the last macOS setup & workflow you'll ever need (I swear)

I'm sorry, guys. It's been way too long.

It's been way too long since I posted a YouTube video here. It's been, I think, four months. The last video I posted was

four months. The last video I posted was on Tensor Core operations. And uh and then we just went completely dark. And I

apologize, but we're back. And as a matter of fact, we're so back. This

one's going to be such a banger, you're going to want to click off sometime in the middle. I don't know. But I'm going

the middle. I don't know. But I'm going to throw so much at you that you actually might click off. Like, be

warned. Okay. Um, I want to go over something like the none of this is sponsored. Uh, these are all my

sponsored. Uh, these are all my preferences. Like I don't get paid for

preferences. Like I don't get paid for any of this. I'm not even well known enough for someone to reach out and be like, "Hey, can we give you money for this thing on your YouTube video?" Like,

no, no, no. Um, but yeah. So, so, so none of these are sponsored. Uh, these

are all just this is my workflow. This

is basically what I'm using right now and going through, you know, sort of the Dunning Krueger effect of just like I think I know what I'm doing and now oh, we're actually kind of ascending that again, I think. Uh, and this is kind of

just, you know, this is going to be basically everything that I use just on the the MacBook here.

Um, before we get into it, please do consider following me on X. You can find link for that in the description, you know, consider subscribing. I probably

won't post ever again, but just subscribe to me anyways, you know. And

then if you do want to jump on a call with me, uh, you know, an hour, however many hours you want, and just, you know, maybe talk vent a little bit, um,

you know, or even just get some general advice on how to up your workflow or any anything code AI related. Um, I'd be happy to do that. Link also in the

description. Um, now first things first

description. Um, now first things first is you need a good background. And this

is what we have here. This is some image that I got uh found off of X from some wallpaper. You need to start off with a

wallpaper. You need to start off with a good background, dude. You have to um but no, for real, for real. First section is going to be

real. First section is going to be productivity and note-taking. All right,

so I'm just going to give this to you raw. This is like unedited at all or

raw. This is like unedited at all or sorry, totally uncut.

First thing is productivity and note-taking. All right. So, there's this

note-taking. All right. So, there's this little tool that I like to use. It's

called Apple Notes. Um, let me open it really quick.

Apple Apple Notes. Let me close it and open back up again because it's being buggy. There we go. This is Apple Notes.

buggy. There we go. This is Apple Notes.

Um, so you know, you all know what Apple Notes is, but I just like this as a general tool because you can I know it's it's hard. It's it's

hard not to laugh. It's Apple Notes, dude. Come on. Um, you can use it on

dude. Come on. Um, you can use it on your phone. I mean, if I just want to

your phone. I mean, if I just want to write something quick down and I don't want to go through any other mess, it's like Apple Notes just works. It syncs

and everything. It's It's covered by Apple, right? So, I can go on my iPhone,

Apple, right? So, I can go on my iPhone, which I'm actually recording this on remotely, is you can you can just use that. And I find this like extremely

that. And I find this like extremely helpful just to jot quick things down and then to port those on later once you actually get back to the computer, right? So, you want to write someone's

right? So, you want to write someone's number down, you know, pick someone up at the mall or you just like, I don't know, need to go and take care of something or you're like, I want to go

to that event, but I don't have internet or something. I don't know. Whatever it

or something. I don't know. Whatever it

is, I like to use Apple Notes for that.

There's not I mean there's like Google Keep but Apple notes is usually what I go to and then there's you know so so there's Apple notes super easy and then I usually use

Obsidian for uh generally just like managing stuff. So let me pull this up.

managing stuff. So let me pull this up.

This is what I've used Obsidian for. So

you know we've uh we've got like this is by the way this is me leaking my entire like Obsidian workflow. This is like extremely personal so please do not roast me over this. I will cry. But this

is where everything happens. Um, we've

got everything from like general notes to um ideas like research, startups, workflow, everything. Um, you know,

workflow, everything. Um, you know, ideas for content creation. Um, a bunch of literature. So

creation. Um, a bunch of literature. So

like research papers that I organize um everything is just in here. And I can I have some plugins that I use where I can actually talk to these. So, I can go like archive, download from clipboard,

and then it'll just port this into here.

Um, all these plugins that you can add to Obsidian. It's very modular. So, I'm

to Obsidian. It's very modular. So, I'm

not going to say my workflow for Obsidian is the best. It's something

that I would advise you to sort of expect to change on your own, but Obsidian has been great for me. Some

other options would be like maybe notion or RemNote. Um, I found RemNote to

or RemNote. Um, I found RemNote to sometimes be a little clunky in how it handles stuff, and I haven't used the space repetition feature as much as I'd like. You can also add space repetition

like. You can also add space repetition to Obsidian. The only con with Obsidian

to Obsidian. The only con with Obsidian is that it does not sync by default with your phone. So, you can't download the

your phone. So, you can't download the mobile app and have it sync by default.

