't remember. It's a number that's so big every device can have multiple IP addresses and you don't need NATS anymore. You don't need network translation tables. Senior, it means I have a lot of experience, right?
Seniors, you got it. Well, we're looking for the experience. Although apply regardless I always say swing at every every everything swing at everything.
Stephen if you need an intern I'll be available. Quantify quantum really okay so that is something that we are considering.
quantify quantum. Let the st send me a DM because I we will be I I'm pretty sure going to have an intern. Uh just one just maybe more than one. Maybe more than one, but right now I'm thinking one for now. Stephen, if you have if you need an intern, you'll be available.
Good to go. Send me a DM.
Looking for an intern. Uh we not there's no official start for this yet, so it might not happen though. I'm I'm pretty sure we're going to Torvvis European. I don't know. Flash uh flashback to LinkedIn. It's still a concept.
Cool. On now, doesn't All right. Nice.
Oh, okay. Yeah. So, very likely intern.
Yes. Uh you're going to message you on Instagram, Stephen. All right. Sounds good. I I'll be ready. I'll be ready for you.
Got your open rolls here and maybe an internship. Maybe. Maybe just one.
Captain America was frozen in Marvel.
Capt. I like how you typed the caption America.
Dreamer. Caption America.
Yeah. So, what we do on this channel is software engineering. Uh, and I will be building AI. We also have a new stream coming up in the next couple of weeks that's going to be focusing on new technology that we're going to be releasing and I'm really excited about it. Really excited about it. Frozen and Marvel movie. So his aging stuff that's why aging mentioned thank you for compliment on my word story telling this will and continue enjoy. Okay. All right. Also available for internship.
You have plus one. Yep. Exactly. We'll let you know. I'll let you know if the opportunity opens. I'll throw it on Discord. It will be a link in Discord.
Can we implement AI into a game in an innovative way?
I think that sounds like a good call.
I think that sounds like amazing. Here's why. Most games don't do that. Most games have AI in sort of like player player opponents, right? Or some sort of natural environmental AI.
They don't that that's as far as they've gone with AI and it hasn't been full like neural net AI, right? It's always been some sort of human coded set of rules for AI. Never mind. Message you on Facebook. I couldn't find the IG. Uh, see, so that works. You can also send send on LinkedIn as well. You can also send me a message on LinkedIn, which I'll throw on Discord here as well. Working on a project for people to test for some feedback. Yes.
Let's see. Reply. Let's see. Uh, yes.
All right. Here's my LinkedIn uh page.
I'm going to paste that here. So, you can also grab that link here.
It's in the box for sure. Yes. Isn't there that one new game that uses AI companion? Is there? Is there? You got to tell me what it is. What's the name of the game? It's a website that lets you spin around the globe and listen to random radio stations. Oh, very nice.
For free.
Unfortunately, as humans, we do not propose uh possess antifreeze proteins.
Ah, right. Exactly. And if we froze ourselves, that would that would be a problem. It would be a little bit of a problem.
Cuphead. Yes, I have. Sam Sam made me play Cuphead uh a few months ago and we played it at his house on his Switch.
Once Genesis from Terminator comes, then AI will have a new face. Uh I think so.
I think so. I I prefer I prefer the utopian futures. Thank you. I would prefer not the dystopian futures. I want the utopian. Thank you. Okay.
So, what do we want to do here?
There's also an AI like demo made a few weeks ago on the Minecraft AI.
Oh, yeah. That's right. That was pretty neat. Like it was like an AI shader.
I think it was an AI shader. Oh, hey.
How's it going there, Daniel Harts?
Welcome on in. Good to see you. You agree you're learning pretty hard on the solar punk modality of optimism. Hey, there we go. I like that optimism. Let's keep that optimism going because we can be we can be destructive and tear things down, which is doesn't really help that much. It doesn't it only hurts people.
What we should do instead is like, "Okay, I see what you're doing over there. Okay, fine. We're gonna go over here and do this other thing that's going to be amazing and it will change everything.
Solar punk is the way. It's the way. You got it. You found one meme. So, I'm going to say in the chat, you have to imagine. Okay. Uh, dreamer. We'll give a shot. We'll give it a shot.
All right. I want to try Muan.
Let's try Muan. Muan.
Let's try that there. So, Steven built an A. Oh, that we can't do that. Yeah, we can. Right. Right here.
Oh, I spelled it wrong. Uh, there we go.
The U before. Uh-oh.
Oh, and we got a problem here. All right. One second. One second. Let's see. This will be Adam W. Let me double double check that this was correct, that we didn't mess anything up.
It should be a drop in replacement. It really should. I don't know why it's not working. LR for learning rate and params. All right. So, that's good.
Let's close that. So, I want to use Muan. We're going to find out why it's not working.
Basically, our cells have uh a ledger that controls cell data for duplication.
By the time it degrades, they are claiming to found a way to stop or slow it down. Really, I want the magic. I want the magic pill. Let's stay alive forever. So, Steven built an AI that learns from user behavior. I'd love to see that taken further. Kyle, you remember that? The AI chase game. That was a lot of fun.
Let's use our technologies to learn and improve our biomes, overall success of our planet. Yeah. Not just our species.
I feel like it's the humanity's higher calling. Not fighting about everything.
Exactly. Instead of fighting, let's build a better world.
AI coming to conclusion in the near future as human beings deemed as obsolete. Oh, code mess. Well, it really depends. If the AI ends up having some sort of personality that's leg real and it decides it wants to take over to protect us, the AI is going to protect us. We got to be it's going to come in and take share take care of things.
I found someone who remembers that a guy always say thank you. Hello. Treat a friend moral. Always say thank you.
Yeah, that's a good that's good advice.
Dreamer. Yes.
Look it up. Was a propaganda. Oh, it was it or a pragmata.
Not actually AI. It's more like Undertale. Oh, yeah. Undertale. I remember that game with a ton of statements. A ton of if statements.
Yeah, that's that's what AI is in the game world.
That's exactly what it is. Uh, hello.
One quest. Who is that program? Yep.
Yep. You got it. We're doing Python.
We're doing some Python. You guys, you know what can uh force us to make a better world? An alien attack. An outside force. Hey, quantified quantum.
You're actually correct on that matter.
Humans have this special ability to unify against a singular enemy. If we happen to get one, it's going to happen.
Oh, hey Christian. Christian, thank you so much for subscribing. Appreciate it.
Welcome on in. You joined the right channel for software engineering. What we're doing today is we're going to build a convolutional neural network and we're going to be messing around with a bunch of hyperparameters and we're also going to be using tensorboard. We're going to be pasting everything on tensorboard.
Daniel Hartz, thank you for the two jewels. I appreciate it. Thank you so much. Christian, thank you. Thank you for subscribing. I appreciate it.
Pragmata is the game's name. Oh, got it.
Okay.
Good old Pietorch. Tokenbased personality. I know, right? Right now, I don't think we have much to worry about cuz we have really clever autocompletes.
We have some really clever autocompletes.
Thank you for the two jewels again, uh, Daniel Harts. So, we don't have to worry about it right now unless someone puts the AI autocompleter into a position that it can make decisions.
We have to be careful about that. Right now, we don't have to worry about it, though. Okay. Let's go to learning rate optimizer.
Uh the muon and then we need to see here muan only supports 2D parameters where we found parameter size with a very big a really a man. Oh I suppose can we change the view to spread it out? No for the optimizer on the output but the output size is supposed to be 10. So it should be fine. only supports 2D parameters whereas we found parameter with size of 6333. So I guess we can't use mu on here because it's multi-dimensional stuff. A oh man. All right.
235 two 2335 alien decided to put a blockade around the asteroid belt stopping the space shipments. Huh. Ah I see what you're referring to there. I see what you're referring to. You wrote a personality algorithm, Kyle. Tell us more. Researchers have unvealed some interesting findings on AI. More to the story to be told. Code me. Always interested. Always interested. What are your thoughts? Okay, let's find out.
Someone someone Let's see what we got here. Let's see what we got. Given a task could be anything like porting, migrating, new features, implementing, bug fixing, very long code change. What is the worst place to use LMS for code generation and the best place? So the LLMs in scaffolding when you have a scaffold like a coding agent that has all the surrounding sort of guidance on how to deploy an LLM's task. It has very good capabilities at tending to the initial part of the prompt and everything that gets too big, it's going to start forgetting. So you want to give it very small prompts. Anything that can do tasks for that, you're good to go.
You're good to go. So small coding tasks, you'll always get essentially near perfect outcomes. You want to do small coding tasks trying to switch off Superbase. Oh, Lun, really? There's a storyteller in YouTube ego videos are amazing about clicker cookie clicker around paper clips almost a story to world optimize paperclip production. Oh right, right.
The metaphor is that when an AI go has a goal, it's going to reach it without knowing the consequences of its actions.
Oh, hey Andrew. Thank you for saying hi.
Welcome on in. Stephen, how good are two billion parameter models? They can be very good unless it's an LLM, then it's sort of okay, good, right? So, for example, this is a, you know, not even a million parameter model and it's going to do very good because it's got a very specific task and it's targeting some very specific goals. So, if it's knows exactly what it's going to do, you can optimize it for a specific task and it's going to perform it very well.
