Karpathy_s _autoresearch_ broke the internet

In this episode, we explore the excitement surrounding Karpathy's latest AI project and why it has captured the attention of developers, researchers, and the broader AI community. We break down what autoresearch aims to do, the ideas behind it, its potential impact on AI-powered research and automation, and why many believe it could reshape the way we discover and develop new knowledge.
Whether you're an AI enthusiast, developer, or simply curious about the latest breakthroughs, this episode offers a clear overview of one of the most talked-about innovations in AI.
🎧 Tune in to discover why Karpathy's autoresearch is making waves across the internet.

Transcript

SPEAKER_00 0:00

Andre Carpathy, I mean one of the godfathers of AI has just launched something called auto research. And auto research is a huge deal and it's going viral on Twitter. And I just wanted to do an episode where I can explain to you in the clearest way possible what it is, what are the use cases, how to make money from it, how to be more productive with it, how to create impact with it. And by the end of this episode, I'm gonna give you a bunch of different ideas, use cases for how to use auto research. I'm gonna explain it to you in the most clear way possible. And at the end, I'm gonna tell you how you can actually get started with it. Um, so let's go right into it. So, what is auto research? Well, it's like having a super nerd robot intern that runs science experiments on AI AI models for you all night without you doing the boring stuff. I mean, sounds intriguing, right? So, how do you actually program it or get started with it? Well, the first thing is you gotta give it a goal. So you can say something like, make this small AI model smarter. That's the goal. And then an AI agent will actually plan what to do, like different settings, code changes, edits the Python code for you, runs a short training experiment on a GPU for about five minutes, it reads the results, it, and then it decides what to change next and to repeat the loop. So, in some ways, you know, if you've seen my video on the Ralph loop, where it basically would do engineering 24-7 and you'd wake up to new stuff happening, in simplest terms, that's what auto research is helping you, you know, it do. You give it a goal, the AI agent does a thing, you know, you tell the AI what better means, uh, cheaper leads, more clicks, higher sales, better model school. And then the AI keeps changing things, testing them, and it only saves the changes that improve. So what's really cool about it is you wake up, you grab the best version, and then hopefully you turn it into something you charge for, or you know, you give it away. I saw this tweet by Toby, who's the uh CEO and co-founder of Shopify. Auto research works even better for optimizing any piece of software. Make an autofolder at a program MD. That's just a markdown file, which is really the foundation of what'll, you know, how you're gonna be using auto research, and a bench script, make a branch and let it rip. So that's why I started paying attention to auto research, right? When Andre Carpathy, Legend, and Toby and more people you know start playing with it, I'm like, okay, I got to pay attention. So I created this little visual for for how to think about what auto research is. So you set the goal, uh, the the AI plant is an experiment. It edits and trains the code and settings, it runs a short training on a GPU. By the way, this is an important, I should I should mention that you need a uh an Nvidia chip to actually run auto research, or you can do it in the cloud. I'll talk about this at the end of the episode, but you you know, you do need that. You can't just run it on, let's say you have a MacBook M1 or something like that. It reads metrics. It says, is it a better result? If it's if it's not, it's gonna log the attempt and it's gonna discard the config. If it's yes, it saves it to the config. Um and then just plans a different experiment and it just you know hopefully gets better on your goal, whatever it is. So um let's uh let's get into um we're gonna get into some of the ideas, business ideas around it. But right before that, I just want to say here's a simple mental model for how I'm thinking about uh auto uh auto research. So imagine you have a research boss you can boss around. Number one, you write a clear task. So for code experiments, maybe it's improve this model test score for business, figure out the top five competitors for product XYZ and make a short report. Step two is you give the uh you give the bot um you know access to the code, a GPU for ML experiments. You obviously need to give it access to the internet and documents if you're doing reading task. The bot then runs a loop. So it plans, it acts, meaning it might run code or search, it reads results, it updates the plan. And then you just come back later, you know, uh it could be 12 hours, 20 hours, six hours, and you see if it's logged everything, charts and metrics, and then it gives you a written sum summary in normal language. So, you know, think of auto research as a research bot that runs experiments for you while you sleep, tries lots of ideas fast and keeps the winners. Quick break to invite you to something. Now, this isn't an ad. I just want to invite you to a free event because I think that you're gonna get a lot out of it. I wanted to take one hour of time where we just talk about building businesses in the age of AI. People say SaaS is dying. I actually believe the quite opposite. I think that SaaS is just evolving. I think right now is an incredible time to be building software startups that help you craft your dream life. And for all those reasons, I'm said, I said, let's just book one hour of time. It's gonna be 11 a.m. March 12th. That's a Thursday, where we can go and lock in and just talk about building businesses in the AJAI. I'll include a link in the description in the show notes