Transcription
Okay, close your eyes and think of the last person on Earth you would expect to create an open-source AI project. No, no, not him. I mean this guy.
>> Hey, how's it going, bros? My name's PewDiePie.
Yeah, he built something called Project Odysius. And in this video, I want to take a look, learn how to install it, and see what it can do. But first off, let's break down what it actually is.
It's basically like a self-hosted AI workspace, kind of like trying to recreate the chat GPT or Colot experience, but running on your own machine, like with your own models and your own data, and a lot more control over what you can and can't do with it.
Now, this isn't like a brand new AI model made by PewDiePie, although he did train one of those in one of his videos, too, but that's a video for a different day. You're not getting like PewDiePie GPT. Instead, Odysius is the interface around the models. Think like the Claude desktop app or the Chatg GPT desktop app. It's kind of like that, but a version that he developed.
It can connect to local models, API models, and tools. And it's trying to pull a bunch of AI workflows into one place: things like chatting with models, running agents, working with files, doing deep research, comparing model outputs, writing documents, managing notes and tasks, and even connecting things like email and calendars.
A lot of people want the power of AI without sending everything to big cloud platforms. Odysius is exciting because it points towards the idea of like your own AI workspace that can remember things, use your files, work with your tools, and live on hardware that you control.
So, tools like this are cool, but I do want to set some expectations before you go and install it yourself. It's going to probably be kind of janky, and it's going to definitely have some bugs. And unless you have super good hardware and really good model setup, it's probably not going to perform like Chat GPT or Claude or Gemini, like right out of the box.
So, think of it like this: Odysius is not this polished mainstream AI assistant yet, like you're going to get out of the big labs. It's an open-source experiment for people who want more control, more privacy, and a glimpse at what personal AI might actually become in the future.
Now, with all that being said, let's go ahead and install this thing and see what it can do.
So, PewDiePie put this whole thing up on GitHub here, and it hasn't been live for that long, but it's already got over 71,000 stars and over 9.2,000 forks, like people sort of branching it off and building their own version of it.
But if we scroll down on the page here, it should show us how to actually set it up. So, they've got a page here called 'setup guide'. So, we'll go ahead and open this page.
And we've got instructions for Docker, which is recommended. You've got Linux and Mac OS. You've got Apple Silicon. This is probably the version I'm going to go ahead and set up, 'cuz I am on an M3. And then you've got Windows setup down here. The setup process should be fairly similar for each one. But let's go ahead and do this one here.
I'm going to open up my Finder here. And then I have a folder inside of my computer called 'Code' where, pretty much, this is where I test all sorts of stuff. I'm going to go ahead and create a new folder in here. We'll call it 'Odysius'. And I'm going to open this folder.
And then I'm just kind of doing this like the real simple noob way. And I'm going to right-click on Odysius and I'm going to open in terminal. And this will open it directly in that folder that I just created. So, this is my terminal here.
Now, if I copy the code that it gave me, I'm going to just paste each one of these in one at a time. So, we'll go ahead and `git clone` the repo straight into our Odysius folder. It actually made a second subfolder called Odysius. So, I'm going to go ahead and change directory into that Odysius folder. And then I'm going to run this command: `/start Mac OS.ch`. And it's downloading everything it needs to download and running the process.
Okay, now it's asking me for a username. I'm just going to set it as 'admin' to start. And it's asking me for a password. I'm going to go ahead and enter something here. And cool. It's opened up to a page here asking for my username and password that I just created. So, let's go ahead and enter these, and I'm in. This is the Odysius platform here.
Recraft recently launched their V4.1 family of models. If you haven't heard about them before, they're a professional AI-native design platform for generating images, illustrations, logos, and vectors.
So, V4.1 is supposed to feel more natural and photorealistic, and the model will understand your aesthetic from just a few words. Now, here's a few examples of some images generated with a Recraft versus other top models. The Recraft images just honestly feel more human. Like, the backgrounds don't look like busy stock photos, and these people don't have that AI look, like the midjourney image. And this toothbrush photo, the Recraft one, just looks way more professional.
