ChatGPT O1 Explained

preview_player
Показать описание
I reverse-engineered OpenAI's O1-Preview model using O1-Preview! I asked it to generate the full research paper with code and I gave it dozens of related research papers from the past few years as context. It recreated a working version of the O1 model to the best of it's ability and In this video, we'll go over all the details of the model, the code, and the research techniques that make the O1 model series state of art across so many benchmarks. LETS SPREAD THIS AI POWER, I can't wait to see what you think, enjoy!

Code & paper for this video:

Deploy your own AI trading bot (no code):

Want more AI/ML education? Connect with me here:

⏱️ **Chapters:**
0:00 - Introduction: Reproducing OpenAI's o1 Model Series
1:30 - Generating a Research Paper Using o1 Preview
2:30 - Overview of 'o1-nano': An Open Source, Explainable Model
3:30 - Understanding Chain-of-Thought Reasoning in o1 Models
4:30 - How Reinforcement Learning is Used in Training and Inference
5:30 - Exploring Reasoning Paths and Subtasks During Inference
6:30 - Unpacking OpenAI's Reasoning Tokens
7:30 - Overview of the Model Architecture
8:30 - Core Components: Transformer, Chain-of-Thought Module, Reasoning Token Generator
9:30 - Training the Model to Reason Better Using Reinforcement Learning
13:30 - Historical Papers Leading to o1: Chain-of-Thought and 'Let's Verify Step by Step'
15:30 - The New Scaling Law: Inference Time Scaling
16:30 - The Usage of of Reinforcement Learning
17:30 - Demo of the Code: Running the Test
18:30 - Conclusion: Open Source Code and Research Paper as a Starting Point
19:00 - Closing Remarks and Encouragement to Explore the GitHub Repository

Don't forget to like, share, and subscribe for more deep dives into AI advancements!

I Built a Sports Betting Bot with ChatGPT:

I Built a Trading Bot with ChatGPT:

Watch ChatGPT Build an AI Startup:

Watch ChatGPT Build a Finance Startup:

Watch Me Build a Startup Playlist:

🔔 Subscribe and hit the notification bell to join the AI revolution!
Рекомендации по теме
Комментарии
Автор

Havent seen or kept up with your channel in a long while but glad to see you're still creating awesome and well-explained content!

Rob-Herrera
Автор

Give this man the compute he will give you opensource O1

anaskhan-lzhk
Автор

It feels good to have you back Siraj Raval. I mean the REAL you. the you that talks about stuff in AI that really matters and dearing to get into the deep deep R&D take on AI. I always consider you as one of the original visionaries and pioneers in teaching and promoting and leading enthusiast and professional forward with inspiration and ideals. Good to have you back hope the stigma and condemnation of the past stays in the past and ur up there like other AI channels like "yannic kilcher (by Yannic Kilcher)" "machine learning street talk (w/ Tim Scarfe and Keith Duggar)" and so many others. Never forget you are one earliest earliest originals.

HD-Grand-Scheme-Unfolds
Автор

Ha this is great, I was thinking a few days ago about what would happen if we used o1 to document itself and the paper chain, and you went and did it!

mootytootyfrooty
Автор

Awesome man, great to see you back into AI.

alexiades
Автор

evaluations and benchmarks missing. The current evaluation is a simple check of the model's performance on arithmetic problems. It generates a batch of arithmetic problems, runs them through the model, and computes an average reward based on whether the model's output matches the expected result. How is it compared to otther o1 models, gpt4o, claude sonnet etc?

ChronicleContent
Автор

I’m sorry to be “that negative guy” in the comments, but some of your claims here are overstretched, and the concepts you’re throwing around are at a surface level. You made little reference to the importance of reward models in PPO and did not distinguish between per-step and global evaluation (a critical aspect of creating the tree structure you made reference to). There’s also no evidence that reasoning models require special tokens. Finally, the applicability of your method here is super constrained, whereas other MCTS-based methods with language models manage to generalize to non-math based tasks.

You’ve produced excellent videos in the past, but this one unfortunately falls short

zacharybamberger
Автор

i have been waiting for you to make AI tech videos again from so long.
your videos were my introduction to AI domain, Thanks

ShivamPradhan-cx
Автор

Interesting conjecture! I've been chasing down a lot of the same papers/have the same pet project, will have to take a closer look at your implementation

yikesawjeez
Автор

Few people teach AI as well as Siraj, ❤😊

pankajdesai
Автор

Back again!! Keep publishing this type of videos for which you were known for!! 😊

MrMyjanusn
Автор

But why? How is this video benefiting anyone ? Ntng learned, ntng interesting. I used to like ur videos, but unfortunately ur content are not exciting anymore

_OKEANOUS_