So if you want to deploy a 200 billion parameter or two billion parameter model, make sure it's very targeted at a specific goal. So you might want to fine-tune it to do something very special, something very specific.
Hey, the F4in, good to see you. Welcome on in. Happy Tuesday. We're just uh trying to use a Muan, but apparently the Muan doesn't want to be used right here.
It doesn't like it because we've got multiple dens. It supports 2D parameters, but we have got 3D parameters.
I thought we should be fine with but a muon is going to optimize on the gradient from the back propagation. So, we have to stick with atom w. We can't use a muon. Oh well.
Uh think if Alibaba made 500 million perimeter model. They I think they did.
Uh yeah, you can do that. So, we're getting pretty good loss rate here. I want it to be a lot better though. I think we're fine for now. I think we're fine. I think so. If if we did a one hot, we should also add a one on in here. Let's do a one hot. One hot output. Cross uh what is it? Criterion for the cross entropy loss. One hot output always always outperformed. I've seen it every single time. It outperformed every time.
That gives you confidence in your current model because it's specifically coded to a small use case. It's code meu. There you go. There you go. If you have it targeted for a very specific use case, it's going to be very good. It'll be very good.
If the AI has goals, which it doesn't, it could easily just blame a human for whatever it does and use us as pawns.
Kyle, we're saying all these things, but you know what? In the future, it might happen.
It sounds scary, but in the future, it definitely might happen.
We got to be careful. Quinn 3 500 million or something like that. Nice.
Yeah, that sounds good to me. That sounds like it work perfectly fine. All right. So, sad face. This only works for 2D data and we are using 3D.
So, that's why.
So, we're using 3D data unfortunately.
Okay.
So relu it. This wasn't as good as Jellu. Jelu better update model convolutions. So we did this then this actually worked even better than we thought. So let me put a green check mark here. So that was successful.
We can do some kernel size and tuning on the images. Before we do that though, what I'd like to do is get our tensor board showing our output for our loss.
That I think that sounds like it would be really fun to do.
Let's see if we can figure out how to do that.
Okay, so let's do a batch. Well, we don't need any more images, so I'm just going to comment this out for now.
I'm just going to hide that.
Oh, Anastasia, good to see you. Welcome on in. Happy Tuesday. My main concern is AI is gonna automate. Oh, autonomy on military operations. That sounds scary.
That's when we get that's when we should worry because then the AIs are wielding weapons and they're making decisions all on their own, right? They want to embed into weapons like in drones AIs that can make selfdeisions because it prevents it from getting cuz right now in order to control those things we have to broadcast transmission and if that transmission gets interrupted through some you know blocking then it can't make any decisions anymore right a human can't make a decision so if we embed AI into them then we don't have to worry about that blocking occurring ever again Bonzupi, good morning. Welcome on in.
Good to see you. You're trying to make the gooey show what the terminal prints out. Yes, exactly. Loons. Exactly.
That's what we're gonna do. That's exactly what we're going to do. It's cute how people think it's not already happening. Oh, actually, it's a good point. It probably already is. It's already happening.
Already happening. Just woke up, Bonsupi. Well, good to have you here.
Welcome on in. Good morning. It is. I've been awake for almost 11 hours now.
Basian theory in your model. Oh, nice.
Zergio, tell us how it goes.
Tell us how it goes.
Okay.
Different song.
That's okay. That's good. Okay. Let's continue onward with our tutorial here.
All right. I want to go over where was that? Here. Uh was around here somewhere.
Here we go. Okay. So, there's an add scaler. I've never I'm not familiar with this. However, we are going to add it.
And every mini batch. Okay. So, I'm not familiar with this. So, writer.addcalers.
And it looks like there's a trainer scaler and a validation scaler. Okay.
So, we're just going to do one of these.
We're going to do a loss scaler.
So writer add scalars I think plural right and we're going to say uh model training loss and apparently we have to pass it a dictionary we're going to say graph equals dictionary loss and what is that called? Oh, pretty.
There we go. Loss average. Loss_abg.
Like that.
Graph. Here we go. So, I think that should work. Is there anything else that we need?
I don't think so. Oh, wait. It wants some It wants some extra data here. Does it really? Uh, it looks like it's mandatory. Okay. So, we need to add an like a EO. I'm just going to add epoch here. Ep well uh what's a writer and the scalers? So the writer is going to be what puts data into tensorboard and the method add scalers is similar. So we've got another writer example up here. So you can add an image, right? So we've got add image.
Scaler is a single number. It's a single number. It's one number. A scaler is a value. Single value. So we'll be adding one value at a time. And I think that should work. So, it's going to append a bunch of data points into a graph.
You want to buy a 5090? Oh. Oh. Oh, you're selling your 5090. Missed it. M.
Uh, missed in mist.
A 5090 sounds like it'd be pretty nice.
I don't have anywhere to put it. Another number. Exactly. Loons. That's what AI is. It's all numbers. Did you guys know AI and math? a whole bunch of numbers.
All those different words, right?
Scaler, vector, array, matrix, they all mean the same thing. A certain collection of numbers. It's just a bunch of numbers.
That's part of the wild clawed supply chain. Oh, the whole the whole supply chain at risk. Yes, that's right. We got to be careful. The F1. Mhm. Got to be some careful there, Stephen. Fun fact, a month ago, we successfully moved a payload of antimatter. Did really?
Between two facilities via truck. How did we do we do like a magnetic? How do we Well, antimatter is just regular matter for for the most part, right?
It's regular matter with an opposite something. It's it it looks like regular matter. The only problem is when it comes in contact with non- antimatter, right? The the positive matter. I suppose it will instantly annihilate and convert into pure energy. Magnets.
Electromagnets. Nice.
We trapped antimatter in a magnetic field. Hey, look at us. We're humans. We know what we're doing until we blow ourselves up with antimatter. We got to be careful. All right. So, is this going to work? I don't know. We're gonna find out. Uh, where can this happen at here?
So if back So we need to we want this back. Here we go. Uh let's average loss.
Let's do let's do let's do uh let's do 50. And let's instead of having it here, we're going to we're going to put this here.
There we go. And we will break. So we'll just say epoch. All right. Let's try that.
It's working.
It's working. All right, we're going to see if it works. CERN talks. Does it CERN? Like the What is that? The particle accelerator facility. Is that CERN? Particle accelerator facility. All right. Go back to TensorBoard.
TensorBoard.
Close. Reopen. And let's open this up. I think I think we already have it open up over here. Here we go. Reload.
Look at that. Look at that, you guys.
Check that out. Oh, that looks great.
Okay. All right. That's a lot easier to deal with. Although, I'd much ra h wait. All right. Let's try this again.
Hold on. Hold on. Let's wait. I I wonder what happens while it's learning. So, if we if we rerun this, let's close and open it. Let's reopen it. And we run it now.
Can we reload?
Okay. Okay. So, it it's not real time.
It only shows us.
So, it goes below three.
Okay. Uh, that's fine. That's fine. It's It works. Works for me. It works for me.
Flush. Boom. Okay. Reload.
That seems fine. All right. But at least at least we get to see this. So, look, this says because the loss is pretty linear here, we've not hit the bottom.
We can keep going. The model will keep learning. It'll keep learning. It hasn't hit its floor yet. Whoa. Kevin Mallister, good to see you. Welcome on in. Happy Tuesday, everybody. Kevin Mallister is back. Yo, Kevin, how's it going over there? Are you back in Lond Europe? Did you go back to Europe? Are you still Are you still where you were when I when we chatted earlier? It needs PubNub. It sure does, Lun. It needs pub cuz then you get to see real time loss. You get to see real time loss.
All right, let's do let's close. Reopen.
See? Reload. Double check. Okay. See, doesn't that look good? Oh, that looks so good. M. Look at that. Oh, that's neat. That is very satisfying. It is.
So, let let's see if we can run that for longer. And maybe we can increase the learning rate. Can we make it bigger? I want more learning. Let's go. Learning rate. Let's go five. See if we can make that bigger. So, what is our current Our current high score is or current low score will be 3.0. All right. Let's see if we can get below three.
Let's try below three with the same number of epochs.
Oh, I don't know. I don't know. It doesn't look like that. That did not seem happy to me. Nice save. Almost docks you. Yes. Back in Europe. Nice.
Well, not like a full on docs like uh Kevin is back from New York, right?
Exactly. All right. So, that one did not perform as well. So, let's go even lower. Let's drop that learning rate even lower. See if that does anything better.
No, not bad enough. It's very happy right here. It likes this right here.
It's very happy right here. This is its happy point right here. That is it's optimal right there. So, let's close that and restart it. Antimatter is basically backward matter. Exactly.
That's exactly backward matter. It's It's like It's the same as regular matter. It's just got like a different spin or something, right? Some sort of a different Just don't touch anything.
Exactly. Basic mirror matter. It's mirror matter. Oh, we did it. We got below two or below three. Sorry, we did below three. Nice. So, is that going to say below three? Is that correct? I think I need to restart. So, this is one of the thing I don't like about TensorBoard is I always have to restart it.
Yeah, there we go. See, now we restarted it. So, let's do even more.
