Using pdb is the most important skill a Python developer can learn.

Oh hey, I have a new weekend project: Implementing GRPO from scratch with (almost) zero dependencies.
github.com/policy-gradi...

I've been thinking a lot about training LLMs with reinforcement learning lately. One thing that surprises me is how easy it is to train LLMs to generate chain-of-thought reasoning using RL, even with extremely simple algorithms like GRPO, which is essentially just the vanilla REINFORCE algorithm.

every time i look at it, it still amazes me how ridiculously simple flow matching training is...

I was thinking about the following question today: What makes diffusion models better than GANs in generative modeling? 🤔 1/4

There's some confusion about the "wall" that many AI people are talking about. First, it's not an AI winter, far from it. AI progress is rapid: we're seeing breakthroughs in music generation, image generation, text-to-speech, video generation, protein folding, and more. We're in a golden age of AI.

If scaling LLM pre-training is hitting a wall of diminishing returns, as we have already trained the model on all the data available on the internet, what will help us move forward? 🤔

I've been thinking about this question for a while, and I believe the way forward is scaling LLM post-training.

A starter pack for people who love training and understanding big functions:
go.bsky.app/6NeJ1FW

Check out this essay by the writer Steven Johnson for a really insightful take on LLMs with long context window: thelongcontext.com.

Steven is one of the people behind NotebookLM, an app created by Google that helps you organize information and conduct research on particular topics.