You actually have to set that up manually or or pay a little bit of money for it. So, that's that's the only

for it. So, that's that's the only downside. But then again, you have Apple

downside. But then again, you have Apple Notes. So, uh this is sort of what my

Notes. So, uh this is sort of what my Obsidian looks like. And then we have, you know, let me actually increase this a little bit. We've got like uh some progress. So just like journaling, you

progress. So just like journaling, you know, going and maintaining the the stability up there, some simple life tricks. Um a bunch of textbooks that I

tricks. Um a bunch of textbooks that I may or may not have downloaded. Uh yeah,

we're not going to talk about that. Um

but yeah, and then we've got like, you know, some basic stuff like how do we how do we reverse entropy? I don't know.

This is just stuff that I need to waste brain power on just to reset and it's really good for mental health. So,

you'll just find the stuff like this around in my um you know, yeah, just absurd stuff. Absurd

stuff everywhere.

Um Obsidian's great. You know, we got Journey, we've got planning for the future, the final boss, I call it, which is in fact the second law of thermodynamics. And then just a little

thermodynamics. And then just a little dump. You know, Obsidian's great that

dump. You know, Obsidian's great that way. Um but anyways, that's sort of like

way. Um but anyways, that's sort of like that's sort of like section one. There

is one more thing though I should probably mention. I use this thing

probably mention. I use this thing called linear and it's good for managing stuff especially as a team. I don't work with very large teams right now but I expect to in the future and having

linear is awesome. Like I can literally just um for instance this is what it might look like. So I can go into projects and there's there's this book I'm writing right now called CUDA for deep learning. It's the stuff that

deep learning. It's the stuff that software layer that runs on the Nvidia GPUs and I can add stuff to this. So I

can add like resources and I can I can like quickly change things. I can go like command K and it's very quick to edit stuff. Can like add a link or

edit stuff. Can like add a link or change project priority to say uh hi and then that will get updated and you know people can manage it. You can add people to it. So I can add uh someone new like

to it. So I can add uh someone new like email, right? It's just it's just very

email, right? It's just it's just very quick and snappy for this kind of management stuff having uh charts everywhere and um you know I can use like command B and

I can uh yeah if I go back to like uh say this menu here where all the projects are listed. I can do command B and I can go

listed. I can do command B and I can go to like backlog or any of these uh different uh like progress categories you could say um you know a timeline. So

I can have a little Gant chart here um and organize things. So if I like say I want to complete this first and then after this is completed I can complete this I can sort of like drag lines there

and have things um organized and then back to normal. So one so one two three um just those options. It's like really nice. I can also go like C and create a

nice. I can also go like C and create a new issue anywhere and I can change like the you know a bunch of interesting bunch of interesting stuff you can do with linear. I haven't fully mastered it

with linear. I haven't fully mastered it yet, but it's one of those things where you can really scale up your workflow, I feel like, and it's come in handy for me so far.

Um, yeah. And then in terms of, you know, being a developer, cuz, you know, probably we're we're maybe maybe you're a dev, maybe you're not, but I personally use the terminal quite a bit.

And so, uh, when we're up here, you know, I have this thing called warp terminal. Um, if you if you look it up,

terminal. Um, if you if you look it up, Warp Warp Terminal, Warp.dev, literally just download it and

it uses it has AIS built in. I don't use the AIS as much. I probably should though, but it's just uh it's just very like snappy. Um, you can like do this

like snappy. Um, you can like do this and ask for a command suggestion. So

like uh list all and then you can say like uh do this and it will just right it's sort of quick that way where you don't have to memorize all the Linux commands. You don't have to read the

commands. You don't have to read the 300page Linux textbook to understand all this. Um and so I kind of use terminal

this. Um and so I kind of use terminal you can like bounce around and stuff.

Uh and then this other tool that I use so probably heard of Neovven before so like a text editor. there's you know VS code or like um you know PyCharm or or

ina in some cases Vim and I use Neovim for some things and uh it quits it comes quite in handy but not by default by default is

um new by default looks can look kind of weird. It can kind of look like this.

weird. It can kind of look like this.

And it looks strange, but if I were to go into like an actual like say a Rust project uh and src main, right? Like it

looks it actually looks like pretty good. Um just like the transparency and

good. Um just like the transparency and everything. So this is something I've

everything. So this is something I've integrated. And the way you can make

integrated. And the way you can make this look nice is through a tool called um Envy Chad. So I use nvychad.com Linux Mac OS. You just get

clone it. Um, and then you do this

clone it. Um, and then you do this runmason install command and then you can go like ts install uh or where was it

again?

NVIM ts uh install info and then you can go up and down and basically install different um language servers for uh each of these

languages. So, I've got like CUDA, of

languages. So, I've got like CUDA, of course, I've got C++. I don't have say like uh I don't know, blueprint or bib text or any of these. Uh but you can you

just it makes it look better when you actually go into your editor and mess around with those. So, NVCAD is great.

And then another tool I like to use inside of NVCAD is um I'm going to go find it really quick. Config and Vim.

And I go to my Lua. And you'll have to actually learn Envim a little bit. So,

there's some overhead there, but I like to use it for like quick little edits.

If I need to go edit my like bash rc file or or quickly like change a like change a code snippet somewhere, I'll just pop into here instead of letting VS Code load and everything or actually

cursor. But uh yeah, we go into uh what

cursor. But uh yeah, we go into uh what was it?

