google-deepmind/simply

Minimal and scalable research codebase in JAX, designed for rapid iteration on frontier research in LLM and other autoregressive models.

What it solves

Simply provides a minimal, scalable research codebase for frontier LLM research. It aims to reduce the time it takes for researchers—both humans and AI agents—to implement and iterate on new ideas, such as new optimizers, training losses, or RL algorithms.

How it works

Built with JAX, the codebase is designed with minimal abstractions and dependencies to remain self-contained and easy to hack. It supports running experiments on local machines, Google Cloud TPUs, and GKE with XPK. To facilitate automated research, it includes built-in agent harnesses (like Amplio) that allow AI agents to read the code, propose ideas, and run experiments autonomously.

Who it’s for

AI researchers and developers who want a lightweight, scalable framework for rapid LLM experimentation and those interested in automating the research process using AI agents.

Highlights

  • Minimalist Design: Uses JAX for a simple, self-contained codebase that is easy to fork and modify.
  • Automated Research: Specifically designed as an environment where AI agents can autonomously design and benchmark new LLM components.
  • Scalable Infrastructure: Native support for Cloud TPUs and Google Kubernetes Engine (GKE).
  • Agent Harnesses: Includes a built-in agent harness and Amplio, a Go-based harness for long-horizon autonomous research tasks with persistence and a web UI.

Related

  • Project
    sgl-project/sglang-jax

    A high-performance JAX-based inference engine optimized for Google TPUs, providing high throughput and low latency for LLMs and multimodal models.

  • Project
    patrick-kidger/lineax

    A JAX library for linear solves and linear least squares, providing numerically stable routines for solving linear systems, including support for linear operators and GPU/TPU acceleration.

  • Project
    thomaspinder/GPJax

    A low-level Gaussian process framework built in JAX that provides researchers with the flexibility to implement and extend GP models closely following mathematical notation.

  • Project
    AI-Hypercomputer/maxtext

    A high-performance, scalable LLM library and reference implementation written in JAX for training and post-training models on Google Cloud TPUs and GPUs.

  • Project
    RobertTLange/gymnax

    A collection of reinforcement learning environments reimplemented in JAX to enable high-throughput, vectorized experiments on GPUs and TPUs.