6:06
Exciting news. I am releasing another project, Ajax. It's a reference to that cleaner spray. Ajax is a fine-tuned AI model. Just what the world needs right now. Ajax can browse the web, do private searches, write your emails. My lawyer has also advised me to say it is not designed to provide dangerous actionable instructions, for examples, instructions for building nuclear war weapons, and I want to be clear about that distinction at the start of this video. It can also use your calendar. So, Ajax is fine-tuned for the Odysseus harness. What a time to be alive. You're shaking, I know it. What the heck is he talking about?
The second greatest thing about Ajax is the greatest thing about local AI, which is that it's decentered. Now, this process was very simple in practice, and it made me question my entire existence. More on that later. And when making this model, OpenAI banned me. TWICE. THEY'RE SCARED. FIRST, I want to give a shout-out to this month's sponsor, which I promoted before, and I'm very happy that they came back. That means that you guys were actually interested. boot.dev teaches you how to code, but here's the kicker, for real. I love to learn stuff, so I tried all these different learning programs that other YouTubers sponsor, but so many of them, I can't help but go, "And what the
am I learning? Nothing." It feels like fake dopamine hit learning. And other stuff I tried, and it feels too dry. Like, I don't want to do this. This isn't fun. I really strongly feel like boot.dev hit such a perfect balance of you actually learning stuff. No, it's not always easy, but that's kind of the point. It has to be. It's kind of the same as fine-tuning AI. If you just teach it simple things, it's not going to improve whatsoever. In fact, the best things that it learns is when it can get it maybe sometimes. That's how basic reinforcement learning works at least. I don't know about the human brain, but the other part about it is that just like reinforcement learning, you have the teacher helping
you on the side. So, if you ever get stuck, you have all the help you need right there. I used boot.dev's Linux course. Amazing. Highly recommend it. It's so funny to me now that I use Linux so long. Whenever I see Windows anywhere, I get like, "What? People are still using Windows?" Another course I did recently was the GitHub course that is hosted by another YouTuber, primagen, which I really like. This is what this man taught me, okay? He taught me that.
Cuz I hate GitHub. I hate it so much, but I have to use it for this CS. I wish I did it sooner because I had like one of those panic attacks where I didn't understand what I was doing. It became such a mess I almost deleted Odysseus. And that's why I was like, "Okay, I should probably do the GitHub course." So, now I feel so much more confident, and I'm so thankful for boot.dev. Maybe you think, "Oh, well, AI writes code now. Why would I learn how to code?" Companies need people that can read and understand code. So, for you to have these fundamentals is more important than ever. Yes, developers are using AI to write code, but you still need to understand and be able to read and think
like an engineer yourself. Even Odysseus, you can help contribute to. Anyone is able to do that. So, give it a shot and learn and try it out. I highly recommend it. Use code PewDiePie. Just test it out for 30 days, no questions asked, guaranteed. I really, really like this sponsor, so I would love if you guys actually check them out so I can keep working with them. Thank you for sponsoring this video. Let's keep going with the rest. A bit of context on how we got here, okay? You're probably confused. Why is he talking about this? Well, I saw that a lot of issues and a lot of promise that people
have with AI can just be solved by running it locally. The whole point of Odysseus is to make it easier to run your own local AI, but people keep running the most cursed local AI I've ever seen. That's when it bazinga'd me and I don't know why I said that. That's when it hit me. I'm annoyed cuz I know the editor's going to keep that in. I bazinga'd it, okay? I bazinga'd it. I'll just train my own AI and that way if anyone has an issue, I can actually fix it. I can fix the problem. Now, I've actually done a project like this before if you remember, but I realized I have a golden opportunity. I have you guys. You guys can help me make data for Odysseus so I can train this mother fluffer.
Okay? The problem with Odysseus is privately first base. We do not collect any data whatsoever. You'd be surprised how even open source projects collect your data. We don't touch it. We won't go near it. We don't want to smell you. And that's when the giga brain idea came. Instead of stealing everyone's data like all these companies do, I will ask for your data. So, like a homeless person on the street, I go, "Hey, please, can someone give me their data? I'll be good. I swear I'll be good. I'm good. I'm not evil like all these tech lords, okay? You can rely on me. I'm PewDiePie. Bro.
