วีดีโอ

Thanks so much! I really appreciate it 🎉😊

nice! yes that's pretty powerful. Rust iterators are performant in this way. They produce values as requested. This is memory and CPU efficient. and I like code readability more especially when the performance is the same

yeah good point, there's a lot of these models that are being trained with censorship directly on the model weights. this makes it more tricky to bypass the safety aspects. Might be good to checkout less restrictive models like github.com/xai-org/grok-1 and github.com/deepseek-ai/DeepSeek-V3

good point 😊 you are right that, this is definitely a read-along overview. yes for sure everyone can read. I tried to add some color to the conversation as well

just testing to see if your comments are auto-generated lol

good test! the answers are human generated 😉 though it's tricky to identify these days

good question! unsure of the solution however

that is so true. HN is biased and heavily moderated. basically any condition where you can become shadow-banned

doing well! thank you for asking. lots of focus on new tech and AI these days 😊

@@StephenBlumoh you have no idea 🤷‍♂️ lol

@@StephenBlum have you ever experienced the oh one issue. it’s pig I’m cooking pork son

So I got a question this comfy UI thing it pretty much provides you the script for a UI and interphase that you can use. I’m trying to understand it because I’m trying to create a custom UI which I think that’s what I’m trying to create. I’m not very tech savvy, but I am super good at using AI so it shouldn’t be too hard to try to figure it out. Plus I pay for the subscription for GPT‘s most advanced features, I’m trying to create a few simple things like a custom UI for my custom tabletop RPG that I’m creating, which has unique mechanics different than most of the games out there but definitely still reminiscing of DND and Pathfinder and trying to figure out how I can create my own program almost From the bottom up, but this is only a side thing I’m trying to do, but I am fully creating a tabletop RPG which you’re probably one of the few TH-camrs that I actually would probably send an actual rough draft of the game. Just ask your opinion, but still And I’d probably run it on a windows system just because more people have that this is just stuff I’m thinking of, but it’s gonna be pretty much just like a hardback D&D book with its own campaign story lines and everything so far

that sounds like a great project! custom tabletop RPG 🎉 good idea using GPT to build this for you. Lots of options and frameworks out there to ask the GPT which one is best. Then suggest using that to the GPT when you are building it. this is a good way to start projects! Thank you for the idea 🙌

that's a really good note! the qkv are the same (self-attention) for encoder. and this is different for the decoder

nim has a few HAL engines like github.com/abisxir/alasgar which can leverage GPU acceleration. It's pretty powerful. And gives you easy access to OpenGL calls. It's low-level. You'd be building the game engine aspects yourself to do things like 3D animation

absolutely! sqlite ❤️ great for speed and a great starting point

good addition! thank you Matt 🎉 bits-ui for Svelte UI 🚀

you make a good point. angular can be easy when you start. Angular and React are both great JS foundations for building applications. React is designed to provide component-level functionality and is considered a JS library. Angular is a full JS framework and it is very powerful. To get the most out of Angular, it's best to learn more about the capabilities

Dude, don't be soo mean!

@johnsmith1953x sorry i had to say that

yes it's a health related reason. i'm working to get better 😊

@@StephenBlum damn you just made me sound like an asshole. Wish you best, take care man and of course thanks for the video

@@nandans2506 it's okay hey don't worry about it! you weren't wrong 😄

Yes absolutely here is the link: docs.google.com/document/d/1v9cR3GDPYtEUywMTUn7_uGYXZg5gxEMgMqBpKDOi1_s

@@koalakakes it's recorded through OBS desktop capture. The colors are slightly different than real life

I live in San Francisco, CA, USA

Totally! It was the good ol' days with lots of friends 😄

you are right! Rust has syntax on top of even more syntax. It's a lot to learn and it makes it a challenge to even know when to use the various statement operations. I would love a simple-rust similar to Python that makes it easy with the same memory and concurrency safety guarantees 😊

