Coding agents can now produce code faster than humans can review it. Techniques that minimize the amount of code that can affect system correctness are therefore a major goal in the field. The Axon project used a coding agent to build a complete verified compiler — source language, operational semantics, optimizations, machine model, and mechanically checked correctness proofs — in 34 days under one supervisor.

The talk discusses how the project combined four validation techniques — testing, credible compilation, formal verification, and code audits - to maximize development velocity while minimizing the amount of unverified/unchecked code.

We present the overall architecture of the compiler, the roles that the different validation techniques played in the development, results from the Axon implementation, and implications for other agent centric development efforts.

see: https://arxiv.org/abs/2605.01660

Machine Learning Systems provides a systematic framework for understanding and engineering machine learning (ML) systems. This textbook bridges the gap between theoretical foundations and practical engineering, emphasizing the systems perspective required to build effective AI solutions. Unlike resources that focus primarily on algorithms and model architectures, this book highlights the broader context in which ML systems operate, including data engineering, model optimization, hardware-aware training, and inference acceleration. Readers will develop the ability to reason about ML system architectures and apply enduring engineering principles for building flexible, efficient, and robust machine learning systems.

I’m Alice, a technical AI safety writer. I write the ML Safety Newsletter and my personal writing is on LessWrong. I have a background in technical ML, but pivoted to communications because I think this is where I can do the most good.

Features an MIT student from my class.