And there are a few different models in the lineup depending on what you're making. V4.1 is the faster model for creative exploration and everyday concepting, while V4.1 Pro is for higher-resolution images when you need more detail. And then there's also V4.1 Vector, which is probably one of the most interesting models because it generates real, editable SVG files. So, if you're making icons or logos or illustrations, you're not just getting a flat image that looks like a vector; you're getting something you can actually edit in tools like Figma or Illustrator, and it'll come out exactly how you envisioned it.
They also added V4.1 utility, which is for when you don't need all the drama, just clean, simple, and predictable images like flat product shots, front-facing compositions, and simple scenes. The big idea here is that Recraft is trying to move beyond 'generate me a single flat image' and more towards AI images you can work with and perfect. It feels more aimed at people who actually need usable design assets for branding and websites and marketing campaigns, product visuals, presentations, and, you know, things like that.
If you'd like to learn more and try it out for yourself, there's a link down in the description. And thank you to Recraft for supporting this channel and sponsoring this portion of today's video.
All right, looking through the sidebar, we've got 'new chat', we've got 'search', we've got 'email', we've got 'brain'. Okay, so this is where the long-term memories and skills sort of get saved.
We've got a 'calendar' in here. We've got a 'compare'. So, we can actually do like our own blind comparisons of models, it looks like. So, you can almost do like your own arena leaderboard kind of thing for models. It seems like the 'cookbook' looks like this is where we can select which models we want to use, install various dependencies. We've got a 'deep research' module in here. We've got a 'gallery' for photos and stuff. Our 'library' where we can actually import documents and things. 'Notes': this is your basic to-do list and also where reminders are managed. And then 'tasks': so, this is like recurring sort of cron job loop tasks.
Here, we also have 'chat' and 'agent' down here. Not quite sure what the difference between the two is. But if we go to 'select model', you'll see we've got no models connected. So, if I hit plus, we can add API models. We can actually send chat straight to a server and let, you know, Anthropic or OpenAI or DeepSeek or Gemini or Grock or any of the models actually be the models that respond by just giving it an API key here. Or we can set up something like Olama and then use local models.
So, we'll go ahead and download Olama here. So, if I just go to olama.com, download for MacOSS. Then, once it's downloaded, we'll go ahead and just run it. Install Olama, and then we'll fire up Olama in the background here. And now I can see up in my top bar that we have Olama running.
So now, if I come back over to our Odysius chat, add Olama. And cool. Now it found Olama, but there's no models. So now, if I go to the 'cookbook', I should be able to download models here, right? Let's go for this Gemma 3 12 billion parameter model. Let's select that one.
All right. So, it finished downloading after about a minute here. So, if I close out of this and select the model, so now it's serving Gemma 312B. I wonder, wonder if it knows itself.
'Explain Odysius to a non-technical person in 90 seconds. Use the analogy of a personal AI control room running on my own computer. Then give me three reasons someone might prefer this over chatbt and three reasons they might not.'
What do you bet it doesn't know thyself? Um, okay. I guess it does. That's surprising. I didn't know if it would actually answer details about what it itself was.
It says, 'Odicius is different. It runs locally on your own machine. Can access and understand your documents, emails, and even your calendar. It's not just answering questions.'
Pros and cons. Three reasons to prefer Odysius: Privacy and security, offline functionality, personalization and integration. Three reasons why you might not prefer it: Needs computational resources, setup and technical knowledge (I mean, it was pretty easy to set up, honestly), model size and capabilities.
'Let's see who made Odysius.' What? 'Odysius was primarily developed by a team at Together AI. They're a company focused on open source large language models.' All right. So, no. I guess it doesn't really know really what it is. I think I probably led it too much with my original prompt, but this all happened offline. This was not actually sent to a server.
Let's give it some of the old-school tests here. And this is more of a test of Gemma 312B than it is Odysius itself. But what the heck?