Uh, let's see. Let's do 10,000 epoch.
All right, let's try that. Oh, wait. You know what? I wonder maybe what we could do is we could flush it every time.
So, let's try that. Let's flush every time.
What happens if you make the learning rate one? it uh then it's 100% and it will never learn anything other than the one thing that you taught it. Basically, it'll learn one thing really well and not anything else. That's what it will do. If you set the learning rate to one, that's basically 100%.
And it can only learn one thing. And if you try to teach it something else, it'll forget the other thing that ever existed and it will learn that new thing. It basically is last right wins.
If I had an AI interview and I think I absolutely flopped it, would you do you want to try answering some of the questions? Yes. Let's see if we got Let's see what our knowledge is. I think there might be some advanced math on there that I have no idea what the deal is. So, Kevin, let's do it.
Uh, opposite of electron spin. Yes. When it comes into contact with Yeah. So it's just storage, not AI. Yeah, exactly.
It's kind of like storage. It's it's it's essentially storage. It's a bunch of numbers that can represent some sort of abstract meaning. When you present it with input data or sensory data like visuals, audio, some pictures, or some some classes of some sort, it will give you an answer based on the features that you tell it about. So, it's got sight and whatever you want it to answer based on what you're giving it sight for, it's going to know. It's basically storage.
My cell is going to keep dropping. So, hey, got to drop. Yep. Dropping those cells. You got fiber internet, but it keeps going out. Oh, I know. When I got Google Fiber, I had Google Fiber. It was so fast. It was zero ping. You know what that means? Zero ping. It means less than 1 millisecond ping. It was so nice.
Everything was instantaneous. I miss it so much. That was when I was living in San Francisco.
You had a multi-layer perceptron MLP with linear activation function. How many layers are needed to learn a polomial H polomial one?
I think just one, right? Well, because it's a it's a a polomial can have waves, right? based on how many parameters you have and it's a two-dimensional line.
So, you're creating an equ you only need one layer, right? It's an MLP. So, is it an MLP with one layer? Like it's a support vector machine.
Let's see if I got it right.
I'm on a cell right now cuz internet is outage. Oh, yeah. That's what I do, too.
When the internet goes out, I grab the phone and turn on the internet tether.
Also, the format itself is queryable.
Uh, Kraken, wait, what? What? Did I miss your one of your messages? Wait, hold on a second. Kraken, I missed I The company I am with is being bought by Google Fiber. Oh, wait. No way. Really? I like my Google Fiber. Steven LB is the best or the coolest coder. Well, Zergio, thank you. We're we're we're trying. I want to have a really good skill. And it was one of my one of my things that I wanted to do many years ago. I was like I met like a super senior engineer many years ago now, more than 20. And I met them and I'm like I someday I want to know everything just like you. I feel like I got to that level. I feel like I achieved it. I got close enough. I know enough now to build whatever I want.
However, it doesn't matter so much anymore because the AI can take care of it for you.
Where can you find PubNub merch? Well, you know what? I have started. So, YouTube, I'm in the YouTube partner program and I was invited to the creator community. They said that we have a checklist of items to take care of, including merchandise. So, right, we got some merchandise here. Wait, is it going to show up? I don't know if you can see it right here. Like for example, I've got a shirt on and you will be able to buy that shirt at some point. You will. You will. I'll put it up there. I'll put it up there. So, I have opened up a store because it was one of our objectives.
They're like, "You need to have an objective." And part of it is you got to turn this button on. And when you turn it on, you got to take a screenshot and show us.
Why pointers exist in C++?
because you need to be able to access memory across different different contexts and you can use it to read and write to the same memory address. It's important because you can then update and pass by reference to data and share that reference in registers so that way you can very easily make mutations and things like that. Uh power went out last night. Oh, Kyle, really got to get that battery back up because there's Oh, Kevin Mallister in the interview. I forgot what a polomial was. Yeah, it's just it's just an equation with a bunch of uh parameters, right? It's then they add each other together, right? So, it's a parameter. So, a coefficient and a variable plus another coefficient, a variable, a coefficient and a variable.
That's polomial right there. And I think you can you can learn that polomial in a single layer. I think you could because it's basically the same thing.
Uh because it's linear activation function, you can't learn a polomial.
Oh. Oh, it's a linear Oh, it was a tricked question. Was that a trick question? It can't learn a polomial, it seems. So, we both I got it wrong then.
Multi-layer perception with linear activation function. How many layers are needed to learn a polomial? It was a trick question. You can't do it.
I I wasn't I wasn't I I I bypassed the linear part of it. I was thinking like a tanh. That's my favorite.
Let me read that. Let me reread that. In a multi-layer perception with linear activation.
All right. I still feel like it should be possible. You want to learn the basics of assembly to know how does a CPU work like in memory? Oh, well. Yeah.
Me too, Kevin. Oh man, I know how to deal with imposttor syndrome. Hey, don't worry. Yeah, yeah, yeah, yeah. Guess what? Guess what? You have AI now.
Everyone, it doesn't matter anymore.
Everyone's so confident cuz the AI is there to take care of everything for you. It'll build whatever you want it to build. You don't have to worry about imposter syndrome anymore. There are people So, there were this was really All right. So, what was it? There was a a YouTube short that someone shared the other day and they're like, "Hey, uh there was this very confident person that was coding with AI and they were certain that they knew what they needed to know." And you know what? They were wrong, but it didn't matter. It didn't matter. They were completely wrong. They had syntax breaks. Things were broken.
It wasn't working at all. And however, even though they were confident in their incorrectness with what they thought was the problem, it didn't matter because the AI fixed it for them anyway. But the semicolon, it was you, Lun. Yes, it was you. That's right.
Reminded me of the video yesterday. That was you, Lun. I Yeah, that was you.
You have AI now. Everyone is an imposttor. Exactly. The F4. You got it.
Uh, computational units from one layer become the facilitator of the next. Ah, okay. Wait, the YouTube creator program needs you to turn on a feature to show them by screenshot. Why can't they just get that from the back end themselves?
It's weird.
That is weird. That is very weird.
You're right.
The data is already in the database.
It's already there. It's already there.
You bring a really good point.
It's It's I They might just be I I don't know. I think my take on it is they're not going to go into the backend system.
They're like a layer on top and they're mostly for to help the creators do what they need to do and they're just training you and they have objectives and goals for you.
When you're an imposttor, it's not a syndrome. It's just an impostor.
Yeah. So, the F4 It is weird. I agree.
They already have access to the data.
So, why why do they I think part of it is like, for example, on my YouTube, I have thousands of videos. So, Steven Blum. So, if I scroll in here, you can see I've got 2,000 videos. And you have to link you have to link into which video it is. And they are not going to know. They're They won't know. But I did add the store.
It's currently currently in the works.
We're currently in the works on the store.
Are we coding or what? All right, GK.
Good point. All right, let's get back to coding, guys. Let's get back to coding.
All right, I want to Where are we here?
Uh, did we added the flush here? All right, let's try this.
Okay. So, if that works, the start tensor board because we're flushing it.
Oh. Oh, is it working now? Yes, it is.
Okay. There it goes. There it goes. It's now we're flushing it so we can see it in real time.
All right. So, that's uh 10,062 13 and it's still learning. You can see it's got plenty of units to learn. 1653 still 60. Oh, there's 1947. Okay, nice.
Stephen added a link or on cominatorial explosion.
Oh, nice. I believe AI software from what your thoughts. There are some there are some problems there.
Look at that, you guys. It's still going. Okay. Oh, wait. What is it currently at? Uh 163. Whoa. No way. All right. I think we've got a pretty strong model now. I think we got a pretty strong model.
I'm pretty satisfied with that cuz look at the rate of of adjustments. It's just learning like crazy.
Think uh AI models are going to get really expensive though. Well, we need more hardware to run them. What's an impostor anyway at this point? Even back in 2010, imposters were using other people's code. All right, we were copying and pasting off of Stack Overflow.
I remember that. That's I That's I did that all the time. Oh, is it going to go below one? Oh, it's going to go below one. Hey, we did it. All right. I'm very happy now. I'm very happy. I like that.
Uh, aren't original perceptrons analog?
Uh, and you had to do back prop by hand.
Yeah, I think so. Uh, maybe. See, I reloaded the wrong window. Oh, it's starting to it's starting to um smooth out here. It's starting the learn the learning rate's reducing a bit. It's reducing, but we did it. We got below we got below one. That's so great.
Hey, how's it going there, Northstar Alpha? Welcome on in. You were on uh Oh, and you thought everyone is AI. Are you AI? Northstar Alpha, we are regular human.
regular human over here. As far as you can tell. As far as you can tell. Okay.
So, you can see it's starting to to to to drop off there. Okay. That's perfectly fine because if it's below one, I'm happy now. No, I'm happy because it's below one.
Sophie, if I miss any of your chat messages, you guys just let me know. Uh we will easily jump back into it. Speak for yourself. I'm AI. Ah, Kyle's AI. All right. All right. So, that was 10,000.
Whoa. Hey, thank you for the G's or the GG's. Hey, Zergio, thank you for the GG's. I appreciate it. Thank you so much.
Let's see. La. Okay. Good, good, good.
