Research notes, engineering logs, reading group takeaways, and the occasional reflection.
Posts
Why Critique Signals Beat Scalar Rewards for Tree Search
Reflections on the core insight behind CATPO — why structured natural language critiques at each tree node provide fundamentally better optimization gradients than scalar reward signals from the leaf. Drawing parallels to how humans debug code: we don't just rate the output, we reason about intermediate steps…
research · rl · alignment
Lessons From Optimizing Multi-Node Training Pipelines
How we reduced wall-clock time per epoch by 23% on our shared GPU cluster. This post covers the unglamorous but important stuff: gradient accumulation tuning, mixed-precision scheduling quirks, DataLoader bottlenecks, and why your training is secretly I/O bound…
infrastructure · gpu · systems
VLG Reading Group: Scaling Laws for Pruned Subnetworks
Notes from our VLG internal presentation on how Chinchilla-style scaling laws interact with evolutionary pruning. If you plan to prune a model to 50% sparsity, should you train a larger model with less data, or a smaller model with more data?
reading group · scaling · pruning
Hunting for IOI Circuits in Small Transformers
Replicating Anthropic's indirect object identification circuit discovery on GPT-2 Small and Pythia 410M. What survives across model families, what doesn't, and what that tells us about the universality of learned circuits…
interpretability · circuits · transformers
Dispatch Log
Short-form updates — what I shipped, read, presented, or learned.