00:00What's going on? What's going on Zinus Nation?
00:02So yeah, we've made all those Mac Studios, all those pictures with Mac Studios.
00:05Yeah, we bought all those for all of our front desk workers and the people in the office.
00:10And we might add more people to the office, maybe not.
00:12We'll see what the office turns into.
00:14But I wanted to take a second, because now I feel like we almost haven't done enough, yeah, to talk
00:18about it.
00:19And I wanted to see, let's see, let's see, what are some uses for a Mac Studio for an M5,
00:32M5 Mac Studio, Mac Studio with 512 gigabytes, gigabytes of unified, five memory.
00:48Let's see, okay.
00:57Let's try a different account.
01:01I like talking to Grok, but if Grok's not going to work, we'll do ChadGBT.
01:15All right, Grok's not going to work, so we'll use ChadGBT.
01:32Let's see.
01:33If you're talking about the M5 Ultra Mac Studio with 512 gigabytes of unified memory, that's an extreme workstation.
01:41Apple specifically positions the 512 gigabyte configuration for enormous local AI models and the most demanding professional workloads.
01:54It has, right, okay.
01:56Here are some things I'd consider genuinely interesting uses for it.
02:00Run very large AI models locally.
02:03This is probably the most unusual reason to buy 512 gigabytes, right?
02:07You could run a large open-weight LLM locally, whatever that is.
02:13Right.
02:15We'll just read it out to see if it matters before the Lord.
02:18You could run large open-weight LLMs locally using things such as MLX slash MLX dash LM or LM Studio
02:27without sending prompts or documents to a cloud service.
02:30Apple specifically says the 512 gigabyte M5 Ultra can run enormous LLMs entirely on device.
02:39For perspective, very roughly.
02:49We'll just skip that part.
02:54We'll just skip that part.
02:57The enormous memory headroom means you can accommodate the model plus context, KV cache, operating system applications, and other models.
03:05You could even have multiple AI models loaded simultaneously.
03:08So you could run, like, Grok on your computer, right?
03:12Whatever Grok is.
03:14Something like that?
03:15Is that a thing?
03:16I don't want to say, like, that makes sense.
03:17Like, I'm explaining something I understand, right?
03:19Now, anyway.
03:21Local image generation.
03:22You could use it as a serious local image generation workstation.
03:28For example, stable diffusion variants, flux family models, comfy UI, draw things, custom image generation pipelines, image upscaling, AI image
03:38restoration, batch generation.
03:42Apple specifically cites up to 4.3 times the AI performance of M3 Ultra and up to 8.2 times
03:48the text-to-image performance of M1 Ultra for the M5 Ultra.
03:53Text-to-image, right.
03:55Yeah, so you could generate images like I do on ChatGBT, but right there on the...
04:01Okay, serious video production.
04:03This is another area where the machine would be ridiculous.
04:06You could work with 4K multicam, 8K footage, raw video, ProRes, heavy effects, color grading, motion graphics, AI video effects,
04:14large After Effects compositions, 3D rendering and VFX, AI-assisted software development, huge data sets, and data science.
04:29Run lots of virtual machines, right?
04:34Want to know something else?
04:36Yeah.
04:37Does Grok still use the Super Colossus?
04:43Right.
04:45Colossus computer?
04:47Right.
04:54Yes, Grok still uses the Colossus supercomputers, but the situation has evolved considerably.
05:00As of October 2026, XAI says its Grok models are trained using its Colossus 1 and Colossus 2 infrastructure.
05:07In January, XAI said Colossus 2 was operational specifically for Grok, and XAI's own January 2026 funding announcement said its
05:16Grok 4 series was built on training infrastructure power, but...
05:19Colossus 1, the original 200,000 GPU cluster.
05:23Right.
05:27Wait.
05:31How does the Grok...
05:36Let's look at another one.
05:38What is a brief summary of how the Grok...
05:45Right.
05:46Of how the Grok...
05:48Grok...
05:50Colossus...
05:51Colossus is really interesting stuff.
05:54Supercomputer...
05:56And the computers...
06:00Gemini...
06:02Gemini use...
06:04And the computers...
06:06And the computers...
06:07And the computers...
06:11ChatGVT...
06:12TUs...
06:13Stack up.
06:15Stack up.
06:17Okay.
06:18This is a fun way to use AI, I think.
06:21Very roughly, all three are on the same league of AI supercomputer scale, but they take very different approaches.
06:28Main hardware...
06:32Grok uses NVIDIA GPUs, right?
06:35Right.
06:37Big infrastructure...
06:38Wait.
06:39Okay.
06:47Should I keep reading?
07:03The TPU infrastructure was interesting.
07:07I read about that.
07:07Google says its new TPU infrastructure is designed to scale.
07:10Because they use like computer...
07:12What is it?
07:15They're like building...
07:16They make like...
07:17Pods...
07:18Like computer pods...
07:19And they make building blocks to make bigger and bigger computers.
07:22And they're like called TPUs or something, if I understand that correctly.
07:30Whereas...
07:31The...
07:31The Colossus 2 is one giant...
07:34One giant computer.
07:36But Google uses a whole bunch of smaller computers to build a giant computer.
07:40Yeah.
07:40I thought that was interesting.
07:43Anyway, yeah.
07:44I think that...
07:45I think our brains are fried for more of this.
07:47It's no-stop.
07:47Okay.
07:48Sure.
07:49Do it.