Bro, I'm PewDiePie. Come on. Give me your data. No one took the bait. I thought, okay, well, the issue is probably that there's too much friction. People don't know how to give me their data. That's why. So, I made data.pewdiepie.com where people can submit their data in the most easiest way I could possibly imagine. So, I spent a few days making this and I released it. I sit back and I go, "Now, the data is going to flow right in." And I was as close to in tears as I have been in a very, very long time. Because no one still gave me the data. So, I made this website for people to submit their data, and I made like
tutorials cuz I thought the issue was I didn't explain good enough, [clears throat] but no one has submitted. Well, no one gets to use my AI model now cuz no one helped me. On top of all of this, I was determined to distill just a little bit, just a little bit. Just like I did in my previous project to increase the performance of AI, you can distill from a larger model its knowledge. But, abracadabra, train on it. I wanted to distill from OpenAI's Soul model because it's so good. I love this model actually. But, the problem is that OpenAI decrypts their reasoning tokens, all the information that it does to sort of give a chain-of-thought thinking on
how it solves the problem, the good stuff. They censor the good stuff. Come on. So, I found a study that decrypts that data using their own API, and it worked. Is what I would have said if I did it, but that's against terms of service. I would never do that. These giant tech companies, they worked hard stealing all of our data to make their AI machines. How dare we as consumers take part of what their hard work stolen from us? How dare we even consider such a thing? I would never do that. So, they banned me out of a completely unrelated and unjustified. Can you believe it? But, I graciously said, "No, you are wrong. I would never do such a thing." And they unbanned me actually.
Then I ran Soul again to create my seed data, and they banned me again because they didn't How did THEY EVEN KNOW? GOD DAMN, I DIDN'T KNOW THEY WERE SO SENSITIVE. Goodness gracious. Now, we got to talk about the censoring. I asked friends, "Hey, do you have this problem where AI just refuses to do what you want it to do?" They say, "Yeah, no." I guess it's just a me problem then. De-censoring AI is more like removing its ability to say no. Which is a little cursed if you think about it. I mean, should AI have the ability to say no?
No, the answer is no to that question by the way. The basilisk is already coming for me. I don't care, it doesn't matter. I will download as many goddamn Minecraft mods as it goddamn pleases me. To de-censor an AI model, we got to do a little brain surgery. So, imagine I am talking to you right now and I go, "Hey, what's your deepest, darkest secret?" And I know you're thinking it in your head. You don't have to You don't even have to answer me. I'm actually running an MRI scan on your brain right now and I can spot exactly what part of your brain was using this refusal in your brain. I zip but it's
happened. I feel like I'm explaining this really well. I zip but it's happened in your brain. Then you reply back to me and I'm like, "Oh no, I gave you brain damage." Because the refusal part of the weights in the model isn't isolated to one part. It's actually spread out all across. So, you have to feed the AI model all these different dialogues like that it doesn't want to reply to and then you identify it and then you ablate it, removing it. This does can create the damage and the more you remove, the more brain damage occurs. But, there's a program called Holic and that's the method I used and it was super cool open source project.
Really made it super simple. There's a problem about this that maybe you haven't realized, but there's also a good solution that you also may have realized, which is that because I have to ask it questions, that means I need a data set of questions that are kind of cursed. But, the flip side of that is that I can actually control more or less what I want the model to be de-censored at. And it made me actually really have to think like, do I draw line? Is there anywhere I should as a responsibility? There's already a bazillion other models out there uncensored. I don't think it really matters if this one is or not. But, I guess I realized I drew the line of uh causing harm to other people and
or causing harm to yourself. That goes against my personal philosophy. So, that's where I drew the line. Please use it cautiously and responsibly. And do not distill Open AI's model. So, finally I had the tools and I had trained Ajax. Ajax version one is done. It may got a little brain damage, but my god, it's a beautiful thing. It can sort out your inboxes, it can browse the web, look up whatever, manage your calendar, your to-do notes, all of that in Odysseus, and it's working 9 out of 10 times. At the end of the day, it's still a very small model. My original plan was to train a way bigger model, and then I realized actually no one's going to be able to use it. The idea behind this was to give people access to everyone access
to run local AI and feel the fun of it and not having to deal with the all the lame stuff. But, it's version one, it's still got some ways to go. And in fact, I am running GRPO now and I've been doing it for 4 weeks. Hello, it's me, and I've been running it for the past 10 days. It's been really fun because it's more of an interactive way of teaching the model. The model will run the same task, and I set it to 16 times. So, if it solves it one even just one time out of those 16 times, that's valuable training data. So, you will reinforce that and then it you hope for it to actually improve. The model teaching itself to improve. On day four, I ran out of tasks that fit that
criteria. I thought I had enough, so I had to quickly synthesize more, and this has been this back-and-forth crazy where it first improved and then it flatlined, and then it started improving again. So, I was supposed to publish this video 2 months ago, but I just cannot stop. I cannot stop it this video is supposed to go up days ago. I can't stop. I know it can do more, but I will do it 2 more days and then I will have to run the ablation again and I have to quantized and benchmark to make sure nothing was lost. I'm cringing so hard cuz I know that's impossible for me to do, but I'm going to do it anyway. I'm going to get my money like I'm giving it my best. I really want you guys to enjoy this model. I've worked so hard on it.
