OpenClaw + Auto Research = God Mode

0:00Building with AI has just changed forever. And in this video, I'm going to be talking and deep diving the GitHub repo Auto Research done by Andrej Karpathy on how it's going to affect how builders build with AI for good. So, quick rundown on who he is. He was the head of AI at Tesla. He built neural networks behind autopilot. He was a founding member of OpenAI. He created nanoGPT, which is basically taught the entire internet how language models actually work under the hood. This is one of the most respected AI researchers on the planet, and we're going to be deep diving his latest project Auto Research, and it will change the entire landscape. A couple weeks ago, he open-sourced that repo called Auto Research. I have it linked down below.

0:45You definitely should check it out. And it went viral like crazy. 26,000 stars on GitHub in less than a week. VentureBeat wrote about it. Reddit lost its mind. Shopify CEO tried it the same night. And the reason why everyone's freaking out is because of what it does. Here's the idea in a nutshell. You give an AI agent a small language model training setup, and then you go to bed, and then while you're sleeping, that agent runs experiment after experiment, testing and testing on the code. It tweaks the learning rate. It changes the architecture. And then it adjusts the normalization. It basically it runs a training loop. It checks the model if it got better. And if it did, it keeps the change and keeps

1:30iterating over and over. All this happens while you're sleeping, keep in mind. If it didn't work, it throws out, and then it tries something else. And it just keeps going. It keeps iterating. Two steps forward, one step back, back and forth until it gets better. All night long, no human involved. Karpathy left it running overnight, and it completed 126 experiments in one night. He left it running for 2 days, and it ran about 700 experiments in total. A human researcher, a really good one, might get through eight to 10 experiments in a full day. This thing ran 748 hours. That's not the same game anymore. That's not even the same sport. We're in an entirely different league when it

2:15comes to AI agents working on a constant feedback loop. Here's the part that really got me, and it should get you, too, is Karpathy has already been hard tuning this code for months. So, this guy with 20 years of deep learning experience, the agent found 20 genuine improvements in the code he thought was already optimized. One of those improvements was a bug in his attention implementation, a missing scalar multiplier that is making the attention too spread out across all the heads. He had missed it for months. And the AI caught it just because it doesn't get tired. It doesn't get distracted. It doesn't decide to go get coffee after the 15th failed experiment. So, stacking

3:01all of those improvements together dropped his time to GPT-2 metric from 2.2 hours to 1.8 hours. That's an 11% efficiency gain on code that one of the best researchers in the world already thought was done. And that is crazy. Real quick, before we keep going, if you're watching this and you want to actually build with some of these tools, not just watch videos about them, then you're going to want to check out our community down below, Shipping School. We have a full Claude code course, a full OpenAI code course, and four live bootcamps every single week where we actually help you get set up from scratch. Like actually set this thing up, not just watch a tutorial and figure it out by yourself. And we also provide one-on-one coaching, so you could book a

3:46call with me. We could share screens, and I can help you get Claude code or OpenAI running on your machine. That's it. No fluff. I built this community because watching YouTube's only gets you so far. We launched it just 3 days ago, and we have over 55 members. You need people around you who are actually building, people who hold you accountable, and coaches who can help you when you get stuck. I'll put the link in the description down below. Get in now before the price goes up. Now, let me tell you why the design of this thing is so smart, and you should check it out and see if it's going to work for your setup if you're looking to train AI models. Because it's not too complicated. It's actually pretty simple. The entire repo is basically three files. First, you have a prepared.py. That's your data prep script on Python.

4:33You don't need to touch it. It sets up the training data. The second thing is you have train.py. This is 630 lines of code for the training, and is the file that the AI agent modifies. Every experiment, the agent reads this file. It comes up with a hypothesis, and it makes a change, and then it runs the training. And that's it. One file, one target. The third thing is you have a program markdown file. And this is the brilliant part. This is a plain English markdown file that tells the agent what to do. Think of it like instructions you'd give a research assistant. You're telling it what to explore, what constraints to respect, what the goal is. You're not writing code anymore. And you shouldn't be in the era of AI.

