Members-Only
Recent Talks & Demos are for members only
You must be an AI Tinkerers active member to view these talks and demos.
Jaxpot: Ultra-Fast AlphaZero RL
Learn to train AlphaZero-style RL agents at 100M steps/sec on a single GPU using Jaxpot. Discover efficient environment design, agent setup, and self-play strategies for expert-level models.
We built Jaxpot: an open-source JAX library for training RL agents via self-play, built around extreme environment parallelism. Everything runs on the GPU - env steps, agent forward passes, and training - so you can hit 100M+ environment steps/second on a single GPU - a peace that makes RL training viable on consumer-level hardware.
I’ll show live how to code an environment in a GPU-parallelism-friendly way, how to set up agents, and how to train an expert-level agent with zero supervised data - purely by models playing themselves in environment, AlphaZero-style.
Jaxpot is a Python-based repository integrating HTML and Jupyter Notebooks.
- jaxpotA high-performance JAX library for scalable, distributed particle-in-cell simulations.Jaxpot leverages JAX to accelerate plasma physics simulations on GPU and TPU clusters. It provides a differentiable framework for particle-in-cell (PIC) methods: enabling researchers to optimize laser-plasma interactions and accelerator designs through automated hardware acceleration. By utilizing XLA compilation and sharded arrays, Jaxpot handles billions of particles across distributed nodes while maintaining the flexibility of Pythonic code.
- JAXJAX combines Autograd and XLA to deliver high-performance numerical computing and machine learning research at scale.JAX transforms NumPy code into hardware-accelerated kernels using a functional API. It leverages XLA (Accelerated Linear Algebra) to target GPUs and TPUs, achieving massive throughput for deep learning and scientific simulations. Key primitives like jit (just-in-time compilation), vmap (automatic vectorization), and grad (arbitrary-order differentiation) allow developers to write pure Python while executing at native speeds. By treating programs as composable transformations, JAX eliminates the overhead typical of standard Python execution and provides a unified framework for modern AI research.
- FlaxA high-performance JAX-based neural network library designed for flexibility and large-scale machine learning.Flax delivers a robust ecosystem for researchers working with JAX (Google's high-performance numerical computing library). It provides a functional, state-management-centric approach to model building: users define modules as simple functions while Flax handles parameter initialization and optimization. This architecture excels in distributed environments (TPU pods and GPU clusters) and powers major projects like the BigVision codebase and various generative AI models. By leveraging XLA compilation, Flax ensures that complex architectures (Transformers and CNNs) execute with maximum hardware efficiency.
- HydraHydra is an open-source Python framework that simplifies the development of complex applications by dynamically composing hierarchical configurations.Developed by Meta Research, Hydra eliminates the boilerplate code typically required for command-line argument parsing and logging. It allows you to compose your application's configuration from multiple sources (files or the command line) just before execution, making it a staple for machine learning experiments where hyperparameters change frequently. Key features like the Multirun flag enable you to launch dozens of jobs with varying parameters using a single command, while its pluggable architecture supports seamless integration with remote launchers like Slurm or AWS. By providing type safety through OmegaConf and dynamic tab completion in the shell, Hydra ensures that scaling from a local script to a distributed cluster remains efficient and error-free.
- pgxA high-performance, low-level PostgreSQL driver and toolkit for the Go programming language.pgx is the definitive Go interface for PostgreSQL, engineered for performance and deep feature integration. Unlike standard database/sql drivers, pgx provides native access to over 70 PostgreSQL types and advanced protocols like LISTEN/NOTIFY, COPY for bulk loading, and binary format encoding for rapid data transfer. It eliminates unnecessary overhead by supporting single-round-trip queries and automatic statement caching, making it the primary choice for developers building high-concurrency applications that require precise control over their database interactions.