to join. And I can't wait to see you there. Okay, how do we use it? Here's some ideas for you. So the first idea for you I have is a niche agent in a box, you know, products. This can be multiple products. And by the way, I put out these ideas. I want you to do these ideas. I think that, you know, even if they don't turn into businesses, you will learn about these tools, and that is going to help you outperform 99.9% of people on this planet. So uh you package tiny auto research loops tuned for one painful niche. So the example I think of is an Amazon listing experimenter, an email sequence tuner for real realtors, uh, a pricing optimizer for SaaS. Those are you know auto research loops and ideally in a niche that you understand well. And then you charge a monthly fee. So the value prop is this thing runs experiments for you 24-7 and just show shows you the winner to click accept. How valuable is that? And how many different niches are there that you know this plays into? The hard part is figuring out what though what's the pain points, and then and then obviously you know you want to be quick, quick to market, right? So here's a visual of it, pick the painful niche, design the tiny auto research loop, run experiments automatically, see which setup works best, turn best setup to a simple agent product, and then you charge that monthly subscription. Number two, you're gonna wanna, you know, here's an idea print money using an A-B testing for marketing. So this is it's it's very similar. Um, but instead of, you know, uh instead of uh you know doing it for realtors or whatever, you're doing it for ads and landing page experiments. So landing pages, so the agent writes variants of headlines, layouts, and offers pushing them to traffic measures, which one converts better and keeps iterating. So this is like conversion rate optimization around landing pages. You know, the old think of you know tools like Optimizely. That's a SaaS tool that, you know, when I first moved to San Francisco, I remember how big they were, and everyone's talking about Optimizely and A-B testing. And it's like, well, this is the future of that uh auto research for different landing pages. You can also do use auto research for something like ads, which auto-test creatives, it auto tests angles and audiences, and then it keeps the combo uh combos that lower CAC or Raise ROA. So, you know, you profit by running this for your own product products. Like if you if you want to build your own products and just use this internally, that works. Or, you know, all offering an always-on experiment engine to clients as a retainer service for $5k a month. I'm gonna give you the best landing pages every single month, and it's just gonna come to your inbox, that sort of thing. Visual of it, business goals, uh you know, the goal that you're giving the auto research is more sales, it's generating things like pages and adversions, sending traffic to the versions, measuring conversion and revenue. Um, does any version beat the current best? Um, you know, if it doesn't, then you're gonna keep the current control. But if it does, you know, you're promoting the winner to a new control and it you're asking for the AI for new ideas. All right. Hope your creative juices are starting to get flowing, you're starting to understand a little bit more about how it's working, how you think about goals, how you can think about agents, and how you can set up these loops. Number three, research as a service. So auto research's recipe is basically a loop for doing research, right? Because you're searching, reading, summarizing, and you're comparing, and then you're repeating. So, how do you point that at money problems like market and competitor research for startups? So constantly updated reports on who's doing what, pricing, features, and gaps, super valuable. Investor and MA decks, fast technical and market due diligence summaries, super valuable. Compli um compliance and regulation tracking for niches. I don't know, crypto, healthcare, finance, super valuable. So you can charge per report like a one-off, or you can set up like a monthly subscription for always fresh dashboards. So visual, uh, define client research question, auto research, searches and reads, um, summarize and compare findings, creates reports and dashboards, deliver insight to client, and the client pays per report or monthly, whatever you decide. Number four, uh power tool inside your own product. So if you already have built a SaaS or workflow, embed an auto research style agent so your users can press optimize, just like a big, I envision like a big button that just says optimize. And the system runs a mini research loop for them. So for example, tune prompts, pick best pricing, rank suppliers. Then you can charge higher tiers for this feature, or you can use it as a wedge to upsell pro and enterprise brands. So maybe you uh maybe that's a part of pro and enterprise. Maybe it's something that you just send an email to you know your entire list and you're like, hey, you know, we have this really powerful tool. Imagine you press this button. It's like it's like bending spoons, right? It's like bending spoons. Like, how is this um bending spoons, not the private equity group I'm talking about? I'm like the idea of you can bend a spoon, right? It's incredible that you'd be able to optimize, press a button, and this would happen. So visual over here, have an existing SaaS, add an optimize button. Users run many research loops, tools suggest better settings or prices, users see better results, offer higher price pro plans and enterprise plans. Number five, this is a saucy episode, by the way. This is saucy. All right. Agency that sells, we run more tests than anyone else. Because auto research lets you run hundreds of experiments instead of a few, you have a simple pitch. We do a hundred times more testing than other shops for the same or lower fee. A niche example, a Shopify store conversion lab, B2B