@@StephenBlum Totally agree. Because of this sometimes you feel you're getting nowhere with Rust learning...almost tempting you to switch to another simpler language, like Go!

yes please! simple rust 🙌 🎉 the language itself has been amazing. We have seen 5-30x improvements each time we replace an existing layer in our stack with Rust. It's performance of C with memory and concurrency saftey make it the greatest language. The only drawback is the syntax is overwhelming. If we can have a simple mode, most applications don't need the advanced yeet's 😊

​@@StephenBlum sure, but is this 5-30x improvement comparing Rust to Python? Then it would not be that fair anyway, I have listened to a bunch of your talks and I know your business needs a performance that "C level" languages offer.

@@PouriyaJamshidi yes you are right. the comparison is between Rust and Python, and Rust and Go, and Rust and JVM. We see amazing performance benefits by using Rust. Note that we also use a lot of C already. And we won't see a boost in performance switching our C code bases to Rust.

@@StephenBlum Right, I remember you talked about Go as well in the Rust in production podcast. I would be very interested to know if you ever write something in Nim and how it compares to Rust. We recently wrote a syslog proxy that handles a few hundred thousand logs per day (at times tens of thousands per second) without breaking a sweat. Though, your scale is different

@StephenBlum I wonder why Nim was not considered for the code base.

Yes good point! The source data needs to attend to itself. So we need copies of the input/output data to compare. That's why the QKV variables are the same. Duplicate data for self-attention so we can multiply the vector with itself to create the attention between each token. Example of translating between English and German: 1- Encoder Self attention Q = K = V = Our source sentence(English) 2- Decoder Self attention Q = K = V = Our target sentence(German) 3- Decoder-Encoder attention Q = Our target sentence(German) K = V = Our source sentence(English)

Again neither the query, key and value vectors are identical nor their weight matrices Q, K and V. The 3 vectors do stem from the same hidden vector h pertinent to the same word, i.e. q = Q.h , k = Kh, v = V.h , but that is it, they are not identical. Is this an experiment/joke or what?

@@yvesbernas1772 Yes you are right. You identified this correctly. This clarification was necessary. Thank you for mentioning this. Yes the three vectors are ultimately different in the way you describe. The problem I ran into early on learning is how they start from the same hidden vector h. This was a huge missing part and it was important that you helped describe this detail. Thank you! 🎉

Good point! Rocket looks clean. Like a simple and elegant web app server framework. Simplicity is key. And it looks like Rocket does a great job 😄

@StephenBlum yes it does and easier to pick up. My opinion though.

thank you! that's great to hear. shadcn-svelte is excellent 🙌

`Manaco 15` is the font 😊

Good idea! Here is the checklist direct from the video: docs.google.com/document/d/1I_9e0kESyWCcoqiLfdCfjctvul7RaO0DLpLx-ynnlY0/edit?usp=sharing

Yes absolutely. Here is the checklist from the video: docs.google.com/document/d/1I_9e0kESyWCcoqiLfdCfjctvul7RaO0DLpLx-ynnlY0/edit?usp=sharing

@@StephenBlum Thank you!!!!

Yes you are right. ReactJS is very slow. Preact can help speed it up. And if you use the preact-compat library, most components should work github.com/preactjs/preact/blob/main/compat/package.json

Good point! Q dot t(K) is the token by token matrix. Each token is a vector, commonly 512 floats per token. Both Q and K are the same array of tokens

Kali Linux comes installed ( or offers easy access ) to all pentest tools. For specific pentesting, you just need to create a list of your favorite pentest tasks. A great way to get a list of ideal tools available, most of the LLMs can provide a list that matches your specific tasks 🎉

@@StephenBlum thanks for ur advise

Great suggestion. Illustrator can export directly to web-ready SVG. And you can use that SVG to make fantastic textures 🎉 just save/export the file to SVG. If there is an optimization option for web export, make sure to select that option so the file is smaller

good question! for "fast" with a million polygons, you have to do it progressively. If you are looking to draw all polygons on the screen at the same time, then it might be a non-starter. If you are only showing some of the polygons on the screen, then using most 3D render frameworks will automatically take care of this for you by only drawing the visible polygons in the viewport. 3D frameworks automatically cull. Here are some types of culling: Back-face culling: A technique that avoids drawing polygons that are facing away from the camera. A polygon's vertices have a clockwise winding if you're looking at the back face, and a counter-clockwise winding if you're looking at the front face. View-frustum culling: A visibility culling technique. Occlusion culling: A visibility culling technique.