Whoa. My gratitude. That was 20 jewels.
That's pretty good.
Questionable intelligence. Best. There's no name. No need for artificial intelligence. Yeah, you don't need to call it that.
Are we just extremely large neural networks? I thought so. That's what I thought, too.
I thought that, too. All right. So, we I'm going to say I'm going to hide that.
We're going to hide that. No, we're going to put this down down below. All right. So, we tried and we tried these.
They didn't work. So, I'm not going to give them a green check mark.
These didn't work. Um, and then we fixed the model. So, I'm going to count that as done. We did that. We fix the model cuz you see right here below one you guys right there highlighted on the screen negative log likelihood loss. All right, reduce image size. So I don't need do we need to do this? I don't think we need to bother with that cuz we've got a GPU accelerator. So that worked. So this is done. All right, this is going to be success. These are done. Done right there.
And then tensor board. Oh, we did that one right there.
Check mark.
We did that. Success.
Okay. See here. What does it finally look like? What do we look like? All right. That's pretty good. That's pretty good.
See here. Um, okay.
You think models becoming more expensive in the future is really resource distribution issue. Not enough clients approach for model accuracy. We're giving them too much power because it's easy. Oh yeah, right. If we can just throw more compute at it and then just crank it up to, you know, a billion trillion, 10 trillion, 100 trillion parameters and just waste all the compute because it's going to give us higher accuracy. We don't have much incentive, right? We're just going to consume 100% of all the resources.
You got it.
Nothing's real anyway. We had fake people before AI was a thing.
I have to say yes on that. You think your YouTube is 30 minutes late? Oh, is it? You really admire Steven's consistency in these tutorials? Oh, really? Nice. Yeah. Mo, I really, really love doing this. This is something that is like a very satisfying for me. I want to learn as much as I possibly can about PyTorch. Loon, you forgot to add no mistakes at the end of your prompt.
Yep, exactly. No mistakes. Perfect.
Works every time.
All right, let's keep coding, you guys.
Uh, wait, actually, before we jump back into coding, Kyle posted something over here. Uh, is this is this work classroom safe?
Navigating the AI combinatorial explosion.
It's probably safe. It's from Forbes.
All right, we'll read just a few first intro paragraphs here.
All right. Businesses world is experiencing cominatorial explosion as a surge of generative AI startups come to the market and it's only just begun.
Creators of the new technology say that businesses will soon be inundated with new products and services. now under development. We're going to have so many new capabilities, you guys. A lot of capabilities. Look at that graph. Way to go. I know. Lion's right. Wasn't it great? We did good. We did really good.
That's what you want to see. You want to see that good. It's so good. I love it. It's so satisfying.
We will look into internals of PyTorch.
Oh, we did. We did. Kevin Mallister, we went through the graph. We went through autograd. We went through uh the gradients. We went through the inputs and outputs and all the data stored when we do forward passes and when we do a backward pass to generate the gradients.
And it's all the stuff that we did when we wrote our own from scratch machine learning model framework. Can the graph reach zero? Yes, it can sort of mo uh it can it can reach zero though if it does that means you may have overfit or it's too it's it's basically it's perfectly perfectly memorized memorized the data you trained it on which is a problem. And that's that doesn't seem like it's a problem, right?
That seems like it should be okay. And it actually might be if we give it 100% of all known data on the planet and it has perfect recall from all that data.
So, it might be okay. Uh the problem is though, we can't do that. We don't have all the data. We don't have all data.
So, we weren't able to generalize with zero.
So that's the problem.
Grover's algorithm be on. Oh, don it, dude. We need a new way to distribute the model's power. Maybe if there was a P2P network for uh a model computation.
That would actually they they already do this. Well, let's go. Keep throwing more at it, right? We could. All right.
Here's here's what we'll do. We'll turn it up to 11 here. We'll we'll we'll bring it up even further. Where is it at here? Where? Uh learn. Let's go epoch.
We'll go we'll do we'll double it. We'll go to 20,000. We'll retrain it.
Let's see here. Is it re is it retraining?
Uh it's retraining here.
Yes. Okay, we're retraining now.
You don't need all the data to know what's on everyone's wall. If you know what's on a few people's wall. Yes.
Yes. That's a good point. If you That's why you need to generalize because you need to be able to not see everyone's wall and come into contact with data that you've never seen before and still get the right answer. That's the whole point of AI.
Yeah, I got that walnut walk. I figured I figured blockchain network for model computation. Everyone running everyone's model. I think there might be some of that. I think there might be double and give it the next epoch. Oo, yes. Double it. Even more epochs. So, I think I think we're doing okay with this. Uh, we found we found the ideal learning rate. We gave it more epochs and we are getting a good descent on our loss. Okay, let's go back to Forbes really quick.
In 18 months, the product space is going to look completely different because right now everything is shifting behind the scenes. That's what Aiden Gomez, one of the brains behind the transformer algorithm that lies at the heart of the generative AI models like GPT4 with a bunch of money of investment having already poured into dozens of AI startups this year and another 10 billion worth of ideas in the pipeline.
What's a seauite executive to do? Oh, we're on seuite. What do we need to do?
There's too many things. I don't need I don't have any idea.
Add me on LinkedIn. Lions. All right, let's do it. Oh, hey, Michael. Uh oh, sorry. Uh yeah, Mitch, is it Michael?
Yeah, Michael, right? Michael, thank you for becoming a member. Appreciate it.
Good to have you here. You joined the right channel for software engineering.
Did we add GPU support? We sure did. We sure did. Device right here. There's GPU GPU support right there. Loons, you're paying attention.
You're paying attention, Lun.
And now you're a member, Michael. Uh Mitchell. Mitchell or Michael?
Michael. I think Michael, right? I'm going to call Mike. M I C H A L.
Michael. Thank you for becoming a member. That's so cool. All right.
Look at that. It's just It's just going burr, you guys. It's going burr.
Learning rate go burr. Let's reload this real quick. Yeah, look at that. So nice.
So nice. So nice. Throw a dart at this new product map and hope you hit a winner or delay buying new products or services and risk being left behind by the bold competitors.
That is the risk we have these days you guys. Uh it's called Male. Oh, Mikail.
Oh, Mikail. All right. Male.
Mikail.
Male.
No, not Mikuel. All right. Well, you got to tell me. All right. The rapid evolution. Uh, it's Michael. Oh, like an I. Okay. So, we got it right. So, it's my Michael.
Michael. Like Mike. You nailed it.
Michael. Okay. Michael. We'll try our best to remember that. Really? We'll see.
Uh, let's see. All right.
All right. Sounds good. Sounds good.
Lion for lamps. Let's, you know what?
Let me Let me do a quick little check here. Let me open a new window, LinkedIn, and we're going to go to our network, and I will add you right now.
Let's see.
Uh, go to show all.
Can you tell me what your first name is?
Cuz I got a lot of people on here.
Got a lot of people here.
Uh, line for lamps. Let me know what your first name is. Got a lot of people.
All right.
Uh, you might already be gone. It's too late. Okay. The rapidly evolving te uh tech landscape, lack of standardization and complex ecosystem creates a paralysis of leaders navigating trade-offs, uncertainties and limited resources while considering the long-term implications of their choices.
This is what uh sesh er from North American chair of B CGX Boston Consulting Group technology built and design division. Okay. All right. So, we got to there's a lot to consider here.
Michael, you joined because you saw there was an intern opportunity. Yeah, I will announce it on the Discord. I will announce it on the Discord, you guys, when we're ready to go.
It's so nice of you. You always try to pronounce everyone's name right. I'm the loss. Yeah. Got to try Got to give it a try. Definitely got to try it out. All right. It's far from clear which AI applications are the most worth investment nor how successfully inte how to successfully integrate these technologies into their operations.
Business leaders are caught like deer in AI headlights overwhelmed by the growing number of AI solutions and their farreaching implications. There's too many you guys. There's too many. How many spots are open? One.
There's one. Oh, look at that. It reached the reached the end. Wow. Look at that. It learned, you guys. It learned. It learned. Look at that. It learned. And see, now now we're getting diminishing returns, which is exactly what you want to look for right there.
We did it, you guys. We did it. One one open spot. I sent you my summary and general Discord. Uh, okay. Let's check it out.
All right. Trying to hard-code every possible conversation creates a combinatorial explosion because human behavior has too many infinite variables to map perfectly. Yes. So, we're going to have an unlimited number of solutions because there's no way to capture them all. Although, it feels like there's certain solutions can fit into some sort of uh solution set of sorts, right? Some sort of solution set.
The system collapses into endless loop of rules and archetypes.
Which is exactly why caging the AI to a single strict utility is the only way to keep it functional and reliable. Strict utility. Interesting. Well, yeah, you got to in order to do that, you have to tell everyone that they can't do what they want to do, right? You have to say, "Oh, you want to use the AI for this?"
No, you can't do that. Sorry. You wanted to build something? Nope. You can't because otherwise we'll have combinatorial explanations.
They'll explode everywhere. Michael, would you take on another if it's not paid? No, we can't do that. It has to be paid by law. It's a legal requirement for an intern to join a company, at least in our country, in the United States, uh, and our states, you have to be paid. It has to be a paid internship, you guys.