Plugins and then I use this thing called Avante Envim that basically uses Claude Claude 3.5 Sonnet to do uh completions

for me. So I can say uh I go like space

for me. So I can say uh I go like space double a and I can say uh hello world. I

can uh send that over and cloud will actually uh or give it a second. Yeah. And then it will complete

second. Yeah. And then it will complete a response there. Uh so it's just it's sort of like cursor but in nvim. And you

can like edit things quickly and I can go like up and down and stuff. Uh it's

pretty cool. But anyways, that's basically it for development and terminal tools. Um, I love I love the

terminal tools. Um, I love I love the whole neoimb chat thing. It's it's just like great. Everything is done in the

like great. Everything is done in the terminal. Um, next up for creativity and

terminal. Um, next up for creativity and content creation, I'm actually using this thing called Screen Studio, which I'm using to record right now, so I'm

not going to open it. But Screen Studio is a recent tool built uh for Mac OS specifically. And you can basically um

specifically. And you can basically um it's like a little I don't even know how to explain this. It's like so easy to just use. You literally download it and

just use. You literally download it and you I think you pay a little bit of money for it subscription, but it's way easier to use than like OBS. So in OBS Studio, you will like start recording and then stop recording and it will just

capture the screen and the camera will stay in like some awkward position. But

in this one, you notice that my camera is actually down and it's got this like nice um sort of bubble around it.

Um this is Screen Studio and I can like I can stop playing and I can start playing again and I can select a specific chunk of the screen. It's just

phenomenal. And I use it to actually record like most of my content now. Um

because it's just so easy to use. And

you can like, you know, edit something like trim it like this and then send it off as like an MP4 somewhere and you don't have to pop into another another tool like say Cap Cut um to actually

finish off editing. But I usually do like to go into Cap Cut anyways because it's such a great tool. Now Cap Cut is interesting. It's it's also a native Mac

interesting. It's it's also a native Mac OS app. I can like create a project and

OS app. I can like create a project and say like one of the time-lapses that I did recently for I do these coding time-lapses on X and let's just say um

we took like 47th, right? Pop it into here.

Um this is when I was on a trip to Helsinki in Finland and we were working all night, all day and all night. And uh

yeah, I can just go into here, literally drag and drop and then like say I go finder gen clips and I can say

like put uh the Arya Math Protoar remix in there and it will just sound amazing.

Um and then I can like go into here, I can change the speed. Um and you'll see that that'll like shrink. Uh we can like sort of do this and it's just like easy

to um you know scroll in and out like everything is as you would expect sort of when you start using a new piece of software and you try to like tinker around with things like drag that into that window or just like try some weird

finger movement like things just work as you expect them to and Cap Cut's great for that.

Um now last tool it's it depends on what content you're creating but when you're technically when you're doing technical lectures like I am this can come really this is can really come in handy so I

use uh this thing called excaliraw and I can go and draw like a new say I want a new diagram say like I don't know

v1 or no go private oops Um, I go into private say and I I create um I create a scene. We'll put this together. Let me

scene. We'll put this together. Let me

full screen. And I can for instance pop down to here and like make this in dark mode. And then I can just like draw

mode. And then I can just like draw stuff and I can change like the opacity of it. I can change sort of the

of it. I can change sort of the different different schemes. I can

change the colors, uh, everything. And I

can use this to put together these nice diagrams to explain how concepts in artificial intelligence work, like back propagation or running code on GPUs. Uh,

this this can come really in handy cuz the intuition you it's literally limitless. You're stuck to this 2D

limitless. You're stuck to this 2D diagram that looks like an actual sketch pad and you can just draw things on it on here and it's like very quick and there's keyboard shortcuts and everything. So, I'll like sketch

everything. So, I'll like sketch something up on the whiteboard and then I'll go into here and like literally copy and paste it and like maybe polish a little bit and then it's like ready to ship and show people. Um, so very useful

tool. I paid a little bit for this. It

tool. I paid a little bit for this. It

was like I don't know $8 or something um per month and uh I literally just get to have my own storage of them. So I don't have to back up files. It will just save it in the cloud for me and it's awesome.

So Scholar Draw is another great tool.

Um but yeah and then I'm going to get into some more like technical ones now.

So system customization and utilities I like to call this I mean this is the recommendation that actually Grock gave me um this is what it the title a good

title was but uh so first off I'm going to start off with this thing called raycast and raycast if you have not heard of raycast it is like if you're not using raycast you have to use it um

it is insanely good um literally you you install raycast I'll just bring up actually the browser again but Uh, raycast

uh, raycast.com your shortcut to everything, right? Download for Mac, right? AI

right? Download for Mac, right? AI

extensions, everything. And literally,

um, I can fly into here and I can go command space and it will just, it has my, it has my thing saved, my profile picture, and I can say go to like um, I

can go to Discord. You probably seen me do this before.

Um, cool tricks inside Discord and I can ask AI with tab and it will complete with GPT4 mini, right?

Um, I can go to like say Discord for instance. Oh, that just popped up. And

instance. Oh, that just popped up. And

now I can go to uh say Spotify. Um, or I can go to cursor, right? Or I can go to uh another a really cool tool actually

is if you go to this store um you search archive if you're into reading papers like I am archive search. So you can literally uh once you install this tool

so you can go search archive and I can go say uh this new inference optimization called SG lang like this and I can pull that up and it will bring up archive paper for it instantly. Um,

and you just have this massive search over the entirety of archive and you can pull any paper up instantly. It's like

very very fast. I was actually going to build a tool for the specifically to do this and then I figured out that raycast has already has one. So if the if there is a tool that you need um just like

literally check the raycast store and it will probably have that tool. Um or just check GitHub or ask like perplexity um which is another tool I'll go into later. But yes, so you can you can bring

later. But yes, so you can you can bring up all this stuff and then yeah, like raycast is wonderful. Um I can move windows around. So if I were to say um

windows around. So if I were to say um like take take my obsidian window, go like left and go right go

um oops go center. Uh I can go center like last third. I can go first third.

last third. I can go first third.

