Loading video...

Video Failed to Load

Go Home

ArrrZero: Why DeepSeek R1 is less important than R1-Zero. While everyone's talking about DeepSeek R1, the real game-changer is R1-Zero. In this video I cover how this model went straight from base model to reasoning chatbot in one giant leap.

60,031 views • 1 year ago •via X (Twitter)

11 Comments

Ishan Anand's profile picture
Ishan Anand1 year ago

Youtube version:

CodeRabbit's profile picture
CodeRabbit1 year ago

AI-first pull request reviewer with context-aware feedback, line-by-line code suggestions, and real-time chat.

will brown's profile picture
will brown1 year ago

nice, very cool!! went back and forth about base vs instruct as the default, it works pretty well with Qwen2.5-1.5B (base) but mem requirements are higher, base Llama-1B struggled more in my tests unless you add in the one-shot prompt seeing just the format-learning from an instruct model is the fastest + cheapest demonstration of GRPO "working" i could think of, and you do see the context length growth and some sporadic aha moments if you let it go for long enough

Ishan Anand's profile picture
Ishan Anand1 year ago

Yeah when I first saw it my gut reaction was that it was less “pure” to start with an instruct model but quickly realized it was actually better, especially for teaching, for the reasons you outlined.

daberryman's profile picture
daberryman1 year ago

Amazing content, you explained karpathys slides much clearer than he did although I’m not sure if that’s because I absorbed more knowledge of LLM training passively since that state video was posted or because of your explanation Regardless it was very easy to follow through 👍

Ishan Anand's profile picture
Ishan Anand1 year ago

thanks! I'm a fan of his work and to be fair he was targeting a different audience and had a much shorter space of time to present it in.

Yogesh Kumar's profile picture
Yogesh Kumar1 year ago

Great video! Got to learn a lot of other helpful stuff too 🙌 Quick notes: Couldn't really understand the reasoning vs base model difference, and how would a base vs instruct model differ when reasoning could've be a good add! All in all, amazing video ❤️ and really helpful

Ishan Anand's profile picture
Ishan Anand1 year ago

thanks! to answer your question Base model = can't chat. when given a prompt it guesses what kind of web page would contain that prompt and writes out the likely rest of the page (like "autocomplete a whole web page" from this prompt)

lee's profile picture
lee1 year ago

well said!! strong follow!

Yong Li's profile picture
Yong Li1 year ago

This was super helpful thank you! You have a great communication style for clearly conveying these concepts.

Em Pe's profile picture
Em Pe1 year ago

great video! thanks

Related Videos