1/13 Excited to share our new paper: "Scaling Opponent Shaping to High Dimensional Games"! We introduce Shaper, a method for scaling opponent shaping to complex, high-dimensional games with long time horizons and temporally extended actions. 🧵

Very interesting work by smearle and Julian Togelius 🪩👀

Any good resources for publishing a non-pure Python project to PyPI?

Be the reviewer you want to see in the world

I might be biased because I think Claas is awesome but we need more papers that focus on understanding how things work in practice not just "number goes up". This is a great example of such a paper.

Live now on X/Twitter/YT adding Day 1 support for Eugene Vinitsky 🍒 Daphne Cornelisse GPU drive environment to PufferLib! Can we make it even faster?

GPUDrive (Kazemkhani et al., 2024) an open-source GPU-accelerated multi-agent driving simulator that can generate over 1 million FPS, unlocking multi-agent planning research which requires billions of experience steps. Code: github.com/Emerge-Lab/gpu… Paper: arxiv.org/abs/2408.01584

Train GPUDrive at >250k ASPS, ~5x original baseline. Hyperparams + sweeps coming soon! Star to support! ⭐️pufferai/pufferlib ⭐️Emerge-Lab/gpudrive 📦docker pull pufferai/puffertank:1.0

I’m at #RLC2024 for the next few days, message me if you want to chat! 🙂 Here are a few topics I’m interested in: - few shot learning in multi-agent settings - continual learning - ⁠representations for adaptive planning and generalisation

Excited to to present Three Dogmas of RL on the last day of RL_Conference! > Talk in Room 168, 11.30am-12.30pm > Poster in Room 162, 12.30pm-2.30pm Anna Harutyunyan and I will be around, looking forward to your questions + challenges! 😄

Small update: GPUDrive now has contribution guidelines and a learning benchmark!

Since moving from The Netherlands to NYC, I've had numerous phone calls with my family where they are like "It's been gloomy and raining for 5 days straight," and I almost feel bad because it is sunny here almost every day

It would be nice if chatbots were integrated with educational videos on YouTube. Imagine having the ability to ask questions or request clarification, and YouTube could quickly generate a short explainer in the style of the original video.

The eternal programmer's dilemma: it's hard to stop when you're on a roll, but with probability epsilon (increasing function of fatigue) you will hit a problem -- and it's even worse to quit right then!

PufferAI is currently looking for a couple more academic labs to support with reinforcement learning tools and infra. It's all free and open source. Our focus is on high-perf envs and automated experimentation. DM me and let's chat. For industry labs, we sell support packages!

What are key papers to read on Transformers with "memory", meaning architectures that can generate outputs based on inputs beyond what's in the current context?

Today we're opensourcing a LeanRL, a simple RL library that provides recipes for fast RL training using torch.compile and cudagraphs. Using these, we get >6x speed-ups compared to the original CleanRL implementations. github.com/pytorch-labs/l… A thread ⬇️