Um, yeah, it's just it's beautiful. But yeah, yeah, Raycast lets

beautiful. But yeah, yeah, Raycast lets you do a lot. It's uh it's quite it's quite impressive, actually. Um, but let

me slide this back down.

Um, this is a little more niche. Now,

this is called Carabiner Elements. This

is a Where was it? This one. Yeah. So, I

go Carabiner Elements with Raycast. Um,

and you can install this on your own, but basically I use this to write macros and map stuff around. I'm not very complicated with macros. I don't use

like crazy but um when it comes to like remapping something, for example, um pressing escape every time I want to like exit raycast like like this, pressing escape every time would be super annoying. So, I just bind

super annoying. So, I just bind everything to caps lock and then I just type all of my capital letters with shift so I like not capitalize when I'm sending people messages. Um, hence the title of this

video. But, uh, Carina Elements is great

video. But, uh, Carina Elements is great and you can just do simple mappings like this that will make your life way easier. Um, I was looking for a tool to

easier. Um, I was looking for a tool to sort of do this remapping easily and Carabiner Elements was the first one that was that actually came up. So, this

tool has been like pretty helpful for me overall and it's got a lot more stuff built into it that I haven't even covered yet and will not be covering.

Um, and then there's this upper tool up here called better display, which uh, you know, I can adjust the brightness with just flickered off. Uh, I can like

turn the brightness of that. I don't

know what it looks like from your end, but I can definitely change brightness there. Um, and it just kind of it makes

there. Um, and it just kind of it makes things easier um, just for individual displays. And you can have virtual

displays. And you can have virtual screens and change system colors.

There's just a bunch of options there which I find pretty nice.

Um, but when you want to say like remote into, you know, another machine, like let's say I start off in the terminal

here and I want to remote into my um my GPU cluster out there in the breaker room um because it's very close to the router and power and everything. It's

just it just works and it's very cool in that room so my GPUs can stay at a nice temperature. Um it won't warm up the

temperature. Um it won't warm up the entire house because it's locked in this little cabin. Uh but basically I can go

little cabin. Uh but basically I can go SSH and uh into this. This is a special IP address that's specifically for my rig. I can go into here, type the code

rig. I can go into here, type the code and I am now in and I can go to like my say my CUDA directory, right? And now we

have everything here. So I go like know cd ptx and I can nim from here and like check uh like check stuff that's

literally like I'm wirelessly connecting to that uh machine in the other room. Uh

and the best part is that you might think oh well SSH is just everything in your local network but no this thing up here called tail scale this um this actually lets you connect from anywhere

in the world. So, uh, a month ago when I was out in, uh, Finland, I could connect into this, um, into this instance and do training runs and do experiments when I

was 8,000 km away from my house. Um, and

Tail Skill is great for that. So, let me just quickly show you what Tail Skill looks like. So, we can log in.

looks like. So, we can log in.

Um, and I'll quickly just This is my tail net and there is hopefully I don't leak stuff. Um, but

yeah, like you won't be able to sign into any of these unless you're actually in my tail. So, don't worry about it.

But I got my iPhone connected. We got uh this instance, which is my Linux rig, and then my MacBook Pro. Um, you know, and then a friend's and then a friend's MacBook as well. I gave him root access

just to do experiments on my rig. But uh

yeah, you can add your friends to this and basically you would go download and then there would what I would recommend you do is actually go talk to an AI called Grock like this and you can just

ask it to um grock.com. You can go and ask it how to set up tail scale. you

download it um through like a a referral link and then you would set up and check off a bunch of settings and then after um some continued you know after some efforts

it's not just like it's not the easiest thing to set up in the world but um you know it's it's on the easier side I would say of all things and you can sort of just get it spinning and then you can

connect all of your rigs together and uh yeah you can just talk to your GPU cluster from across the world when it's in your house and you'll be the only one who can talk to Uh, so pretty cool tool that I use. Um,

and then there's one thing I did want to cover was like sort of my directory structure. Um, I've been working on this

structure. Um, I've been working on this for quite a bit and I thought it would be worth like sharing with y'all. So,

basically what I do is um, and I don't know you you might not care as much about this, but I'm like kind of I'm going through that organizational phase where everything needs to be in perfect order. And

literally this is what we start off with. So, we got like applications, etc.

with. So, we got like applications, etc. just the basic stop on Mac OS and then uh we'll have gen which is where everything happens. So I'll go into gen

everything happens. So I'll go into gen which is general and then uh I can go like ls in here. So I have this chat app that I'm building. I've got all these

clips. So I go like ls chat um we got

clips. So I go like ls chat um we got all this. We got like ls clips. Uh all

all this. We got like ls clips. Uh all

this like MP3 files that I can play um ls cursor. So I've got like these cursor MCP tools which by the way do not do yet. the cursor MCP tools are just

do yet. the cursor MCP tools are just not it right now. I would add this to the workflow, but there's just too much overhead to make it make sense. Um, some

people might find it valuable and in that case, you might want to look into cursor MCPS. Um, it's called model

cursor MCPS. Um, it's called model context protocol, but I'm not covering that in this video. And then there's like um a Grock desktop app, which I

installed, but I don't use that much.