5:19You're more focused on writing directions nowadays. Karpathy basically said that the human's job is no longer writing code. Don't think you need to getting into this industry. You don't. The human's job is now more focused on writing research directions. You're the research advisor in this age of AI. The agent is the PhD student who never sleeps and never complains. And there are some really intentional design choices here that make the whole thing work. The fixed 5-minute time budget is pretty big. Every experiment gets exactly 5 minutes of GPU time. That means all the results are comparable. You actually can tell if one change was better than the other because the conditions were the same. It's like a science fair for AI. You know, fair

6:05conditions, clear metric, let the best experiment win. The metric is val_bpb. The validation bits per byte. Lower on that is better. That's the only thing that matters. One number. Did it go down? Keep it. Did it go up? Throw it away. It's really that simple. The agent doesn't need to understand why something works. It just needs to know if the number went in the right direction. And the single file modification thing, only train.py gets changed. That constraint is actually what makes it so powerful. Because when you limit what it can change, the agent has to be creative within those boundaries. You're not giving it the keys to the whole house. You're giving it one room and saying, "Make this room as good as possible."

6:50Now, here's where it gets more crazier if it isn't already blowing your mind. It's not just Karpathy using this. The whole community jumped into it. Tobi, the CEO of Shopify, took this pattern and applied it to an internal query expansion model overnight. He woke up the next morning to a 0.8 billion parameter model that scored 19% higher than his previous hand-tuned 1.6 billion parameter models. That's insane. A smaller model beat one that was literally twice its size because the agent optimized the architecture for his specific hardware instead of just defaulting to bigger is better. That's pretty cool, especially if you're running local models on smaller pieces

7:35of hardware that can't handle the robust, you know, Quinn 3.5 and such. And then there's this company called Hyperspace AI. They took a single agent loop and distributed it across a peer-to-peer network. On the night of March 8th, which was just what, a week ago, 2 weeks ago, 35 autonomous agents on their network ran 333 experiments on completely unsupervised network. And what happened next is the insane part. The agents on a huge H100 GPU used brute force to find aggressive learning rates. But the agents running on regular laptops with just CPUs, they had to be clever. They couldn't just rely on raw compute, so they focused on initialization strategies and

8:22normalization choices. The underdogs got creative because they had constraints. And they were sharing discoveries in real time using something called a gossip protocol. Never even knew that existed. When one agent found that a specific initialization technique dropped loss by 21%. That discovery spread through the network like wildfire. So, within hours, 23 other agents had incorporated it into their own experiments. And in 17 hours, these agents independently rediscovered ML milestones that took human researchers at Google Brain and OpenAI nearly 8 years to formalize. 17 hours versus 8 years. Hmm. This is getting scary and

9:08crazy at the same time. People on Reddit are going back and forth about this. Good place to check to get some opinions, right? Some folks are saying this is basically, you know, an early singularity. Karpathy himself joked about it. He said, "You know, who knew the early singularity could be this fun?" That's crazy. Other people are more cautious saying, you know, we need to see how the AI cost stack up when you're running hundreds of experiments. That is a fair point. But the pattern is undeniable here. You know, somebody on the Claude code subreddit already built a skill that applies the same pattern to any task, you know, not just machine learning training. It's across all of Claude code skills. You could apply this. And any task where you have a clear metric and you want to iterate, which brings me to the part I really want to talk about. Because even if you

9:54never trained a language model before, and most of us have it, you know, even if you don't have an Nvidia GPU, and no one has five grand to drop, you know, the money on a, you know, 5090, you know, even if the word PyTorch makes your eyes glaze over, this pattern should matter to you. Think about what's actually happening here. You know, you have an agent, you give it a clear goal, you give it a metric to optimize for, you give it some constraints, and then you let it run experiments. You're basically keeping what works, and then you're throwing away what doesn't. And the compounded improvements over time is insane. You know, that's not just my machine learning research, that's everything. You know, your marketing, your landing page, your email subject lines, like your content strategy even.

10:39Let's take a look at your pricing page. You know, every single one of those can be framed as an experiment with a clear metric in mind. Open rate went up, keep it. Conversation dropped, throw it out. Click-through improved, stack it on top of the last winner. The feedback loop is the product. That's the insight. Karpathy just showed it in the context of training a language model, but the pattern applies everywhere throughout your whole project. This is basically what my content pipeline already does. You know, I've got Ghostface who goes out every morning and scouts the trending news. The Rizzler writes the scripts and the outlines. Inspect the Deck handles all things related to X. You got generating thumbnails, and every morning that E-Val checks what performed and what didn't. We connect to the back-end YouTube API, and then, you know, it just adjusts as we go. And as