So, so yeah, Kyle, I I guess one of the problems is you have to tell everyone no, they can't use AI because they will build something crazy.
Bizzle reputation of what Steven is meant mentioned starting his career as an unpaid intern. Oh, I was an unpaid intern though. Hey, I was because that was many years ago. They've made laws since then, you guys.
In the USA. Exactly. You can no long Exactly. play the Efron.
Let's take a look over here at subject of AI's idea real quick.
Super idea. Whoops.
It's your idea. It's your idea and you can build it. You can build it and make it however you want.
And you can make you can make anyone do whatever you want them to do. What's the craziest AI project you have been have you seen on Aszi?
H the craziest probably you probably one that you guys here on chat pasted in before which was the vibecoded operating system. You don't know what you're going to get. You're going to get something that works.
You're going to get something that works or more likely something that doesn't work.
Paying interns is a good thing. I think so, too.
Too many people took advantage of the unpaid interns and just fired them when they started asking for money. Uh-huh.
Uh-huh.
You want to move to Japan? Oh, hey, you want to move to Japan?
You That sounds pretty nice. Japan sounds like an amazing paradise to me because everyone is for the most part, the majority, not everyone. There are still people in the fringes in Japan.
Everything is tidy and clean and orderly. And I subscribe to that. Thank you. Thank you so much. I subscribe to that.
It needs P. You're right.
All right. You're saying the AI can't build what's right for humans because the combinations are too vast to determine what's right. Utopia will always be an individual's idea.
Oh, okay, Kyle. I get it now. You can never have a utopia because everyone else has a different idea of what that is.
Well, I suppose Japan is the closest thing that we currently have to that, right? Or because it's the the laws aren't too strict, so humans still have freedom and everyone is tidy and cordial. So, I think Japan's the closest to the utopian society that we got right now, unless uh there's something that I don't know about. If you're going to go through uh focus on learning the subtle cult uh cultural things so that you can become Japaneseish Japanese.
You live there for a while. Okay. So, it's still there's it's still not a utopian. Oh, don't don't get me don't get me wrong. It's still not a utopia.
It's still not.
Can we have the pub song, please? Okay, Lun, you got it.
PubM song on the way for you. It's tidy.
There are a ton of issues, though. Okay, got it. Yeah, that makes sense, Michael.
That makes sense. All right, so did we achieve our objective in our code?
I think we might have achieved our objective.
>> All right, let me do let me get status.
Get add get status again. Get commit.
All right, so success below one score. We did it.
>> Party popper.
>> Did I spell success right?
>> Below 1.0.
All right. Push to main. Boom. Japan is the it's the people that make it that South Korea is trying. It's all about anime. I know, right?
>> Determining a good human experience might be the very thing that limits.
It's a paradox.
>> Yes, it's the thing because we all have a different idea. And if we want to achieve paradise, everyone's idea of paradise is different. For example, I want a bunch of anime in my paradise.
While a lot of people don't care about anime, right?
If you want to go there, you can easily go live. Go live there. Go live there for half the year. Rebellion aunt, welcome on in. Good to see. Welcome on back. Happy to have you here. Good to have you here.
>> Okay, let me see here. Looking good.
Okay, so I think we achieved our objective. Did we Did we do it? Did we win? Let's see if we can finish off the tutorial.
>> So, LA did that. We don't need to worry about these other things. Uh, we could do a one hot. Let's do a one hot really quick. I want to do a one hot really quick on the labels and then we can move over to finishing up the tutorial, which would be here. So, we achieved the scaler, which is exactly what I wanted.
It looks really good. And then we're going to go on to visualizing the model, which apparently we can do. Look at how in neat this is. We can visualize the model, you guys.
>> Request sent. Oh, lions for lambs. Oh, wait. Okay. Um, one second.
>> I found you. I found you, Nicholas.
>> All right, we're connected now.
>> All right, here we go.
>> Thank you for connecting. Exclamation point. We're connected now. Lions for lambs. We did it.
>> Oh, Bonzupi, your assistant. Stop saying that. Uh, request sent. It triggers you, right? It triggers you.
>> Don't hire another assisted.
>> PubNub is an internet 2-way radio.
Birectional JSON radio. Isn't that great, Lun? I love it. 100% think we should be going into a genius develop some AI enabled human routine. They might even sell it as a business model.
I think we're going to see a few of those. I think we'll see a few of those.
One can be in a country of 180 days total. Only 90 at a time. They are very strict.
Okay. You can only rent apartment for 6 months.
It's time. It's time. It's >> it's the lyrics >> where metrics like happiness will be how they sell it. The metrics of happiness.
It's so simple. It's so simple. It was right in front of us all along. Kyle, you found the ultimate metric.
You like anime? I do. Michael, I'm watching I Was Reincarnated as a Slime right now. And the uh what do you call it? Um Ephren. No, wait. Free Freren.
I'm watching Ferrer.
All right.
I'm trying No, that's too that's too intense.
Is this fine?
We'll go with this. We'll go with this for now. I'm I was I was looking at different music to to to listen to here.
Uh reminds me of the uh the Doctor Who episode. I like the Doctor Who where they sell emotions. Oh, yeah. And it was like 11th Doctor, you think? I remember that episode, too. There was a stand and then there's like these little uh circles, right? There's these circles and I forgot. I think you just you just Was it like a you put it in your mouth?
Look at Link Share. All right, Lun.
My keyboard knows the lyrics. No way.
No way.
That's crazy.
Are you serious, loons?
That is I wonder. Yep. Someone's done some training on that. Oh, I love that.
That's great.
That's crazy, Lun. What do you know?
Global companies, but competition is really hard. Competition is rough. It is. Now that everyone is competing, competition is vast, you guys.
The emotion stickers. Oh yeah, better and better Dan here. Oh, how's it going?
Better and better Dan. Good to see you.
Especially hard if you are also already in the relative niche. Security architecture entries are rather low. So, you know, we can get into some security, you guys. We could do we could do it. We can get into some security here.
Oh, another link share.
Did I do that? Link share. We did it.
It was amazing. Okay, so let's go back into this. I want to do a one hot really quick if we can.
Let's see if we can do that. So, the one hot classes.
So, I wonder if we can.
Is that going to work? Maybe.
Maybe. Yes, I think it will.
All right. Where is this at? So, I need the transform and then I there's another there's another feature here called transform.
Uh something else we need we need to find it. It was around here somewhere.
gp-- r transform should be in here somewhere.
It was amazing.
So that's why uh that's so that's why the way I use AI in your software is utilities focused.
Oh, so we're getting we're getting sort of an insight into how you use the AI really trans. Did I spell that wrong?
Right here.
grap transform star.
Oh, it might be somewhere else. It might be in a completely different directory.
Okay, so we need to look this up and then rewrite it.
So, it was called transform_targets.
Let's see. I need to I need to look this up. How do we set We need We need We need help here. All right. PieTorch.
How to set Lambda transform on targets when importing data from the torch vision data set.
Okay.
Donut, dude.
You were about to ask if we have to delete that. No, no, you're good. You're fine. Oh. Oh, got it. Got you. Uh, I I missed it. How's it going there, Christian? I saw you posted a message that got moderated there.
All right. So, I need an answer really quick about the docs and then we're going to use that. We're going to use that right here. Target transform. All right.
That's what it is.
target transform. And I don't think we're using this right now. So, I'm just going to I'm just going to hide that for a smidge because we're not we're not using that.
And does it have an example? Okay.
Transform.lambda.
And then we can This actually might This is exactly what we're looking for. Um, let me I'm not going to copy and paste it. I'm going to try to type it. So I need a torch.z0ero number of classes. Okay, so this is going to equal a trans. This will be target transform target transform equals transform.compose compose and we're going to do a two tensor and then we're going to do a lambda that will be a transform dot lambda lambda with capital capital and then we pass it in uh an actual lambda.
Okay.
And is it one value? Yeah, it's one value. Okay. For the targets, you're probably overthinking it, Stephen, but yeah, I do that with everything. It seems like you've got something that works. That's what I can at least tell you on that. From my perspective, the only thing I do is just ask it to do something and it does it. That's all.
That's all I do.
Lu, thank you for clicking the hi button then again.
You are the number one. Hello button clicker. You are the number one.
Oh, let's see here. Um, sh link share YouTube short classroom. Okay.
Okay. Anyone can recommend a content course for vector arithmetics.
Array arithmetics for vector vector math. You want to do vector math. I mean, I just say do do arrays in Python or something, right? Do some uh map reduce in Python. That should do the trick. A little bit of math reduce there.
Let me see here.
I had to get the music a little bit better. YouTube is great. Not for short live format. Oh, yeah. I It depends.
Like YouTube's getting pretty good.
Like, it really depends. What you need to do is give the algorithm enough of your view time so it learns what you like. It keeps bugging. Oh, it does. No way. Really? Oh, for embeddings. Okay.
Um I don't know what the best is. Oh, three blue one brown. Bonzupi. Yes. Okay.
Bonzupi, thank you. All right. What Bonzupi just recommended is your go-to.
You want three blue, one brown. Has a course on calculus.