'How many Rs are in the word strawberry?' Okay, so it got that right. Um, what about this one?
'I need to wash my car. The car wash is 100 yards from me. Should I walk or should I drive?'
'Given the car wash is only 100 yards away, walking would likely be the better option. It's a short distance, and you'll save on gas and wear and tear in your car.'
All right, that again, that's more of a reflection of the demo model than Odysius itself.
If I came into our settings and added an API model, like, let's go ahead and add an OpenAI model here. I'll jump over to `platform.openai.com/ap openai.com/api keys`. I'll create a new key here, and I'll just call this one 'Odysius'. I'm going to actually delete it after this video, so I'm not too concerned. We'll create our key. Copy it. Jump over to Odysius chat. Paste in our API key and add that in.
If I reload again here, I now have all of the OpenAI models available to me because I just plugged in that API key. So, if I was to go and use, you know, GPT 5.5, ask the same question. We get a much better response.
'Drive. If the goal is to wash your car at the car wash, you'll need the car there.'
I like that it's telling me the cost, too, because this did call the API. This one wasn't run locally. And it's telling me that it used, you know, 210 of a penny, I guess that is.
This might actually bog down my recording, but I actually want to see what happens if I put like a really, really strong model that's local on here. So, let me go back to the 'cookbook'.
I kind of want to test this Quinn 3.5 122 billion parameter model. Now, it does require 77 gigs of VRAM. I am working with 204 gigs, apparently. So, let's go ahead and test that one. It's got a 98.5 score, meaning, supposedly, it works really, really well with this app. Looks like it's going to take quite a while at 76 gigs. So, hopefully, if I close this, it'll keep downloading. Let me close out of this and double-check. Looks like the download's still going. That's good.
If I open up with a 'brain', let's see what it knows about me.
'The user owns a car. The user is interested in personal AI assistance. The user has sensitive information they need to protect.'
Interesting. These are all details that it pulled in just from my past chats. I don't know what this is referencing based on these chats, but I mean, it's essentially doing the same thing that like the earlier versions of chat GBT and COD were doing with memory, where it just kind of remembers details as you chat, so that in future chats it can pull that information back in.
All right, let's try the 'compare'. Let's use GPT 5.5 as one, and then we'll use the Gemma 312B as the other one, and we'll do them blind and parallel. And let's go ahead and start, and let's ask it a question like: 'Is local AI actually useful for normal people yet, or is it still mostly for hobbyists? Give me the strongest argument for both sides, then give me your honest conclusion.'
Let's see if we can figure out which model's which. It should be pretty obvious. The one on the left's most likely the logo model just because of the speed. The right model is going a lot more in-depth, which to me screams GPT 5.5. I also think this: '$11 per 1,000 tokens'. Wait, that can't be right. Oh, that's $11 for 1,000 responses like this. This one's 9 cents for 1,000 responses like this. So, this one on the left is obviously the local Gemma model. This one on the right is obviously the GPT 5.5 model. Let's see how they actually responded, but you can pretty much tell which model they are without even needing to read a word.
All right, so let's break down the usefulness of local AI for everyday people. Here's the strongest argument from both sides, followed by an honest conclusion.
Strongest argument for local AI is privacy and reliability, with increasing concerns about data security and the potential for cloud services to be disrupted. Strongest argument against it being generally useful right now is the significant performance and resource limitations. I'd agree with that. Yep.
Honest conclusion: Right now, local AI is trending towards usefulness for normal people, but hasn't fully arrived yet. It's currently most attractive to a niche of tech-savvy users, those comfortable with a bit of technical configuration who are deeply concerned about privacy or need offline functionality. For the average person who just wants an AI to help with basic tasks, cloud-based AI solutions currently offer a smoother and more reliable experience.