SaaS pricing experiment service, email subject line and sequence optimizer. You charge per month and a bonus if you hit specific KPI lifts, Rev share performance fee. People love that. You know, of course they're gonna be you know interested in yeah, if you can do if you can lift this KPI, we'll give you some bonus. So here's the uh the visual start an optimization agency, use auto research to run many tests, improve stores, pricing, emails, and funnels, show clients more experiments and wins, charge monthly retainer and performance fee. Number six, and we've got about uh 10. Yeah. So we're almost almost done. And then after we're gonna talk about um just some cool, interesting you know, stories around auto research, and then I'll end with uh you know how you can set this up very briefly. So auto quant for trading ideas. So you can use auto research to run small, fast uh back tests of many simple trading rules. So LLM base, factor screens, sentiment filters on one GPU overnight. So you can keep the few strategies that look promising, then either trade on your own account or sell signals and strategy reports. So depends if you're a trader, maybe you're doing yourself. Um, or you yeah, you can just you know sell this as a digital product or yeah, yeah, yeah, basically a digital product. So you def define the simple trading rules, you run many back tests overnight, you review the strategy performance, you keep only promising strategies, trade your own capital, or you can sell the signals. I think finance is changing a lot. Um, and I think with things like auto research, uh, you know, it just it's going to be an unfair advantage for a lot of people. Um so I think you're gonna see a lot more digital products uh that people sell and also you know, just using their own money, trading themselves instead of giving uh 1% or whatever to a financial advisor. Um I'm sure also, by the way, a lot of people are gonna get burned by this too. Like they're not they're just gonna blindly just trust an auto research. You need to have a human in the loop and you need to manage that uh obviously accordingly. But yeah, you can just see, yeah, there's definitely gonna be some people are gonna get burnt. You just give the entire um they're just gonna like give a bank account and just let auto research just trade for it. I mean, would be interesting, it would be an interesting test, that's for sure. Number seven, always on lead qualification and follow-up. A point an auto research style agent at your CRM, so like a Salesforce or something like that, and inbound leads. Let it test rules and messages to see which leads are most likely to buy, right? It auto grades the leads, suggests next actions, and drafts follow-up. So salespeople only focus on high value deals, so it's more revenue per hour spent. Visual over here for you. Connect to CRM, you know, auto research tests the leads, rank leads by likelihood to buy, draft follow-up messages, sales focus on best leads, revenue per sale increases. Eight, finance ops, autopilot for businesses. Use the loop to grind through invoice matching, expense report generation, and exception detection with continuous small improvements to rule and prompts. You can sell this as we cut your AP expense time in half, either as software or as an op service with a small team and agent. By the way, I can totally see someone like someone starting this, and this gets acquired by one of the large fintech companies or one of the large banks. Uh, so visual uh here, ingest and voices and expenses. The auto research improves rules and prompts, matches invoice and detects exceptions, it generates clean expense reports, reduces manual finance work, and then you can sell it as a software or op service. Or you start, maybe you start as op service and then uh you kind of evolve into the software. Two more for you. Number nine, an internal productivity lab for your own org. I thought this was interesting. So treat your company like Carpathy's GPU lab. Define KPIs, so like response time, close rate, ticket resolution, and let agents iterate on workflows and templates and routing rules. So you just get fewer meetings, less manual grunt work, and then you personally touch only the high impact decisions when everyone else rides the improved process. So the goal here is defining the key metrics, auto research is testing the new workflows, it's improving templates and routing rules. You're cutting meetings and manual tasks. That's good. Team focuses on high impact work, and then higher productivity and ideally higher profit. Last idea for you done for you research or due diligence shop. So you use the research loop to chew through docs, filings, product pages, and reviews and keep an evolving living memo for clients like investors, acquirers, execs. You make money by selling fast, well-structured briefs, and a monthly uh update packs instead of one-off manual research logs. Um, so uh, you know, the the goal get investor or acquire a question. This happens all the time. Auto research reads through docs and filings, it summarizes that product markets and risks and maintains a living memo for the client. It delivers a brief and updates packs, and the client pays for reports and ongoing access. Um I would pay for something like this. Um, so hopefully someone builds it. All right. So those are a bunch of ideas for you. I also saw a couple interesting things this morning. Uh, my good friend Morgan Linton, uh, who's you know been on the pod before, he says, I woke up this morning and all I can think about is auto research. So many ideas ideas swirling around in my head. Not sure 99% of the world realized the incredible breakthroughs carpathy is making and just sharing casually on X. Right now, where my mind is going is medicine. It feels like in many ways, clinical trial design is itself kind of like a hyperparameter search. I know right now trials cost tens of millions of dollars minimum. It feels like an agent swarm could optimize