So, I just mainly keep everything at this. There's like a hierarchy. There's

this. There's like a hierarchy. There's

like all the main Mac OS stuff and then there's Gen and then it goes down into say like more special stuff that I want to have more organization over. So like

clips or cursor specifically or like images or um ISOs. So I've got like um you know got like Iuntu, I've got a

Toyo, I've got Windows 10.

Um and then I'll have like prompts. For

instance, Pokemon is Pokemon is a current uh client work project that I'm doing. So, I'm not going to show that.

doing. So, I'm not going to show that.

But, and I've got like prompts. So,

inside of here uh I've got like this expert say uh list prompts uh or cat prompts ex expert conductor read me. And

it will be just like um prompts for language models on how to do better. Uh

let's see. Ultra deep think read me.

uh or sorry we'll do this and then cat. So

ultra think ultra think mode basically you can take a model like opening I03 or 04 mini uh that's what the models are right now at least and you can just say go even deeper uh and it will like

search everything and try to you know it just basically ups the quality of the model. Then I've got another one here.

model. Then I've got another one here.

Um, prompts. What was it again? Uh,

hyper rationalism first principles. You are hyperrationalist

principles. You are hyperrationalist first principles. Problem solver with

first principles. Problem solver with zero tolerance for excuses.

Rationalizations or or Pure focus on destructing uh deconstructing problems. Basically, um, imagine you're working uh imagine you're working at SpaceX for Elon Musk and you're trying

to break down a problem. It's like you literally have him right here and it will just like rigorously break down everything, challenge all assumptions, everything. No fuzzy thinking

everything. No fuzzy thinking whatsoever. Specivity, specificity on

whatsoever. Specivity, specificity on all things. You get the point. And you

all things. You get the point. And you

can actually find these um on my GitHub repos. So,

repos. So, github.com/invatoshi/hyperrationalism

github.com/invatoshi/hyperrationalism first principles. Like you can find this

first principles. Like you can find this here. Um if you go to my repositories

here. Um if you go to my repositories you can see there's like ultra deep think expert conductor um hyperrationalis and first principles

down here. Um so that's kind of what I

down here. Um so that's kind of what I use for those. And then what's another good one? There's uh pi. So I'll go in

good one? There's uh pi. So I'll go in here. I have all my Python projects laid

here. I have all my Python projects laid out just like literally if we were to list just like one of them say uh I don't know llama VA which is llama voice assistant. I go ls. It's just

voice assistant. I go ls. It's just

literally a Python project setup. It can

be very like sloppy, whatever, but just like the prototype for anything I do goes into there. Um, and we can just quickly iterate uh locally on this MacBook

Pro. Uh, got Riy, which is Raspberry Pi.

Pro. Uh, got Riy, which is Raspberry Pi.

So, this little Raspberry Pi over there that's doing a time lapse of me working for the entire night. um is actually where you'll see me on X is probably you know time-lapses of me working all day

and night is uh yeah all that stuff is done like here. So uh I don't have any I have a time-lapse maker which I haven't actually dragged into here yet but sometimes I'll just take a song from

YouTube for instance and I'll just drag it into here. I go like M player uh and I'll go like I don't know Dreams or something and it will just play the song on loop all day and I could just vibe to

that. Um, so this is a little thing you

that. Um, so this is a little thing you can hook up like your speakers to that or like onto here or something. And I

don't know, just kind of like a a placeholder tinkering directory for the for the Pi. And then, you know, we've got Rust of course. So, you know, different things we could do here. So, I

can go into like say um CD tensor is uh cargo cargo

run. Sure.

run. Sure.

I know that's going to take too long.

Uh, but you get the point. All the Rust stuff goes in there. So, we have Python, we have Rust, uh, we've got CUDA. So,

bunch of CUDA products that I have, projects that I have. So like I don't know for instance tensor cores with like

nvim and you could see like tensor corores.md or um a cuda file which I use

corores.md or um a cuda file which I use um which I use cutless tensor cores. You

can sort of like play around with stuff and I include a lot of this here just to tinker with. Of course you can't run

tinker with. Of course you can't run CUDA kernels on a MacBook because it does not have an Nvidia GPU in it. It

has Apple Metal GPU. Uh but you can just store it on here, do like quick iteration locally and uh you know push it up to push it up to your cluster or something.

Um so I just like to do like quick you know or or store like a template on here of something uh so that I can quickly send it over to any machine that I need to. Um because if something is over

to. Um because if something is over there then I have to like send it here and then send it over to that machine.

And when you have all your templates and examples just at one centralized entity, then you can send that to any GPU cluster environment that you want.