11:26time passes, we have this evaluation loop happen. You know, what's not working? What is working? It's the same loop. You know, we scout, we create, we measure, and then we keep the winners, and we ditch ditch the losers. It's just on repeat. Karpathy just did it for neural network research with a 630-line Python script. The difference is he formalized it so cleanly that now anyone can apply the same thinking to their domain. You know, you don't need to have a PhD. You don't need to have an H100. If you have one, great, but need one. You just need a clear metric and a willingness to experiment and tinker around a little bit. And you need to have the discipline to actually throw away what doesn't work instead of getting emotionally attached to it. And I know the last part is hard. I've been there when you spend so much time working on something, you really want to get it to work, but I've gone through

12:12the same thing building out Easy Flip. I spent 30 days on it straight trying to get it to work, and I ended up just scratching the project and starting over. And within 1 day, I got the project to where it was, and it's working even better after just scrapping it all. Sometimes you just got to hit the reset button. Don't feel bad. You know, um for another example is like you spend 3 hours on a video idea, and you don't want to admit it it's not going to perform, but the agent doesn't have that problem. It doesn't care. It ran the experiment, the number went in the wrong direction, it threw it out, and then it just tried something else. No ego, no sunk costs, just the metric. If you're struggling to keep up with content, well, I'm about to save you about 40 days worth of work. [music] I built something called Content Machine. It's 10 AI agents that run on the Open Claw

12:58orchestration, and they handle everything. Scripts, [music] thumbnails, X posts, blogs, outreach, clips, newsletters, all of it. So, I went from 1,000 subscribers to [music] 4,000 subscribers on YouTube in 7 days using this exact system. Every single morning, I wake up, [music] and the content's already done. I spend maybe 15, 20 minutes reviewing and approving them, and I move on with my day. [music] It works for any niche, fitness, finance, real estate, marketing, whatever you are building, and it is 100% completely customizable to your [music] use case. So, you get the mission control dashboard, all of the cron jobs, everything I've built over the last 40 days helping me gain more and more people to subscribe and join the

13:44community. So, you plug in your own thing, and it molds it to you, it learns how you talk, and it writes [music] so it doesn't sound like AI slop. $97 one-time, it's not a subscription. I'll put the link down below, and you'll thank me later. Karpathy wrote this sci-fi intro into the repo that honestly gave me a, you know, it rose my eyebrows, gave me a little chills. He said, "One day, frontier AI research used to be done by meat computers in between eating, sleeping, having other fun, and synchronizing once in a while using sound waves in a correct in the ritual of a group meeting. That era was long gone. Research is now entirely domain of autonomous swarms of AI agents running across compute clusters, mega structures in the sky." So, he's

14:30basically describing a future where human researchers don't do the experiments anymore. They set the direction, they write the program markdown file, they decide what questions are worth asking, and then the agents do what actually works of answering those questions at speed and at a scale that no human can match. Could just a pause right here, could you imagine how this could be applied to the medical industry? It's going to shift everything. It's going to change everything as far as testing medications and testing trials, what's working, what's not. It's going to be insane. I know some people hear all of this, and they get scared, especially beginners. And, you know, I hear it, and I get excited, because this is the same opportunity I've been talking about since day one. The people who learn to work with these tools, who learn to set

15:16up these loops, who learn to think in terms of metrics and experiments and autonomous iteration, those are the people who are going to win. The repo is MIT license, fully open source, and you could go look at it right now. It's three files. That's it. So, if you have an Nvidia GPU, you could run it tonight. And if you don't, the community has already forked it for Mac, Windows, and AMD. It requires Claude Code or Claude Code X, or really just any coding agent that could read a markdown file and modify a Python script. Honestly, the specific repo isn't even the point here. The point is the pattern. Give an agent a clear objective, a measurable metric, constraints to work within, and then let it run. That's the entire formula in a

16:02nutshell. Whether you're training a language model or testing landing page headlines, or even optimizing your YouTube thumbnails, the loop is the same. So, if you want to learn how to set up these kind of automations and loops for your own business, that's literally what we do inside the community. We have over 118 members just in the last 12 days, and people are building, people are shipping, people are asking great questions and getting good feedback from other people in the community. You know, we're in there, we're building with AI agents, we're doing six live calls every week, we're real people shipping real products. And I'll put the link in the description down below if you want to be part of a like-minded community of non-technical builders building and shipping and learning new things every day. We do have courses in there, too, for Open

16:47Claw and Claude Code. We're really focused on community and actually human-to-human interaction. So, you could check that out. If you guys haven't subscribed already, we do about four to six videos every day. And if you like this AI news and AI introspective of what I'm doing in my journey and how I'm building things, then subscribe. Hit the notification bell, and we'll see you in the next one. Have a blessed day.