All right, I'm not going to read this whole thing 'cuz this is just, uh, very wordy here, but yes, but unevenly. Local AI is already useful for some normal people, but it is not yet universally better or easier than cloud AI. The honest answer is it has crossed from mostly hobbyist toy into practical niche tool, but not yet into default mainstream assistant. For ordinary users, local AI can already be genuinely useful for summarizing personal documents, searching through notes, PDFs, emails, journals, or work files, rewriting text, drafting emails, brainstorming. So, it goes into a lot more depth of all the stuff that you can actually use it for.
The killer feature: I can ask questions about my own private stuff without sending it somewhere else. And that's probably the main reason. It's going to be people that want privacy without it going to a cloud. People that want it as cheap as possible because you're not paying per tokens or API fees; you're just paying for the power usage, essentially. And people that maybe want to work offline completely.
Arguments against it: just a lot more steps. It looks like a friction problem. And there's a quality gap. There's definitely quality differences between the two. Local AI is useful now, but mostly for people with one of these motivations: Privacy matters to them. They work with lots of personal local documents. They want offline access. They dislike subscriptions or cloud dependency. They're mildly technically or willing to use a polished app. They have capable hardware already. For everyone else, cloud AI is still usually the better default.
To be honest, I feel like they're both equally good answers. This one's just way more wordy and gives a little bit more details. I'm going to actually say tie because, like, they both have their pros and cons, right? Like this one, it got to the point quicker. We know which is which. Yeah. Gemma 3 is the one on the left. GPT 5.5 is the one on the right. No surprises there. The compare tool is pretty cool, though. It's like you could have your own arena.
What I wonder is, does it rank them anywhere? Can I actually like see how I've ranked over time? That would be kind of cool. I don't believe that feature exists yet, but it would be kind of cool if you started built out your own personal leaderboard of which models you like better just by prompting and testing constantly.
This is interesting. It even gives you some eval prompts that you can test. So, if I was to do like 'draw SVG' and give it to both of them, the one on the left finished first, which I'm guessing it's kept it as Gemma, I'm assuming, but let's go ahead and copy this. I'm going to jump over to this site called 'HTML online viewer' here. And I'll just paste in the first one. And here's the SVG that the left model made.
And then if I copy this one, which again, you could tell by the price here and how many tokens it used, this was definitely the GPT 5.5. Again, let's open a new tab here. That way we can compare. And I'll paste this second one in. Oh my god, that is a huge difference. So, this is what Gemma 12B made. This is what GPT 5.5 made when we asked it to make an SVG. Yeah, there's quite a large gap between capabilities, I'd say. All right, I think, uh, I think B1 on that one.
Oh, the model comparison's fun. All right, let's try, uh, let's go to 'deep research' here. So, this is interesting. We have a bunch of settings for how we want it to do the deep research. We can give it a certain amount of rounds we want it to run. So, let's do, I don't know, five rounds. For format, I guess I'll leave that on 'auto' because I don't really know what 'product' means. Search engine, so we can have it search for us. Guess we'll just leave it as default endpoint.
I want to use Gemma 'cuz I want to see how the deep research works when we're doing local models, 'cuz the whole point is, if I want to do deep research with OpenAI, I would just go and use the chat GPT deep research. So, how does it do deep research when it is using a local model?
Let's see: 'How does Odysius from PewDiePie compare to agents like Open Claw and Hermes?' And I'll leave it with these settings, and let's see what it does.
All right, so it's doing some research here. It's planning a strategy. It's on round one. I imagine it should be pretty quick because so far Gemma 312B has been quick every time we've prompted it. It's just going to go through multiple rounds.
To me, this could be the ultimate way to use something like these local models: is to actually use a local model like Gemma 312B, ask it your prompt, and just let it take its time responding, right? Like, it could go through however many rounds. Like, let's say I set it on 10 rounds, it does all the research, and I guess I'm assuming it prompts it like 10 different times and then comes to a consensus or something.
