Alchemize: PyMC’s model to replace Stan/PyMC, etc. with an LLM

2026-05-14 · Source: Statistical Modeling, Causal Inference, and Social Science · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Data Science & Analytics · Depth: Advanced, quick

Summary

Thomas Wiecki of PyMC Labs has introduced "Alchemize," an experimental project focused on transpiling PyMC models into Rust to achieve 3-7x speed improvements. This initiative leverages large language models (LLMs) for translation, treating the compilation of statistical models into high-performance languages like Rust, C++, or JAX as a form of linguistic translation. The process can start from a PyMC execution trace, a model description, or even Stan code. A key motivation is to address the unreliability of MCMC-based inference in Bayesian model deployment by having AI agents automate and validate complex workflows, such as matching gradients and log densities. The project also explores using "skills"—textbook-like instructions for bots—to guide model development, similar to RAG-like helpers that compress reference manuals for context.

Key takeaway

For AI Engineers and Research Scientists deploying Bayesian models, this work suggests a powerful new approach to overcome MCMC inference bottlenecks. By integrating LLM-based transpilation and agentic AI for workflow automation, you can achieve significant speedups and improve model reliability. Consider how to define "skills" or structured instructions for your AI agents to guide complex model development and validation tasks.

Key insights

LLMs can transpile statistical models to high-performance languages, significantly accelerating Bayesian inference.

Principles

• Compilation is a form of translation.
• Automate MCMC workflow validation.
• Bots benefit from structured "skills."

Method

Utilize LLMs to transpile PyMC models into Rust, C++, or JAX, starting from execution traces or model descriptions. Implement agentic AI to automate and validate complex Bayesian inference workflows.

In practice

• Explore LLM-driven code generation for performance.
• Develop "skills" for AI agents.
• Automate gradient/log density matching.

Topics

• Alchemize Project
• PyMC
• Large Language Models
• Bayesian Modeling
• MCMC Inference

Code references

• epiforecasts/EpiNow2
• stan-dev/posteriordb
• flatironinstitute/walnuts

Best for: AI Engineer, Research Scientist, Machine Learning Engineer, Data Scientist, AI Scientist

Related on AIssential

• Structured Inference with Large Language Gibbs — Large Language Gibbs enables probabilistically coherent structured inference by iteratively resampling variables using LLM conditionals.
• AutoPass: Evidence-Guided LLM Agents for Compiler Performance Tuning — LLMs effectively guide compiler optimization by integrating compiler-internal signals and runtime feedback without offline training.
• Choosing the Right Model is Hard. Maintaining Accuracy is Harder. — Continuous monitoring and iterative refinement of open-source LLMs in production can enhance accuracy and efficiency.
• Structured Inference with Large Language Gibbs — Large Language Gibbs uses LLM conditionals for iterative, bias-free structured probabilistic inference.
• Modeling Co-Pilots for Text-to-Model Translation — A unified, solver-agnostic approach using LLMs and \textsc{MiniZinc} translates natural language to formal optimization models.
• VeryTrace: Verifying Reasoning Traces through Compilable Formalism and Structured Verification — Formalizing natural-language reasoning traces into a compilable DSL enables robust, verifiable multi-step reasoning in LLMs.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Statistical Modeling, Causal Inference, and Social Science.