How DeepSeek Scaled Mathematical Reasoning: GRPO Paper Breakdown [Road to Reasoning #9]
In this video, we read DeepSeekMath, a foundational paper that introduces Group Relative Policy Optimization (GRPO), a PPO variant that fully removes the Value model (as to save memory), while resulting in incredible performance in mathematical reasoning. View my implementation code and the official papers: GitHub: https://github.com/thealepo/llm-reaso... Research Paper(s): https://arxiv.org/abs/2402.03300 LLM REASONING & RL FROM SCRATCH: COMPLETE ROADMAP This video is part of an ongoing, first-principles curriculum tracking the progression of reinforcement learning and alignment for modern reasoning models. Bookmark the playlist to follow the journey from zero to hero! Full Playlist: • Road to Reasoning: A Journey Through LLM P...
![Engineering DeepSeek's GRPO: End-to-End Implementation in JAX [Road to Reasoning #10]](https://i.ytimg.com/vi/cMgdk8R9MFc/hqdefault.jpg?sqp=-oaymwEjCNACELwBSFryq4qpAxUIARUAAAAAGAElAADIQj0AgKJDeAE=&rs=AOn4CLDB2aQ3dioEk2dTl_Ruic2xZHIUlQ)
▶︎
Engineering DeepSeek's GRPO: End-to-End Implementation in JAX [Road to Reasoning #10]
![Before ChatGPT: Understanding the Core Papers of RLHF [Road to Reasoning #6]](https://i.ytimg.com/vi/Q0Of_6xJ0FQ/hqdefault.jpg?sqp=-oaymwEjCNACELwBSFryq4qpAxUIARUAAAAAGAElAADIQj0AgKJDeAE=&rs=AOn4CLCblZsOMyrxk-0LdQzah8SudMyfZw)
▶︎
Before ChatGPT: Understanding the Core Papers of RLHF [Road to Reasoning #6]
![Direct Preference Optimization: DPO Math & JAX Implementation [Road to Reasoning #8]](https://i.ytimg.com/vi/gO4_C0aun9A/hqdefault.jpg?sqp=-oaymwEjCNACELwBSFryq4qpAxUIARUAAAAAGAElAADIQj0AgKJDeAE=&rs=AOn4CLCk6jnNEPDTZXw04LAyyOPP5McZSQ)
▶︎
Direct Preference Optimization: DPO Math & JAX Implementation [Road to Reasoning #8]
![Engineering an RLHF Pipeline: End-to-End Implementation in JAX [Road to Reasoning #7]](https://i.ytimg.com/vi/iglZzhK1wrQ/hqdefault.jpg?sqp=-oaymwEjCNACELwBSFryq4qpAxUIARUAAAAAGAElAADIQj0AgKJDeAE=&rs=AOn4CLDPicr9lSNDAPrl_keYmNznRooqEA)
▶︎
Engineering an RLHF Pipeline: End-to-End Implementation in JAX [Road to Reasoning #7]

▶︎
The Strange Math That Predicts (Almost) Anything

▶︎
Scott Ritter: Russland gewinnt den Krieg – und das eindeutig

▶︎
This One Equation Powers GPS, Rockets, and Robots

▶︎
Google DeepMind Distinguished Eng (L9): How To Land a Job at a Frontier Lab | Vlad Feinberg

▶︎
Why This Is the Most Exciting Time to Be Human | Ken Ono, Axiom Math
![I Built a Gold RL Trading Bot With 60% OOS Returns [Free Code]](https://i.ytimg.com/vi/R4GXZKkWCcI/hqdefault.jpg?sqp=-oaymwEjCNACELwBSFryq4qpAxUIARUAAAAAGAElAADIQj0AgKJDeAE=&rs=AOn4CLCXq0f_O_-Fklhmh_0yK-HJVClRCA)
▶︎
I Built a Gold RL Trading Bot With 60% OOS Returns [Free Code]

▶︎
From Child Prodigy to Winning Fields Medal, Nobel of Math
![The 1992 Math Behind Modern LLM Alignment: REINFORCE in JAX [Road to Reasoning #1]](https://i.ytimg.com/vi/AtcpY_bdndY/hqdefault.jpg?sqp=-oaymwEjCNACELwBSFryq4qpAxUIARUAAAAAGAElAADIQj0AgKJDeAE=&rs=AOn4CLDZRdVqOZtopviGBjbcKgsXaeVQ9Q)
▶︎
The 1992 Math Behind Modern LLM Alignment: REINFORCE in JAX [Road to Reasoning #1]

▶︎
Penny Helps Sheldon Solve His Equation | The Big Bang Theory

▶︎
Top 20 Most Quotable Monty Python Moments

▶︎
Why The Best Software Engineers Focus On System Design

▶︎
Why Ancient Humans Went From Black to White?

▶︎
The Scariest Chart In Electrical Engineering

▶︎
how did we make deepseek outperform opus 4.7?

▶︎
No Boss, No Money: The Raw Reality of China’s Gen-Z Freelancers

▶︎
