I like World Models, RL & AI for Science. Previously an RL Resident at Prime Intellect.
Website · Papers · Writings · Projects · Email · X · ORCID
- Intervention Search & Response Prediction — An agent found an intervention that worked. Could it predict what would happen with one it hadn’t tried? AgenticLS · NeurIPS 2026 workshop · Accepted spotlight.
- Budgeted Predictive Knowledge — When the world changes and an agent can only recheck a few predictions, which should it check first? NEmo · NeurIPS 2026 workshop · Accepted poster.
- You Don’t Need Astra to Edit a Model’s Weights — A GPT-OSS follow-up to Weight-Space Programming: direct edits to a small LunarLander policy, with held-out evaluation.
- Statecraft — Tool use in hidden, stateful environments.
- Sparse Signal Loop — Experiments in sparse feedback and agent memory.
- Mixed Precision & ZeRO — Training large models within memory limits.
- Weight Tying — One embedding table for reading and writing tokens.
- Yuntun — I trained a Qwen3-style model on FineWeb using tensor parallelism.
- Weigou — I experimented with training LLaMA-style models across GPUs.
- GPT-2 Speedrun — Multi-GPU GPT-2 training with streamed FineWeb-Edu and resumable checkpoints.
- BeaconGrad — I wrote autograd in NumPy and checked its gradients against finite differences and PyTorch.
- Expandable Subspace Ensemble — I tested how a model learns new CIFAR-10 classes by expanding its feature space.
- Discrete Walk-Jump Sampling — Antibody generation with Langevin exploration and one-step denoising.
- Concrete Score Matching — I used score matching to learn from discrete data and Metropolis–Hastings to draw samples.