Um, what's a good one? What's a good one? Maybe maybe

one? Maybe maybe scripts. So like MCPS, if I want to

scripts. So like MCPS, if I want to previously when I was getting my MCPS going, I can just launch this thing in the background and it will literally just basically a shell script that will just run things. or like when I want to

create a new GitHub repo, I have a script that's aliased to the command gcreate. Um it's like a single word that

gcreate. Um it's like a single word that I do uh gcreate like this. Um and it will just it'll literally just push put up a new one and put all the contents in my

current directory and just dump it into a GitHub repo and I can make it private or public. So just having scripts like

or public. So just having scripts like this are very useful part of my part of my workflow. Thought I should share that

my workflow. Thought I should share that with y'all.

uh templates. This is a good one. So,

CUDA templates. Um there's like TC which is tensor cores and then we've got like ls torch. So, if I want to say like have

ls torch. So, if I want to say like have a have something in PyTorch that I then call to a GPU kernel. I can I can compile that GPU kernel um with like a

setup script. So, I'll go like Python

setup script. So, I'll go like Python setup.py and then I can call main. I can

setup.py and then I can call main. I can

call this main uh script here and that will reference those compiled binaries and then it will call them whenever I need to do an operation. So if PyTorch doesn't have something uh natively if a

new research paper comes out or a new neural network operation I can go and write my own kernel for that and make it optimized and then I can take it back to um I can take it back into like one of

these single files and then just have a bindings.cu to you uh and then you know

bindings.cu to you uh and then you know just run it faster by default in PyTorch um and just have my own implementation for that. Uh so it's just kind of nice

for that. Uh so it's just kind of nice when you want to be on the absolute bleeding edge of things and I know as a as a systems as a systems guy this is very useful to have because you literally are on the bleeding edge

nothing like a lot of what you do will be uh a lot of what you do will be referencing stuff that's old old literature because that's what hardware

is based on. nothing is like absolutely bleeding edge like software is. Um but

you know you will you will have a lot of bleeding edge software too that no one has seen and it just really helps to have templates like this.

Uh another template might be like uh I don't know ls python so like nano gbpt or mlx benchmark uh or like rust which I haven't put

anything into yet because there are I don't write much rust code uh web development for just like you know you know little react website or

something and then this y repo which ah it why not basically basically just everything existential crisis like how do you uh basically that that thing in

obsidian that I had where um we would basically would go through like solving intelligence and traveling at the speed of light and uh having unlimited energy like all this stuff and

just like why like why are we doing everything that we do and something in along the lines of that is like questions to or answers to the universe

itself and like why are we here? um why

does the universe exist? Just like

everything that stems down from there in sort of a web format. Uh and that's sort of what I've ported into this little thing here. I felt it had enough

thing here. I felt it had enough importance that I'd put it in Gen. Uh and then yeah, basically that's

Gen. Uh and then yeah, basically that's my whole directory structure there.

Hopefully that like helps you understand a little bit how I organize things. You

don't have to do it the same way as I do. You might be more structured and

do. You might be more structured and like less sloppy and prototypy with things, but that's how I generally do it. Uh and then yeah, let's so I already

it. Uh and then yeah, let's so I already went over those prompts. This is a this is a section I want to sort of discuss

um how to use AI in your workflow um and specifically what I do to use it. So,

one example of using AI would be like if I'm talking to Grock and I want to be like, uh, write me a CUDA kernel that does, uh, that does the top K operation

really fast for a small batch size. What I just did

size. What I just did there was I pressed the function key and it opened that window. And this is called super

window. And this is called super whisper, right? You can drag this

whisper, right? You can drag this around, you know, make it smaller, uh, whatever you'd like. And, uh, yeah, literally

just exit. So, I can go to super whisper

just exit. So, I can go to super whisper and I can click up at the top here, go settings. There's a bunch of modes. So,

settings. There's a bunch of modes. So,

I can have like a default mode, uh, message mode. There's you know there's a bunch of things you can configure here but um I have selected

the ultra mode which is uh right here actually um or I can't remember which ultra I got but anyways uh the whole point is you can do this like very very

accurate and fast um speechtoext translation. So when you're inside of

translation. So when you're inside of cursor uh and you want to iterate quickly and you don't want to have to type everything out and everything is just prompting for you. It's sort of like this vibe coding thing if you've

ever heard of it where you just let the system be like chaotic and you just keep asking the AI to generate new code and new code and new code and you just build the thing up just by prompting it. But

when you're typing like for instance I could go on like typing test and maybe get like 120 maybe 140 words per minute like on a good day. But uh if I were to just talk to it like this how I'm

talking to you now that would that might be like 180 to like 220 words per minute. It's like very fast. Um, and you

minute. It's like very fast. Um, and you can do that consistently and it's just like streaming your thoughts. So

everything will just come out the way that you like there's there's no planning, right? So you have to watch

planning, right? So you have to watch out in text you can actually plan how you write the characters. Whereas in

this example, um, I just can literally feed into Grock. I can say here's my idea. Um, now turn this into a better

idea. Um, now turn this into a better prompt for a coding model. Um, so that way it would you would just send it and then it would generate the thing. you'd

paste that over into cursor and then you would get a really good output because the prompt is optimized for those for those coding models. Um, and they're very straightforward and contain less tokens. There's more specific details in

tokens. There's more specific details in there that it needs to know about as opposed to you just vaguely blabbering on about things. Um, and so Super Whisper has been very helpful there.

Um, but yeah, so Super Whisper is great.