Multiple videos on vector math and machine learning. Nice. Do you remember when you said you were working on documentation tool for the game tower networking? You finished enough of it and make it useful and you put a link at link share. That sounds good. Therefore, and yes, do that.
Torch.
See, comma. And this will be number of classes. Where is it? It's just called classes.
And did I define that?
I defined that down here. All right. So, we have to define that up higher.
We'll define it up here. Classes, comma, uh, what was that? Uh, oh, dype, right?
DT type equals torch dot float 32 maybe and there is a scatter. Is there another way we can do that though? I think we can do that differently, you know.
It has graphs and stuff. Ah, yeah, it does. It's a web app. Pretty nifty. If I do say you so yourself. All right, let's take a look.
Tower network manager.
Oh, is this like Anible tower?
Um, seed seed demo. All right. Data layers.
I don't know how to do anything.
It doesn't see anything.
I can't get it to do anything.
The instructions are under G plus minus arrows.
Yeah, I see him. But nothing's happening in the link share instructions. Oh, okay.
Uh, save test. Load raw test.
I'm clicking every button.
Uh int backwards v1 exclamation in backwards uh t n i4 I user error user error I All right I have to give up sorry I give up too uh hit tilda for command tilda.
I'm pressing it. You can see. You can see I'm pressing it. It's not working.
Paste the example into the import.
What ex? There's no example.
Oh, like this. Okay. Copy project. Import.
import. There we go. Okay.
All right. Now we did it. Now we did it.
Okay. See? Very nice. This is actually really smooth. By the way, this once I once I got it working, all this other stuff like behind the scenes, like I'd never know how to do this. I would I would never figure it out. However, it is pretty neat.
It's clean.
I like this. Very nice.
What's that photo opened? Oh, that was uh I'll share that really quick here.
Let me reload this.
See, here we go. So, someone who used PubNub uh and their team built a electric car that is I believe um AI AIdriven. So, it was autonomous, right? It was autonomous. And this team was successful at building a car only with a $10,000 grant. and they beat a $100,000 team and they came in first. The team the team that had less funding beat because uh I think maybe because they were using PUB they had a really smart idea. They figured it out. Exactly. That's cool.
Autonomous race car. Yeah.
Basically, it's a networking documentation. Oh, got it. Okay, that's a little more clear now.
See here. Um, what are each of these things? Network address. It's all like abstracted. Oh, I see ports here. Okay, I get that part. I get the ports.
I don't see any addresses or cers.
Oo, I like that. I like how it's got the arrow.
Nice.
Oh, there's some nodes out here in the middle of nowhere.
Okay.
F4 for very nice. Tower networking manager.
Very nice. Okay. After it took us a little while, but we figured it out. We took us a little while. We figured it out.
It's pub for sure loons. Yeah, I thought so. Uh may I was thinking I was thinking maybe. Yeah. So, that was uh that was a pretty neat story. That was a pretty neat story. You know what? We had that story for a while. It just took us a while to get it posted.
Oh, DNS. Okay, got it. Okay, so it's a it's a it's an alias.
If you change the data layers, it shows stuff.
Oh, look at that.
Oo. All right. Now it's like getting way too complex. Like what's going on here?
Oo. It looks pretty though.
RJ45.
All right.
How come it's all the same port? We get 39 39915. is all 3 915. They're all 3 9115. Every single one is 3 915.
The dots are the ports. They're all the same though.
Open source. You mean uh the race the race car the autonomous car I believe is closed source because they're planning on entering again this year and they've got a next level that they're going to do. They're going to make it even better. All the ports are on the device 399. Okay, got it. All right.
Very neat. The F4. I enjoyed this. It took us a few seconds to figure out how to deal with it, but after we did, I uh got at least some visuals happening here.
Interesting.
Interesting.
What's this? Network address F2R2. All right. Very neat. Uh what's the difference?
Source available versus uh open source.
Oh, yeah. Yeah. Source available versus open source. One's license and one is the model. So you can have a license that says you can use our source code, but you can't modify it and you can't you can't you can you you can read it. You can see it, but you cannot modify it and re-release it. Right? So that's the difference. One is a license.
One is allows you to do whatever you want with the code. And the other is no, you can't use it. You can see it, but you can't use it. You can see, but you can't touch.
All right, let's add our graph in here.
Oh, wait. We were going to do a one hot.
All right. Uh, I did I have a better Did I have a better scatter? I suppose we could just do the scatter.
Lambda torch torch zero. Yep. Torch zeros is what we're going to do here.
Scatter underscore.
And we want the number. Let's see. We need it to be Y.
What did they do? A torch sensor Y.
Uh, and then we needed value equals 1.
Is that it right there?
I'm I'm I'm walking through this here real quick. Uh and then you know, okay, I'm just thinking out loud. uh will basically be mapping human experiences for others. As software engineer, you enable that you will be successful probably regardless of how and why. So when you build just keep that in mind.
I kind of didn't catch that. I I read it but I didn't understand it.
Uh okay. Thank you for the explanation.
Thinking out loud. Uh, currently my source is all rights reserved because you have not put a license. Oh, you got to put the license in there. Source available means that source code is available but does not necessarily mean that you can modify or redistribute it.
Yes, exactly, Bonsupi.
Some company will just copy copying it and putting into their tool without paying me. I know, right? You got to be careful of that. So, probably going to pick one of those ones that requires publishing your changes, but you can do whatever. Mhm.
I I always do MIT or Apache because that is the most generous. It basically says you can do whatever you want, but you can't blame us if something goes wrong.
It's like a disclaimer.
All right. Uh, so this is target transform. Let's see if it Let's see if I got that correct.
equals target transform. I'm just going to comment that out for now.
Okay, let's see if that runs.
I have an error in here.
Why is that a problem? Is that on the right line? Line 26. Yeah, right here.
Seems good to me. Oh, no. We have to do we have to do this.
Here we go. All right. There we go.
Compose object has no attribute. Yeah, it does.
Oh, it's got to be transforms. All right, fine.
And do we does it always have to be transforms?
There we go. All right. On our way.
Let's see here. Uh, basically, you make the interaction flow engaging and good enough, it won't be that relevant what your actual product is. Oh, yeah. If you just bake in a whole bunch of happiness, I didn't even think of that. Just make it doesn't matter if it does anything as long as it makes people happy. I want to see this live, but it's 3:00 a.m. Oh, it's 3:00 a.m. Guru Taha. Hey, well, don't worry. It'll be recorded. You can you can watch it after you do the MIT also with attempts to convert it to MIT and close source them in companies. Right now, be it does not feel good. So, you have some you have to you have to do something a little extra. Okay.
All your code on GitLab is CC4 Creative Commons. Oh, I never even knew that was a All right. I just do MIT. You have credit if you modify it. Oh, you have to credit. You got it. The MIT does that, too. You still have to like add credits.
Okay. So, let's add this in here. I wonder if that will do the trick.
Target transform.
Okay. So, that's target output. Let's see if our criterion will automatically ingest this. Nope, it didn't. OH. OH, RIGHT, RIGHT, RIGHT. OKAY, so we we've got a few things. I I think we'll be fine here. See here. One second.
Where's the problem at?
Oh, what? Oh, no. Really? On that?
Really? You're all right.
It's just It's just disappointing.
All right. Data iterator. Um, it should work fine. Why does that not work?
H.
So, it's right here is a problem.
So, what if we did break and we said data?
Is that going to work?
Pick should be a pill image or an ind class.
I'm stuck now.
This doesn't make any sense. It should be able to iterate over it perfectly fine. Where's the deal?
19% battery. Oh no, you're running low.
Got to be careful. Well, do you have like a a power cable? Maybe like Okay, so it doesn't like it. It does not like it right here. It's just unsatisfied with this.
Okay.
Uh batch a batch, data equals next data iterator.
Uh wait wait wait wait. Just say data.
Uh, this doesn't make any sense to me.
Image target H.
I mean, I get it. I see what's going on.
It's just that I don't know how to deal with it.
There are large variety of c creative common licenses specifically that prohibits commercial use like you were looking for. Oh, Bonubi, that's a good idea. You could uh CC open available for everyone. However, if they plan on profiting from it, you can say no. You can't profit on my technology unless you pay me money. You can pay me for a different license.
CC byNC4.0. Oh, nice.
This should work perfectly fine. All I'm doing is just transforming the matrix right here. It's not It's not a problem.
It's just transforming the matrix.
All right.
What if we just said tensor?
We don't do one hot.
Still, it's a problem. Are you serious?
All we're doing is making it tenser.
H I mean this looks perfectly fine to me.
So that's going to work.
Ooh, battery degradment.
Careful to charge it between 20 and 80%.
So yeah, exactly. So that way you keep it in its peak capacity.
GPL on the other hand, oh yeah, disallow for commercial use. But a whole lot of other cases of GPL, it's practically radioactive. The companies, yeah, we don't use any GPL. We don't, we intentionally were like, is the license good? We validate every single thing.
Like if there's a license that's a GPL, it is essentially what you it it infects all the other software that you are operating on including your own software that you wrote. So if you want to include that library. So we don't do any GPL. No GPL for us you guys.
This would be nice to be able to get this working.
Uh should be fine.
