In this video we talk about Andre Karpathy’s Auto Research, a GitHub project where an AI agent runs LLM training experiments overnight on a single GPU by modifying code, doing 5-minute training runs, and checking if results improve. We dig into what’s actually happening (is it “training” or “research”?), the key constraint of a fixed 5-minute time budget per experiment with a single file and single metric, and why it’s built on NanoChat as the simplest harness for training LLMs. I also walk through the practical flow mentioned in the repo: download it, install dependencies, run the prepare script, then kick off a new experiment from the program.md instructions and see if your system is set up correctly.
—
TIMESTAMPS:
00:00 Karpathy’s Auto Research
00:21 What It Actually Does
00:57 5-Minute Experiment Loop
01:47 NanoChat Connection
02:18 Setup Steps
02:58 Why It Runs Overnight
03:14 What You Get Out Of It
03:22 Command AI Plug
—
Unlock the full potential of your online presence with Kabarza and Samuel—experts in web design and development (respectively), powered by cutting-edge AI solutions. We blend creative design with advanced tech to deliver smart, high-impact websites that stand out. Ready to elevate your business? Contact us today and see what AI-driven innovation can do for you!
LINKS & RESOURCES:
Website: https://cmdaishow.com
Check out Kabarza's amazing work: https://kabarza.com
Visit Samuel's website for more: https://samuelgregory.co.uk
📷 Follow on Instagram: https://www.instagram.com/cmdaishow
—
HASHTAGS:
#ai #podcast #aidesign #aidevelopment #vibecoding #webdesign #webdevelopment #ainews #webnews #designnews #devnews #AI #LLM #AndreKarpathy #MachineLearning #AIAgents
Transcript
From one loosely researched piece of news to another, Kapathy has developed auto research. Command AI. I never know if this is Karpathy or Kapathy. I'm sure I researched it and it said Kapathy, but it's spelled Karpathy. So, sorry, Andre. Let us know. Give us a text. Let us know how to say your name. Yeah. Andre Kapathy released auto research, a project where AI agents auton autonomously experiment on a single GPU LM training setup overnight, modifying code, running 5minute training experiments and iterating towards better models. So this is what confused me. Is it training or is it research for better models? So I released this GitHub repo which you can download it yourself. The idea, give an AI agent a small but real LLM training setup and let it experiment autonomously overnight. It modifies the code, trains for five minutes, and checks if the result is improved. So, I is slightly outside of my wheelhouse really on the understanding of all of this. But if I re um revert or resort to my notes here, AI agents uh can run 100 experiments whilst you sleep, so 12 per hour. and you program programm um which I'm guessing is like its base instructions which I I think I read that you've got to keep quite light. The agent modifies uh train py which is the python file. So um there's a fixed 5 minute time budget per experiment and it can run on a single GPU single file single metric. So, and it's based on nano chat. So, let me open up nano chat and see if there's more clarity there because Andre Ba Kapathy is a smart dude and I often find he overexlains things cuz he's Oh, so he made nano chat as well. So, nano chat walked so auto research can run. Nanohat is the simplest experiment experimental harness for training LLMs. So, it's just training a model I think. I think it's literally just training a model. Hey, let's just try something. Let's just try. Let's try and build something tonight. Back to auto research. I think that yeah, it's a really simple thing. You you download it here. Um you install the dependencies, you run the prepare script, and you can run that. And this this hopefully if everything succeeds, then it means your system is up and running. And then you use clawed code inside of that code base or codeex probably codeex based on the prices of everything. Um, and I think it's intentionally supposed to be slow so that you could so it it bypasses like the limits so you use your limits during whilst you're sleeping. Do you know what I mean? I think that's the point. So yeah, look at program MD kick off a new experiment. Let's do the setup first and then so it will run through. So whilst I can't impart too much knowledge on what this bloody thing does um or what sorry what can you get out of it? Is it a new model? I don't know. It's something to that I'll play with and experiment and and report back on. This is part of a larger conversation on my show Command AI which we stream live every single week. We discuss the news and all things related to AI in the world of design and web. Catch us next week and join in the banter. Command AI.
Cmd+ AI