You know, transcription like it's it's it's better than the Mac OS dictation tool from my experience. Um, it's

faster, higher quality, um, and they're adding some really cool features to it.

Um, and it's just overall just, I think, better. And it it does cost a little

better. And it it does cost a little bit. I think it's like $8.50 a month,

bit. I think it's like $8.50 a month, but, uh, that's that's been worth it for me from the amount of time of typing that saves. Uh, it's quite it's quite

that saves. Uh, it's quite it's quite worth it and scales well. Uh, and then there's this like perplexity tool, which you probably heard of already. I can

just be like, uh, what's the weather today in Edmonton?

started a little early. It messed up with the typo, but uh yeah, it's temperature like 5 degrees. Yeah, it

feels like 5 degrees outside. I live in Edmonton.

Uh 5 degrees outside. Yep. Perplexity is

great. It's a little search tool. You

can argue that it's better than the search browser itself, the actual search engine built into like Brave or Google Chrome. Um it's great. And then this

Chrome. Um it's great. And then this other tool, uh let me see. I'll bring up this example here.

see. I'll bring up this example here.

This is cursor. Uh this is a cursor.

This is a tool that I use to uh basically do all of my programming. And

so if you haven't heard of cursor, I assume you probably have, but if you have not, I'll give you a little demo here. So you it's it's the same as

here. So you it's it's the same as Visual Studio Code. Um it's literally a fork of VS Code, except you get this little chat menu here, and you get these autocompletes. So if I were to say, um

autocompletes. So if I were to say, um add a dropout layer and it'll add dropout.

Um, and it's just sort of quick that way and will autosuggest things to do and you can sort of just like escape escape.

No, no, no. Or you can tab tab tab dab tab and have it autocomplete everything.

But the most important feature is if you slide this out, you have three options.

You have agent, you have ask, and you have manual. So manually decide what

have manual. So manually decide what what gets added to the context. So I can say manual and I can go uh at inference.py pi

inference.py pi uh explain this and it will use this model here claude 3.5 sonnet um and there's a bunch of models you can have so it can auto select the models it can

use thinking ones thinking is where you'll prompt it and it will go into its own little um it will basically in HTML you might have like open paragraph and

then close or like open a div and close a div. Um it will it will do exactly

a div. Um it will it will do exactly this. It will start with like an open

this. It will start with like an open think and then a closed think and in between it will literally just spew a bunch of text that is its thought process on what to do. So instead of you giving it the the prompt and then it

going and just immediately trying to solve it, it will actually think out out loud to itself on how it should approach this problem the best. Um so models like

DeepSeek R1 deepseek actually does this and you can see that chain of thought.

Um, and then after it's done the chain of thought, when it has the close think token, um, when it like ends, it will then actually start to give you the the response that you want it. So, be like, okay, based on our thinking, this is

what we're going to do. And then it will literally just give you the code or or whatever task you uh ask it to complete and it will it will do very well at like math and reasoning and even coding

tasks. So, I usually enable thinking on

tasks. So, I usually enable thinking on models. Um, but there's only some

models. Um, but there's only some thinking models that are available. So

like Gemini 04 mini 03 um pretty expensive actually but yeah and you can change the models if you actually go to

settings. Uh so you can change I can use

settings. Uh so you can change I can use like say deepseek or um you know I can use Gemini flash or Gemini Pro or GPD

4.1 or whatever models and uh you know at the bottom of course if you want to use a separate one like if you're using say uh Grock right like very very fast

inference you go in here and you can say like use uh this model say uh what's the most difficult math problem that you know in number theory

and it will stream. It will think and then it will and then it will end. You

can actually see it spew out text ridiculously fast. This will go at like

ridiculously fast. This will go at like 275 tokens per second. So, it's actually ridiculous. And you might ask like how

ridiculous. And you might ask like how do I actually add Grock, right? I can

take the product and take like the the platform and I can if I want to go and say like add in a model here. Uh if we go to like documentation and I want to

like say take um I don't know llama 3.3 versatile or maybe uh what's a what's a good one maybe like deepseek right if I want to take a

distilled deepseek model and I want to put it into cursor to like talk really fast how do I do that and the answer to this question is you would go into this openai API key you can override the base

URL and you would slide the gro API URL in there or like open router for instance Um, and open router is also another tool that I'd like to quickly show you as a

part of this whole AI workflow.

Um, which basically you can you can access access all models from here. Uh,

so if I wanted to say access um, I don't know R1, I go deepseek R1 uh, and I could find like which providers are doing what. So let's say

uh, Deep Infraurbo will offer this model at 300 tokens per second.

Um, and so I could go to like API and then I would take this base URL, see OpenAI, then the base URL there, that's what you would paste in. So when we go back to cursor, you'll see literally

this open router.ai/

API/v1 and that will be exactly here. So

I literally just copy and paste that into cursor and I can talk to any router on any model on open router that I want to and then of course my API key in there from open router because the

models aren't free. Um, but that's a that's a you know pretty solid way to speed things up. Open router is great and it just offers everything. You go

Grock or you can go like I don't know I think Grock is one of their providers actually. So go to providers on like a

actually. So go to providers on like a distilled model you'll find uh Grock there. But if I wanted to say like use

there. But if I wanted to say like use Grock 3 um you can you know find it here from XAI right and pretty solid. But uh let me

breathe. Okay, let me give you a demo in

breathe. Okay, let me give you a demo in cursor here.