H pick should be a pill image or indie ray. Got int class.
All right.
I'm not seeing anything obvious.
Yes, exactly. That's what I want. I want to do that right there. And it's really simple. It's really simple.
You're not trying to save battery. Oh, the batteries. It's too late. Your phone is heating so much. Waiting for it to break to change it. Not now, but later.
Okay, loons.
Are there any that prohibit feeding my code into an AI to make a spec for the purpose of making a Chinese firewall port with a different license? So, uh, that foreign what companies are doing, even if you release a license, they are taking that open source and they're having an AI rewrite it and reimplement it.
And apparently that bypasses the law that is beholden to whoever captured the code with the original license. They're stealing everything. Yeah, exactly.
They're stealing everything. We got to be careful, you guys.
They're feeding code in AI to make it extremely detailed spec and then tell another session to make the spec in Python and then it makes almost exactly that.
Uh-huh. Exactly.
It's what happened to IBM back in the day. Oh, it did that. Really? Oh, that's unfortunate. That's the world we live in, you guys.
All right. I really want to here.
I'm going to look around.
One second.
I want to look around to see if we can get this here. Uh la activations maybe.
Yes. No. Maybe.
So let's try this. Let's try this tutorial. Okay. CD desktop pietorrch.
All right. Grep r uh gp gp for target underscore.
There we go. There we go.
Data loading tutorial.
This is what I want. This is what I want right here.
Uh, let's see. Target.
Do we have a lambda?
No. Okay. Let's script for lambda.
So, I was doing this the whole time.
right here. Okay. So, custom made it by custom.
All right, we're going to try cop. So, one hot.
Oh, right. Let's try this.
Okay.
Okay, that's what we were looking for.
Okay, good to go now. Good to go now.
Maybe we're going to find out.
See if that works.
Uh, yep. Lambda. It's cuz we have to put it over here.
Okay, that should do the trick.
Oh, good. Oh, we did it. Okay, so what was the difference?
Uh, classes What's the difference between these two?
Well, one's a compos. So, that's part of the problem. So, let's just look at this though real quick.
So, let's compare let's compare these two here.
Okay. Uh torch zero torch zeros classes dtype float scatter zero yh well seems is it's not it's not that much different.
Okay.
Okay.
So, we got the data. Where's our enumerate at? We got to go back to our enumerate here.
Yeah, that's what I'm looking for. Okay, good, good, good.
Now, let's see if we need to transform.
Nice. Okay, I think we did it. I think we did. We achieved it. We achieved the goal. Perfect. Perfect. Perfect. Check mark. Done. Let's move these things under the done section.
Perfect. Okay.
Success. Success. I'm very happy with that. I'm very happy with that. So, did we do all the things we wanted to do?
We've got one last thing, you guys. One last thing to do today. Did you put out the fire on the cabins? Uh, hey Scared.
How's it going? Scared or werewolves?
Welcome on in. Happy Tuesday. We're building convolutional neuronet networks is what we're doing.
And then I want to add a writer dot.
We're going to learn something new here.
We're going to learn something new.
We're going to visualize the model. Net and images. Net and images. Huh. Add a graph will trace the sample input to your model if we run it once through.
Oh, add graph net images. Okay.
All right, let's try it.
All right, add graph. Oh, graph. Okay, so we're going to do this.
So, this will draw our model graph. So, it should it should basically render the whole the whole deal. All right. So I want it's called model, images.
We need to grab our data iter here. And we're going to say uh images, labels equals next data iterator.
And that's it right there. I think we just did it.
Uh, loons, I'm I'm uh Phoenix wanted their BIOS to work on IBM computers. An engineer just looked at the IBM code and described it to another engineer and they copied it. They copied it.
Okay, let's So, let's do fewer than uh 2,000 epochs.
20,000. All right, let's do just just Yeah, keep it that. All right, so now in about a minute, we should see our graph. Although, I suppose we could see it now. There's no point. We could have done that in the early part. H Let's just wait a smidgen.
Okay, visualizing your model.
Uh, I hope it works. It might just crashed.
Let's restart our tensorboard.
Okay.
How's it going there, Vinnie? Bitty vichi. Oh, Vinnie Bitty. How's it going there? I remember that from uh from Doug. Is it possible to go from an idea to a product or a SAS tool with only free tier AI models? Yes, it is. I think so. If you use the GIMA 4 AI model with the open uh the sort of like the open what is that, you know, like the claw code and the Gemini and the codeex. There's an open source version of that. What is that called? What is open source version of clawed code? It's uh open open code. THERE YOU GO. IT'S OPEN CODE.
So use open code with GMA 4 AI model and you should be do you should succeed pretty well. I think it's not state-of-the-art open code. Exactly.
Exactly. Loans.
So that's what I recommend. Use open code plus GMA 4. Open code plus GMA 4.
And then you'll get as much freedom as you want.
Has some good models for free, too. Oh, nice. So, Open Code comes built in with a bunch of free models. Here you go.
All right. Did it crash? It crashed. No, we didn't do it right. Oh, we forgot to convert it over. Ah, darn it. Okay.
Okay. Okay. Fine. Fine. We'll start over. Did the Did did it capture any of the loss though? Okay, it did. Okay.
All right. So, we have to do two device.
I think that's it right there. I think that was the only thing that it was waiting for. All right. Epoch.
We'll say it's 200. 10 times faster.
Neotron and GLM5.5.
Really good models. Hey, it didn't crash that time. All right.
Where's our Where's our output, though?
Where's our output?
No, no, no. Where's our output, guys?
We We put it right here. Right here.
Look. Oh, we need a flush, I guess.
All right. Apparently, we got to flush it. Let's try it again. And then we will restart Tensorboard.
Vinnie Vidy Vichi, the thing is you've made some tools with free AI models. You've even posted Trading Indicator about 2 months ago and got about 113 downloads so far. Hey, very nice. But you're starting to second guessess everything.
That's actually good because you're starting to realize you can make it better.
You can make it better. You're on the right track. Don't worry about second-guessing yourself. It'll be a guarantee that you do that.
Where is our model? Where is it?
We flushed it out. We added the graph, you guys. Where is it? Did we do it right?
Okay, it's it's updating. It's all there. It's just not showing us the graph. You got to go, Lon. Sorry. No worries. It was good. You stayed for a long time. You started You got here at the beginning and don't worry. We're We're wrapping up now. We're wrapping up. You stayed here for the whole stream, basically. That was awesome, Lun. Thank you.
See you. See you. See you tomorrow. All right. Uh, I want to I want to solve this one thing.
Free tier models. You're not sure about how good it is. It's pretty good. Torva, you too. You've been here. You guys are so great. It's so much fun to be with you guys here. Okay, so I want to figure out why this is not working.
Net images instead of model. Well, so my in the tutorial they called you. You're keeping your eye on this.
You you got good you're good Zergio. So net for them is the model. So if I scroll up here, this is their model. So they called it net. I called my model model, right?
Here's my model definition right here. I just called it model. I didn't call it net. I just called it model.
Thank you. Yeah. Yeah, you're keeping your eye on it. You're right. So this right here, I want it to work. I want it to work. I don't see why it's not though. Let's see. So, we iterate. We got a new iterator. We grabbed one. Oh, images labels. Did I do that wrong? No, I got it right. Images, commas, labels.
Add graph will trace the sample input through the model, which is I what I wanted to do. And render as a graph. It sure didn't though.
What if we don't do any of that here?
Let's let's let's only have nothing.
Okay. Thank you for the hearts, you guys.
The F word. It was good to see you, too.
We'll see you next time.
Okay.
No, there's just nothing. There's nothing there.
Scalers, images, histograms, all.
There's nothing. Did it break? time series scalers custom images audio.
That looks neat and histograms. No.
Huh.
Well, apparently this doesn't do anything.
We We copied the code perfectly.
Hey Stephen, do you know of EA Nazier?
I don't know. I don't I don't think so.
Visualizing your data set with embeddings.
Select LA. Select random class labels.
Feature images.
That actually would be pretty neat to try this. Wouldn't the more interesting stuff deal with reinforcement learning and like how tokens are clustered something?
So that's where you get into the finetuning. So what you do is you have you're talking about the initial so the initial training is unsupervised and it just learns patterns over the repeated data set and then you want to fine-tune it with reinforcement learning to give it objectives and you you have the model run until it meets those objectives and you try to have it run on fewer and fewer amounts of iterations to be successful so that way it can get to the answer more quickly. quickly with less time and less compute. So you basically give it a goal. So Kyle, I that is in fact pretty exciting.
Some scammy merchant selling super subpar copper.
What?
Immortalized in uh what is that word? Some sort of tablets. I don't know what that is.
He may have sold bad copper and treated your messengers with contempt, but you've heard of him. Oh, really? Okay.
First customer service complaint in record history. Oh, wait. No way. Is this a live Python stream? It sure is.
Hey, how's it going there, Big? Welcome on in. Good to have you here. We are going through PyTorch and we're currently on our very at the very bottom here of this Tensorboard tutorial. I've never used TensorBoard before in this tutorial and it's okay. It's fine.
However, there's this part right here which is so simple. It's not working. It doesn't do anything. It's empty. It's completely empty. And it's not like it wasn't working before. Cuz we have we have stuff that works.