I'm actually not quite sure how the rounds works exactly. There's a little question mark here, like I would assume would there'd be a tool tip pop up, but I'm not seeing anything. So, I don't, it doesn't really explain what the rounds are, but my guess is that it's running this prompt five different times and then kind of coming to a consensus at the end of all of the times it ran it, and you'd get a really good response out of that.
So, if you're not too time-constrained and you want like a really good answer, but from a local model, using the deep research with the local model could be the way to go. Of course, you would have to be online because it does do some, you know, web search and stuff, but it's not sending any data to a cloud anywhere other than your search query.
Okay, so it finished. It took about 7 minutes, and it actually created a visual report. Ooh, look at this! This is actually a fairly nice-looking report that it generated. It even has like a table of contents and all that. This is actually much, much better than I thought it was going to be. And this did the whole thing with the local model.
All right. So, 'emergence of autonomous AI agents', blah, blah, blah, blah. 'Odysius aims to be a comprehensive self-hosted AI workspace integrating chat agent capabilities, research tools, and productivity features. Openclaw functions as a versatile gateway and orchestration layer. Finally, Hermes focuses on development of a personalized learning agent, prioritizing long-term memory and self-improvement.'
Odysius's primary strength lies in its holistic approach. It combines AI chat, agent capabilities, research tools, notes, task management, calendar integrations, and local model management into a single unified interface. 'Model control plane, aka MCP.' Well, MCP stands for 'model context protocol'. So, I guess that's a point against the response here, but the MCP is a crucial architectural component enabling agents to interact with external systems like Gmail, Google Drive, and GitHub. This allows agents to perform actions beyond simple text. So, I mean, what an MCP does, it described well what an MCP stands for. Yeah. 'Model context protocol' not quite the right thing.
>> Just just come can you just come on again this is more a reflection of the model that was being used and not a reflection of the Odysius app.
This output, though, where it's got the table of contents and, you know, all of these details here. It used duck.go for the search. All of this is a reflection of the Odysius app because the Odysius app kind of told it how to design this page. The actual text that's in the response here, that's a reflection of the model, not the app. Pretty impressive, honestly.
If I click on my 'library' here, I can actually see all of my past chats. 'Odysius AI controlled preferences'. I guess that's why it had the history of, uh, you know, 'I've got things I want to keep private'. Uh, 'documents'. I haven't uploaded any documents here. 'Research'. It shows that research that we did, and then I haven't archived anything.
Okay, so it does look like it finished downloading my Quinn 3.51 122 billion parameter model. Let's go ahead and launch that model instead. Or does it run them both? Okay, so it looks like I can only run one at a time. So, let's go ahead and stop and launch that model.
All right, I'm actually very curious about this SVG thing here. So, let's go back to 'compare', but this time we'll use GPT 5.5 again, but I want to see it use our new Quinn 3.5 model. Oh, there is a scoreboard! Okay, we'll check that out in a second. But I wanted to do this SVG thing again.
Okay, both models are moving pretty fast. So, the one on the left finished, would cost $35 to do 1,000 responses like that one. And it used 4,274 tokens. Let's go ahead and copy this into our HTML viewer. Okay, that's got to be GPT 5.5, right? 'Cuz it's like almost identical to the last time GPT 5.5 gave us something back.
And then here's our other one, which says it would be $5 to generate 1,000 like this. So, this has got to be our Quinn model. And let's go ahead and copy this. Let's see how Quinn does with SVGs. I mean, not horrible compared to our Gemma model earlier. It's quite a bit better, right? Obviously, nowhere near what GPT 5.5 is capable of with SVGs, but I, it's, it's decent. This is all local.
All right. Now, I need to go back to this 'compare' here. Okay. So, we know A was better. So, I'll go ahead and vote A on that one. And of course, yeah, that was a GPT 5.5.
All right. Now, I want to go and check out that scoreboard that they showed earlier because I'm curious what that looks like. So, if I go to 'compare' again and click on 'scoreboard'. Okay. So, we can see it's just a list like this here. So, GPT 5.5 has won two and tied once. Gemma 3 has lost once, tied once, and Quinn 3.5122 has lost once, but obviously we haven't done a lot. So, pretty cool way to like test a bunch of models against each other.