@ so it meant to show it effectively - don’t show what is not visible at moment? I had in idea for example to have 20 layers for 20 zoom levels and cluster polygons when zoom out , and when zoom in hide what’s not visible in the area. Something like that

@ by the way, happy New Year! Very nice channel

@@sergannn thank you! happy New Year! 🎉

@@sergannn that's a great idea! that would dramatically improve performance. Multiple LOD level-of-detail layers/zooms so that when the camera is zoomed out, a simple model with fewer polygons is drawn instead of the full 1m polygon model

you're welcome! 👍😄

that's actually a really good point! as of today, most of the software we need has already been built and it works mostly well. we can make improvements and fixes. other then that, it's already been written

great to hear you say the video was helpful, your blog is really well written 🎉😄 your blog helped me out a lot thank you!

​@@StephenBlumThank you

@@siddharthmanumusic thank you! 😄 Yes absolutely you are right. C++ is relevant and heavily used in advanced AI frameworks

No problem! glad you found the video 😀

Any time 😊 thank you!

agreed, the Gemma models and Google AI models are heavily censored for safety. This can become a problem even for seemingly simple tasks, the AI will guard it's answers

`Monaco 15` is the font family. running on a Mac terminal emulator window

@@StephenBlum thank you

you're right, that's an important perspective on interest rates. with cheap money, large green (new) projects start with lots of software development. Without cheap money, companies stop making new projects

Smart question! you figured it out. Yes processing videos on servers is smart that you figured this out. AWS charges for each second the EC2 instance is turned on. The time is about 2-3 minutes per video ( about $0.50 per video ~ )

Cool Thx!

I appreciate that! I put a lot of effort into creating this video 🎉

@@StephenBlum you have really good content. If you use canva or Microsoft designer for video cover. Makes them more clickable. And call to action in your videos to subscribe. Love your content. Should be hitting 50k subs with a few tweaks. Especially on topics like pandas, ML and LLMs

you have good video tips! this is helpful thank you! 😊

one option to get 100% up and running with your at-home GPU and remote access is to use PubNub.com API to send/receive the text data. it does take some setup and coding. you can create a chat app interface that delivers the prompt to your GPU running the model, and then publishes the response back. user app -> PubNub Publish API user prompt -> your server subscribes to the events and submits the prompt to the GPU model -> your server publishes back the answer to -> user app is subscribed to the response channel and receives the answer. Copy and paste the description above into an AI tool to build the app for you 🚀

you make a strong point. the AI is just a tool. let's imagine for a second that there is an AI tool that creates perfect code. You still need a human software engineer to direct it because there's scenarios like you describe where the software needs to be maintained. The tool needs guidance

you are right, the title is wrong because it's mostly a discussions video and a mini getting-started intro with a code walkthrough. that doesn't really count as a tutorial in this case, more like it's an overview

@@StephenBlum more like it's a clickbait.

@@true_visual good point. the video is more of an overview with some code reading of the tutorial. It's more of a video reading the public tutorial rather than me creating one from scratch

​@@StephenBlum So why didn't you name it "Elysia JS Overview"? You waste other people's time by putting a false title for engagement.

@@true_visual good idea. I'll update this now.

the managers think that it saves more time than it actually does. there is a lot of hype. the GPT tools do help a bit. it's great that it saves some time we using AI tools. It allows us to focus on the more creative parts of our work