It works. Oh, it works. Heck yeah, dude.
Nice.
Mind taking a look at my projects when you can? I'd love to know what you think. Sure, Vin. Vinichi. Absolutely.
Vine bitty beachy. You got it. So, see you can see our loss which shows up really nice, right? You can see that pretty nice right here. Looks good to me. I like it. The problem is that's the only thing it shows. It doesn't show anything else.
And I want to see other things.
It's not showing. It's just not It's not showing you guys. I really wanted to see it. I wanted to see what that looks like.
Oh well. When you switch over to Tensorboard, you should see graphs tab.
Oh, there's a graphs tab. I just got to keep reading it. Graph. Oh, there it is.
Hey.
Okay. We just got to read the documentation. All right.
Is that the learning rate? The learning rate is 0.001. 001 I think. Right there you go. Learning rate, we found it. All you got to do is read the ducks. Okay.
Although that's not that impressive.
It's It honestly It's not that impressive.
We found it. It took us a while, but we found it.
Oh, that's kind of neat. So, you can see where the device is. Trace.
Okay. Our graph is too simple.
Apparently, our graph is too simple, you guys.
Well, in that, I do believe that we essentially achieved our objective for today. Let me see if there's anything left. So, there's this one last thing here.
that I'm just going to copy and paste because I think I achieved everything I wanted with tensorboard.
Let's see.
All right, we're going to comment these things out here.
We're going to leave that. We're going to paste it in. Set. Paste.
And we're going to see what happens.
Do we have some of these things? I don't know if we have training set. We do.
Okay, we have this uh looks good. Looks good. Okay, images.
Looks good. Yep. Okay. All right. We should be able to see whatever this is.
I don't know what this is. Let's read what this is. And this will be the last thing. The 28x 28 image tiles that we're using can be modeled as a 784 dimensional vector. It can be instructive to project this to the lower dimensional representation to project it as a set of data in three dimensions with highest variance and display them as an interactive 3D chart.
No way. Method does this automatically by projecting the 3D dimensions to the highest variance.
If you switch the tensorboard to the projector tab, we should see it.
although has no object training set data. Okay.
So, what I'm going to do instead of doing random is I'm just going to say next.
I don't care about this.
Let's try that.
They're overwhelmingly trained on English. Ooh, training set. Yeah, that should work.
That's perfectly fine. I'm using that right here. Oh, here. I've got a better idea. All right.
Have you used dimensionality reduction?
Yes, absolutely. Label uh images.
What? What? What? What's What's weird here?
Images, labels, class labels, class labels, label, label for label, and labels. It's right there. Labels and labels.
Classes. Oh, I don't have a class. Do I have classes? All right, hold on.
Oh, got it. Okay, we don't have that because it's we're not using we're using animal pictures.
classes labels. What I'm going to do is I'm just going to make that a number.
Okay. Like this. Watch this.
Uh that should work. I don't know.
Maybe.
Okay. Um shape 784 is invalid for input size of very big.
Oh, right here. Okay.
So, we need to convert that I believe.
Do we have another view? 28 by 28. Uh let's try this.
See, this gets close. At least it'll get us a little bit closer. Also, epox.
It's going to be like 20.
It's still not big enough because the image is massive, by the way. It's massive.
All right. It's going to be huge image.
So, I've got an idea.
There we go.
Easy easy mode. Maybe label should be equal with data points.
What line? Where are we at here? Add embedding features. Oh, okay. Writer to add embedding features.
I got a very big features.
You just don't think AI is a sci-fi utopian tech some people think it is.
Yeah, we're not there yet. You're right.
You're We're not there yet. Maybe when we get normalized AI robots. It'll be different when we get the robots, you guys. Once we get the robots. Okay. So, I'm on this very last part here of this final bit of the tutorial. It'll feel so good to finish it. It'll feel so good to finish this. Uh, I just need to figure this out. All right, so the features are too big.
Labels should equal with data points.
That doesn't make any sense to me.
Labels should equal with data points.
Oh h images should equal with data points.
Yeah, I don't get it. I don't I don't get it here. What we're going to do though is I'm It's because they want fashion emnest. So I'm just going to grab this and we're going to overwrite it.
Where is our training set? Here we go.
Okay. So, we can restore the view.
It is 128 28. Okay. 28x 28, which is 784.
Okay.
Uh is 784 invalid for input size of 90.
What are you talking about?
Here we go. All right, I figured it out.
It's just a little bit a little bit at a time, you guys. A little bit at a time.
Labels. Oh, it's right there. It's right there. It's right there.
Int object is not iterable.
It is a label. Oh, it's a label.
Uh, all right. Let's just comment that out for a second. Don't care about that. And this, it's not a label. It's a label.
So, it's still not big enough. Okay.
Oh, did we not trans Oh, so we need to define a transform up here. One second.
Our transform's different now because we changed it. So, let's wrap this little bit of it up where it is. Here's our transform. Okay, we go. Looks good to me.
Okay, now it should be good. How's it going there, Tatana? Good to see you.
Welcome on in. Hey, it worked. It worked. You joined at the right time, Tatana, check this out. Okay, so we should be able to see or tensorboard graphs.
Uh yeah, time series scalers model training loss. Yeah. So what what which one are we supposed to look at again? Where was that at? All right.
This is the last thing on the on on the tutorial. This is the last thing you guys. We got it to work here. So now if you switch to the tension board projector tab. All right. So we need a projector tab.
Projector. It doesn't. There isn't one.
I don't see a projector tab.
Oh man. Oh man. Really? There's no projector tab here. Let's Let's rerun it.
Restart tensorboard.
Reload. Okay. Thank you for the reactions, you guys. Okay. Where is our projector? Project. There's no projector.
Custom scalers, images, there's no audio, no graphs, maybe distribution. No debugger. That looks neat, but that's not it. Time series scalers.
Did we flush and close? Yeah, we flush and closed it. Maybe we need this part here.
There we go.
We're going to try that.
So, that should work.
Okay.
Projector. There it is.
Well, it doesn't look like anything, but we got it.
We got it. We got the projector tab. Can we say batches batch size equals 8? So this will actually be labels now.
Try that.
Mnest got an unexpected argument. Batch size. Oh, it won't let me do batch size here.
Oh, that's on our data loader. Okay, one second.
Where's our data loader?
Uh oh. Uh, it's training set has been replaced. And we've got an iter. Oh, all right. So, we need a data loader. Data loader loader. Uh, where is that at? Here.
There we go.
So, let's use this instead. All right.
Data loader. And this is going to use training set. Yep. batch size of eight.
Okay. And then we don't need Do we need this? Hold on.
Yes.
We're going to go off of this.
And that should get us a batch of eight, which should give us more data. Okay.
How's it going there, Muggy? Good to see you. Welcome on in. Thank you for clicking the high button. Training loader. Did we get this wrong?
Yes, we did.
No. Wait. Where are we supposed to use this at?
Here. That should do the trick. Okay, let's try that.
I really didn't like that.
H.
Okay.
Yeah, cuz there's only one.
All right. Well, I'm going to call that good enough for now. So, we got our tab.
It shows up. Let's go to uh and we see a very pretty nothing.
We see a very pretty nothing. We copied the code in. And also I kind of don't care about this one. I kind of don't care. I don't need to visualize that.
All right. Um I'm going to commit this. Get status.
Get commit. Am we sort of added projections and push that to master. Okay. All right. That's good enough for now. Hey everyone, thank you for joining in for today. Tomorrow we'll be back and we'll be doing some more PieTorch. We finally got past this. I'm going to call I'm going to call that done cuz we've been on TensorBoard for so long. We've been on Tensorboard for so long. Now we're going to go into this next segment and this is where we'll start with tomorrow.
I sent you an email, sir. All right, Benvich, sounds good. What are we creating? Uh, Chana, we are trying to become PyTorch experts. So, we're walking through all of the tutorials.
The last tutorial that we were just on was a convolutional neural network that we messed around with a lot. We went way off of where the tutorial was and we kind of built it our own. We used animal pictures instead of clothing pictures which was a nice change and then we changed how we transformed the data through our connet.
So, the plan is tomorrow we're going to go and continue on with our journey of becoming PyTorch experts and we're going to move on with this next tutorial.
Sounds good, Bonzupi. Thank you everyone for joining in. We're going to be back tomorrow. See you loons. See you Bonzupi. See you Kyle. See you. See you Maggi. Good to see you guys. Zergio, good to see you. Have you answered the rest of your day? Oh, have an awesome the rest of your day as well. See you guys. Have a good one. We need PieTorch expert badges. Yes, we do. Yeah, we do.
Good plan. Good plan, Kyle. So good to see you, Kyle. All right, we're heading on out. We'll be back tomorrow. Bye, everybody. Have a Oh, uh, what was it?
Noel. Noel, thank you for subscribing.
You join, you clicked the button right before we ended. Jatanya, wishing you luck. Thank you, Tatana. Noel, good to see you. All right, bye everybody. See you tomorrow. Have a good one. See you next time.
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
5 Mind Blowing Omni Uses Cases
PaulJLipsky
1K views•2026-06-02
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
