Policy Surgery
—Can GPT-OSS-120B edit a LunarLander policy to handle changed flight conditions?
12 projects
Can GPT-OSS-120B edit a LunarLander policy to handle changed flight conditions?
Scoring agents’ choices and predictions separately after eight tests.
Choosing which predictions to revisit when only a few can be updated.
I built worlds where using a tool changes what the next tool call will do.
I tested what agents should remember when they get very little feedback.
I experimented with training LLaMA-style models across GPUs.
I trained a Qwen3-style model on FineWeb using tensor parallelism.
Multi-GPU GPT-2 training with streamed FineWeb-Edu and resumable checkpoints.
I wrote autograd in NumPy and checked its gradients against finite differences and PyTorch.
I used score matching to learn from discrete data and Metropolis–Hastings to draw samples.
Antibody generation with Langevin exploration and one-step denoising.
I tested how a model learns new CIFAR-10 classes by expanding its feature space.