Let's see. There's a bunch of other stuff in here that I haven't tested yet, like the 'calendar'. I don't really know if this uses AI. I think it's just like a calendar that's an alternative to using something like Google Calendar or iical or something like that. I think you can just add your own calendar events in here. It's not really an AI feature; it's just a calendar inside of the platform.
You've got a 'gallery' in here where you can upload and organize photos, save them into albums. Again, I don't really think this stuff here is specifically like AI features. It's just a place other than Google Photos or, you know, Apple Photos or something like that where you can store and organize your photos and do it on your own software that's not one of these other companies software.
There's an image editor in here which I think does actually maybe have some AI features in it. There's also some AI tagging which you can connect a vision model and, you know, tag images and stuff. But let's upload an image into the gallery here. Here, I'm just going to upload one of my recent thumbnails here. And then let's click on 'edit'. And yeah, we've got like a little mini editor here that I can go full screen on. And it does have like inpainting and stuff, which I'm guessing is just built straight into the app.
Let's see. If I go to 'cookbook' and then I go to 'dependencies', there's some options here like 'remove background' for the image editor. So, I can install that. There's a 'dnoising'. So, if I install that, and then let's install the transformers here.
>> Two hours later.
Okay. So, I've been trying for probably the last 45 minutes to try to actually get the impainting working inside of this image editor, and it's just beyond me. I, I don't know why I can't get this working. I even got it so it shows an image model. I have Flux Klein set up here. Um, when I give it the impaint prompt, like 'add text hello' and click 'generate', I get this error: 'Inpaint failed rejected endpoint URL disallowed address one'. I don't know why I can't get it to remove what I highlighted here.
I got the background remover to work. If I click on 'background remove' and I click 'background remove', it kind of sort of removes some of the background, but I can't get this damn impaint to work. So, what I've tried so far was I've gone to 'cookbook'. I downloaded the Flux Klein 9 billion parameter model. You can see that I even have it running right now under 'dependencies'. I've installed the diffuser and transformers here. But whenever I come into like edit an image and I try to use this inpainting feature, I just cannot get it to work. And I don't know what I'm doing wrong. And I've spent way, way, way too much time messing with this.
So, I can like fill this image here and then say 'change text to say hello', 'generate inpaint failed'. 'No image generation endpoint configured. Serve a diffusion model.' Okay. So, I don't have the model selected. So, select the model 'flux', 'generate inpaint failed'. 'Rejected endpoint URL'. I don't know. I've spent way, way too much time on this, and I'm kind of giving up on trying to get image generation to work in this. I'm sure there's something obvious that I'm missing.
I've actually tried two other models. Uh, I tried the ideagram model 'cuz that was available on hugging face. And I tried the flux 2 dev model 'cuz that was also available on hugging face. Same issue every time. So, I, I, I'm lost on getting this image, uh, in painting to work. Same happens with outpaint. I can't get it to outpaint either. So, I don't know. AI image generation. I'm just kind of, uh, giving up on that.
As far as some of the other features, you've got a 'notes' feature. Again, I don't really think this is much AI. You can just add a to-do list. So, it's like a, you know, free open-source version of like a to-do list or something where, uh, you know, I can add a bunch of to-do items here and then check them off as I complete them. It's got a 'tasks', which isn't the same as to-dos. This is like recurring sort of cron job tasks, things that you have set on a schedule.
And then another thing that I didn't really test was the 'agent' feature. And the agent feature is essentially giving Odysius access to your computer and tools. So, like I said, it's got an agent feature. I can't figure out how to use it. It's got an image editing feature. I can't figure out how to use it.
But saying all that, I'm pretty impressed with some of what Odysius can do. It's cool that it's got this 'brain' memory, so it will remember a lot of your chats and then ideally remember some of the stuff you've talked about in the past. I like the 'compare' feature where you can create your own sort of internal leaderboard of which models you like better and sort of test a bunch of models over time and figure out what suits you best. I think that's actually really, really cool.
