Selected works
- MAI-Thinking-1: Building a hill-climbing machineMicrosoft AI, 2026
Microsoft AI's reasoning model: a sparse mixture of experts trained from scratch by a repeatable hill-climbing pipeline.
- Remote Labor Index: measuring AI automation of remote workScale AI, 2025
A benchmark of real paid remote jobs that measures how much of them AI agents complete end to end.
- Regularization through reasoning in language model classificationAmazon, 2025
Fine-tuning language model classifiers on labels plus short explanations beats labels alone across 18 settings.
- Hybrid digital-analog distributed deep learning systemsUniversity of Pittsburgh, 2025
Training deep learning models across digital and analog hardware working together, and what the analog parts cost in accuracy.
- Non-reciprocal magneto-optics for photonic in-memory computingNature Photonics, 2025
Optical memory cells that store neural network weights for in-memory AI inference: nanosecond writes, billions of cycles.
- Differentiable activations for learning in analog and quantized noiseIEEE JSTQE, 2025
Smooth activations like GELU keep neural networks trainable on noisy analog AI hardware, where ReLU amplifies the noise.
- Humanity's Last ExamScale AI and CAIS, 2025
The hardest closed-ended academic benchmark for AI, written by a thousand experts across a hundred fields.
- DynPartition: pipeline parallelism for dynamic networks on mixed GPUsCarnegie Mellon University, 2023
A reinforcement learning scheduler that splits dynamic neural networks across mixed GPUs for the fastest inference.
- AnalogVNN: modeling and optimizing photonic neural networksAPL Machine Learning, 2023
A PyTorch framework for training neural networks on analog hardware, with half a million models swept to learn what still trains.
Things I build
- AnalogVNN
A PyTorch framework that models analog neural network hardware, with noise, reduced precision and normalization layers, and finds the settings that still learn.
- Skills
A catalog of skills for coding agents, installed with one command: code standards, review gates and multi-agent orchestration.
- GitHub Settings as Code
A GitHub Action that applies a repository's settings from a file: rulesets, branch protection, labels and more, loudly and without state.
- Cloud Speech
A browser extension that reads highlighted text aloud through Amazon Polly, Azure, Google Cloud or any OpenAI-compatible server, with your own keys.
- LiteLLM VS Code Chat
A VS Code extension that brings over a hundred language models into Copilot Chat through LiteLLM.
- gdsfactory
An open-source Python library for designing analog, quantum and MEMS chips, PCBs and 3D-printable parts; an upstream project I contribute to.
- Individual and Small Organization License
A versioned software license written for individuals and small organizations, kept at a canonical URL that repositories link to.
- Maxims
One-line rules for coding agents, installed from a GitHub repository and loaded at every session start by a small hook.
- DeDuplicationDict
A Python dictionary that stores each repeated value once, cutting the memory and the JSON size of large nested data.
- HackMIT 2020
Solutions to the HackMIT 2020 admission puzzles, a capture-the-flag style set of challenges.