Uh let's take which model should we take?

We can go to let's use Gemini 2.5 Pro at relu optimize this relu kernel via thread corsening.

I need to change that to if I go down here and change this to uh pro-experimental uh

pro-experimental and it will start thinking to itself inside of cursor planning next moves and this is the actual thought process itself and this model runs quite fast obviously because it's hosted on Google and then it will

port in this new and you can sort of just apply and it will you can say like command enter and it will make those changes or I can go like control zed and then I can reject it. Um, but you can generally move like very fast when

you're just you can see how fast I was moving and Gemini 2.5 Pro is actually the best model best coding model in the world right now. Uh, and it is this fast. So when you're talking to it and

fast. So when you're talking to it and you're making these rapid changes um to like say very important like kernel level GPU code uh you can you can you

can work wonders really. Um so this is cursor um hope you like it. I use cursor a lot.

Uh and then there's uh this other one that I want to introduce to you called chorus. So chorus is where I sort of go

chorus. So chorus is where I sort of go off to plan stuff, you know, using using this groi is like it's okay and it works but and it has like a memory feature and

deep search and think and all these modes. But what it does not have is uh

modes. But what it does not have is uh some of the things that chorus has. So

if I go ask and I say expand this and I make a new chat, right? I can I can like look at the side panel there with commandB a little keyboard shortcut. I

can go into uh here I can select models.

So I can go like Gemini or I can go GPD4.1 as you can see there. I can go to say like club 3.7 sonnet and I can be like

sup how can I assist you today? Right?

And it will just take your API keys and it will use those. It doesn't you don't have to pay for anything in course.

completely free and you just use your API keys. You It barely costs any money

API keys. You It barely costs any money to talk to models like this. This this

probably costs like I don't know 0.001 cents to talk um this this specific request like very cheap. Um and you can like branch out. You can use uh what do

you call it? Let me let me reopen course. It's a very new sometimes it is

course. It's a very new sometimes it is a little buggy but I don't mind for how much features and how simp how simple it is. So, I can like open a new chat and

is. So, I can like open a new chat and then I can go say brainstorm. I'll say

like give me ideas on how to plan for a party on uh on May the 4th and I can enter and it will uh

brainstorm and spin up a bunch of random a bunch of ideas here in in parallel. Uh

and it's sort of like really cool to watch. Uh and then you could go to say

watch. Uh and then you could go to say uh May the fourth. Yeah. And then you can yeah give it a second. You can go to

like say uh compare and then I'll use Gro 3. I'll use

Gro 3. I'll use GPT4.1. I'll use

GPT4.1. I'll use um I'll use maybe a new llama 4 model.

And then we'll let these all run on their own. Um let's see.

their own. Um let's see.

How would you approach optimizing uh a floating point 8 matrix multiplication kernel on hopper GPUs from scratch? No o credentials found. Okay,

scratch? No o credentials found. Okay,

so I need to put my API key in there, but you get the point. And so all the models will just kind of go off on their own. Uh and they will just so you can

own. Uh and they will just so you can just watch these responses in parallel.

You can stop one of them, you know, stop this one, maybe copy that response or make this one bigger, uh, etc. And it's just it's just great for just like everything. Uh, and you can I can just

everything. Uh, and you can I can just like quickly Oh, wrong key. I can just make a new chat like that. Command N,

command K. I can like navigate these and go like settings and, you know, API keys are laid out. Uh, appearance, you can like change the font and everything.

Ambient chat is another feature I like.

you can option space and then you can say like sup hey not much what's up with you and then you can like close it with escape.

Um so it's just kind of it's just kind of like nice to have these. So instead

of command space for raycast I'll go uh option space for uh for chorus and just make a new chat with command n like this and uh yeah that's that's that's that's

chorus for you. Great. Um, and then, you know, last but not least, I use the Brave browser. You know, everything is

Brave browser. You know, everything is nice and secure in here. It's uh, you know, it's also it I mean, it is based on Chromium, so it's not the most memory efficient thing in the world, but uh,

it's pretty good uh for for what you get uh from Brave. Exactly. So, yeah, that's uh hopefully I haven't clicked off at this point. You probably have, but this

this point. You probably have, but this is just my sort of rant about all things I do. You know, one thing I would also

I do. You know, one thing I would also recommend is actually getting a good sleep at night. So, if you're staying up, if you're staying up all night, uh, and just like watching this video, like, go to bed. Literally go to bed, wake up

in the morning, um, watch the video again, or just literally, you know what?

take take one of my take one of my stupid videos and literally transcribe the whole thing and then just feed it into chorus and have it have like eight models just go in parallel and attack it

and say what's the most important thing here and then say um and then you could take you could copy all of those responses and say what do the models agree on is the most unique thing from this video and then you could just you

could be so fast and creative with things it's unreal but anyways I am going to stop ranting now. Um, and I'm actually going to go to bed soon. But,

uh, yeah, thanks for thanks for tuning in again. Follow me on X, you know,

in again. Follow me on X, you know, really quick. Follow me on X at

quick. Follow me on X at um at Elliot Archer. Um, you know, I'm kind of famous a little bit, but uh, yeah. Cheers.

yeah. Cheers.