I really liked the 'deep research', especially the page that it output. This visual report that it designed is really, really cool-looking, even if, you know, it's got some inaccuracies. Again, the inaccuracies are more a reflection of the model than the Odysius platform. You've got your 'gallery' here, which, you know, it could replace like Google Photos or something for you, but I can't get the AI editing to work at all, really. And it's got notes and tasks and that sort of stuff, which I don't know if I'd ever really use that much myself. It's kind of more if you want to get off of like Google's platforms or something like that. And I'm kind of fine just using my to-d doist. So, I don't know if I'd ever actually use this feature.
They do have some other themes. So, if you want to change the colors to like 'cyberpunk' or 'terminal' or 'ocean', those actually look kind of cool. You know, I'm partial to cyberpunk, but again, this is a tool for tinkerers, people that don't mind sort of getting their hands dirty and, uh, maybe pulling their hair out sometimes trying to figure some stuff out. I'm sure I could get the image generator to work and I could figure out the agent if I wanted to spend a lot more time on this, but I've got stuff that works right now. Like, I can use Olama directly and chat with models locally directly here inside of Olama if I want. When it comes to generating images, I can use something like Comfy UI. Although that's kind of a pain in the butt on its own as well. But more than likely, if I do want to do stuff with images, I'm going to go use Gemini, the nano banana platform, or I'm going to use OpenAI's image gen.
I feel like Odysius was really designed for people that want to completely get off of all cloud platforms. You don't want to use Google Calendar at all. You don't want to use Google Photos orphos like at all. You use their gallery instead. You don't want to use a tool like to-d doist or any sort of notetaking tool. You want everything to be off every sort of cloud platform. That's kind of what Odysius is for. And you can use local models. None of the local models are going to be as good as what you're going to get from OpenAI, Anthropic, Google, Grock. Those models are all going to outperform any of the local models that you put on here. But if you're doing like real basic stuff, it, it'll do the job.
I am very, very excited to see more and more stuff coming to local computers. So, I'm excited to see that we're getting large language models and image generators and things like that that are working better and better on local computers. Unfortunately, I mean, right now, you need a pretty solid computer to run any of the really good models, or you're going to run a model that's, you know, quite a bit dumber. And I just don't feel like any of the local image generation models are that great yet. Ideogram seems like it's probably the best image generator that you could run locally right now, but I couldn't get that one to work in Odysius. I tried and it just didn't even recognize it as an image generation model.
But I am really, really excited to see more and more of these models come locally. And we're starting to see more and more computers that can run these models locally. Microsoft and Nvidia recently announced their DGX computers, which are designed to do more and more of this stuff locally on your computer. So, it will be cooler and cooler to see these bigger and bigger and bigger models being able to run locally where we are less relying on API cost and sending things to the cloud, and we can run models offline. I'm really excited that we're finally moving into that direction more and more. And seeing PewDiePie make something like Odysius to sort of lean into that is also really cool because I think it's bringing some more mainstream awareness to the fact that you can actually do stuff like this.
So, all really, really cool stuff. But I wanted to make a video where I played around with Odysius and put it through its motions and tested it a little bit and see what it was capable of. And it's impressive in some areas and leaves a little bit to be desired in other areas. But, uh, that's what I got for you today.
If you enjoyed videos like this and you want to see me test and break down more tools in the future, like this video and consider subscribing to this channel. Also, every Friday I make these AI news breakdowns where I break down all of the news that happened in the world of AI for the week. I drink from the fire hose all week and keep up with all the news so you don't have to. You just tune into one video on Friday and get the whole lay of the land in the AI world. So, if you like that type of video, that's also a good reason to subscribe to this channel.
But that's what I got for you today. Hopefully, you enjoyed this one. Hopefully, you learned something, and I'll see you in the next one.