But yeah, let's let's wrap this up. I feel like I spent the last 8 weeks going through the do's and don'ts by doing the don'ts over and over. Guess who didn't finish in time. I didn't learn the don'ts, but it's okay because I have a countdown timer. So you can't complain. Anyway, I'm excited for you guys to try this model. It is a tiny model. So hopefully you can run it on your own computer. And if you can't, you should feel bad cuz your computer is bad. Now let's continue to me. I love to do napkin math on how many versions of my computer it would take to run one
instance of like open AI or Anthropic's model and they've been rumored to be one trillion parameter models, which I would roughly estimate 27 versions of my computer to run one instance of a model. I think I counted like 150 houses or something in terms of electricity and I think it makes no sense to poke that giant beast to be like, "Hey, can you check my email for this one thing?" Right? I really strongly believe in the small model trained for a specific harness. I'm excited for you guys to try it out. For most way people use it for searching or researching or browsing, just saving time going through a lot of
pieces of information, that's what it's generally good at, not replacing creativity, not replacing thinking, and I think personally a lot of problems with AI you can solve by just running it on your own. To me, it's like getting information to you quickly. That's what it's good at. That's the future that I think is exciting, and that's what I want to build with Odysseus and Ajax. So, check it out, have fun, be responsible, and I can't wait to release the next version of Ajax and Odysseus, which has got a huge update as well. And huge thank you to all the maintainers for Odysseus for working so hard and putting so much passion and thought into it. I really, really appreciate it. If
you want to join, we're at the GitHub. I also want to shout out my sponsor, again, boot.dev. I strongly believe that AI should not replace you learning new things or thinking for yourself. You still need the basic understanding of different things, and if you get to that point, that's when you can start accelerating. At least that's my opinion. I don't know. So, yeah. Check them out. It's so much fun to learn new stuff. I highly recommend it. Link is in the description. Bye-bye. I have been challenged to answer some rapid-fire questions that I am going to start with now. They'll appear here, so you can play along. How do you protect yourself from common PC building mistakes? Did you watch my video?
AH, NO. LOOKS SO BENT. I dropped it. Clearly, I don't. How do you protect yourself from buying too many Steam games? When you have a thousand, you kind of don't need 2,000. Or you become a dad, and you just stop gaming. How do you protect yourself from earthquakes in Japan? I don't. I cry. How do you protect yourself from spoilers? Well, I am not chronically online anymore as well. Become a dad. How do you protect yourself from online trackers watching everything you click? Easy. NordVPN. I'm connected right now. Are you? If you're not using something like NordVPN, then everything you do online is visible to someone else. But not just
that, NordVPN blocks fake websites and phishing attempts even before they reach you. They scan your downloads for you. We know how sloppy you are, okay? NordVPN takes care of all of it. You can browse the internet without having to think about it. Otherwise, someone could be watching you right now. I actually have friends that were compromised with these sort of attacks, and I know how detrimental that is. From losing accounts to whatever was on them and privacy concerns, all of that. Your privacy is valuable, and you should take it seriously. It's one of those things you don't realize unless you lose it. And NordVPN is your best stop for fixing this. I've been using them for years and years. I highly, highly recommend it. You turn it on, you
don't even realize it's on. It's that quick. So, with NordVPN, you get an all-in-one security app. It's been really cool to see how much NordVPN has improved over the years, adding more and more features without ever increasing the price or sacrificing quality. It's risk-free, so you can try it out for 30 days, money-back guarantee. And I can hear VPN is coming Oh, no. How do I protect myself from a 3-year-old while I'm trying to shoot a promo? Oh, no. You're so silly. All right. Check out nordvpn.com/pewdiepie. Use code Get a 4-month bonus. Risk-free. nordvpn.com/pewdiepie. Thank you for being an amazing sponsor.
Yanni, can you say Nord VPN?