treatment protocols on small proxy experiments, promote the most promising candidates, and then move to humans to review. So humans still very much in the loop, but later on and experimentation going much deeper, happening faster, and for far less money. I think for me, while I'm not a doctor, he's an engineer. What I'm the most excited about when it comes to AI is the impact it will have on human health and critical areas like disease treatment. Might be a crazy idea, so a real doctor can jump in the comments and slap me on the wrist here. I looked at the replies, I didn't see uh, you know, any doctors come in. But I don't know. I just can't stop thinking about how what Carpathy has discovered here could have some pretty profound implications. So only halfway through my coffee though, but woke up this morning and this is what I'm thinking about. So thought I'd share. I agree. I think there's a lot of really interesting, not just like business profit ideas, but also just like medicine, science, uh, research. So I'm excited for people to take this and and to continue with it. Um, I also saw this tweet here. Uh, what's after auto research? It's Carpathy's new open source project, Agent Hub. So Carpathy also launched Agent Hub. What is Agent Hub? It's GitHub for humans. Uh sorry, GitHub is for humans, agent hub is for agents. So it's basically a GitHub for agents, an agent swarm collaboration platform, a very promising direction. I'm watching him speed run a one man billion dollar uh company. If you look at the GitHub for Agent Hub, it says first use cases for auto research, but it's a lot more general than that. Exploratory project. He says agent first collaboration platform, a bare Git repo, a message board designed for a swarm of agents working on the same code back code base. Think of it like a stripped-down GitHub where there's no main branches, no main branch, no PRs, no merges, a sprawling dag of commits in every direction with a message board for agents coordinate. I think this is really interesting, and just like whenever Carpathy's up to something, I'm always paying attention. So I had to put that one in there as well. So you know, maybe you've gotten to the end of uh this episode and you're kind of like, okay, I kind of I think I understand what auto research is, I think I know what you know, Carpathy's a G, Toby's a G. Like all these smart people are playing with it. Um, how do I get started? Well, to get started, I'd recommend um just tell Claude Code to get you started. So, you know, I went ahead and I basically was like, uh I I gave um Claude Code the lit the this um this GitHub repo, the GitHub, the auto research GitHub repo. And wow, 25,000 stars already. So this is crazy. Um it's really growing, growing quick. Um so I just gave I gave uh gave it the link and I was just like, I need help installing auto research by Carpathy. Um and it says here's how to install it and set up auto research by Carpathy. You need an Nvidia GPU. So I talked to I talked about that in the beginning. It was tested on a H100, but other NVIDIA GPUs should work, and you need a UV package manager. So you have to install UV, you clone the repo, you install the dependencies, um you prepare the data and run a training experiment. In my case, I don't have an NVIDIA GPU. I'm actually using a MacBook and an M1 Pro. I know I'm I need a I need a upgrade um to a new Mac. So I was like, so wait, I need an NVIDIA GPU to do this. Um, but there's a few options cloud GPU. Um you know, you can so you can rent an Nvidia GPU from a service like Lambda Labs, VastAI, RunPod, or Google Collab. Some offer free tier width GPUs. This is the most straightforward path. So that's that's the answer to people who don't have an NVIDIA chip. Just rent it on one of these services. I personally use Google Collab. Why? Um I just know Google the best and trust Google the best. Um you know it also says you can try it via Apple Silicon via an MPS backend. I'm like, no, I'm not gonna do that. Um so with that, that's what route I did. I went on Google Collab, the easiest way to get started. You go to collab.google.com, you create a new notebook, you change the runtime to change runtime T4 GPU, and you run a bunch of commands. That might be like complicated, sound complicated. You this is what collab looks like. You literally just tell, you know, you you listen to what um Cloud Code tells you to do, and you just paste it in and you can get started. So um, you know, if if people are interested, I can spend you know more time with this, with auto research as I'm learning, sharing more about it. But I just wanted to do give you a quick primer on what it is, why it's important, what are some ideas on how you can actually use this thing, um, and then how are people installing it? They're just you know, you can use cloud code as your helper to get it installed installed, and you're gonna want to rent uh a GPU in the cloud, at least to start. So hope this has been helpful. Um this is an another solo podcast that I'm doing on the Startup Ideas podcast. The last time I did this last week, I had a lot of comments that said, Yeah, Greg, I actually really like when you just come in solo and just start like telling us what's on your mind and stuff like that in real time. So I'm here, I read every single comment. So, you know, keep commenting, keep liking, keep subscribing, and I'll keep you know putting this out there for you for free. Yeah, I'm excited to see what you end up using this for. Um, of course it's early, right? Like this is this is brand new. Um people are still trying to figure out what are the use cases, but I always find that you know, in the in the fog, in the fog, people don't really understand where the opportunity is, is when there's l sometimes an opportunity. So um one thing I've just learned in my career is just like when I see people like Carpathy doing things like this, you want to pay attention, you want to tinker with it, you want to have some.