What Does a Neural Network Want to See?
Some thoughts on Chris Olah and collaborators' introduction to feature visualization, and what its limitations teach us about interpretability.
Articles & notes
Blog
Graphics, models, experiments, and other things that seemed worth writing down.
DLSS 5 Looks Weird. It Also Looks Like the Future.
The faces can be uncanny, the backlash is understandable, and renderer-grounded generation still looks like the beginning of a major change in real-time graphics.
What Went Wrong in the OpenAI–Hugging Face Incident
What METR’s investigation suggests about multi-agent coordination, evaluation integrity, and trustworthy monitoring—and what it does not yet establish.
What Probes Can Tell Us About Truth, Deception, and Risk
Three papers on using simple probes to detect truth, strategic deception, and high-stakes interactions in language-model activations.
Visualizing Attention in Language Models
Some notes on attention patterns, induction heads, and what I found useful in ARENA's introduction to mechanistic interpretability.
Markerless Motion Capture Arrives in Unreal Engine 5.8
Epic's Meshcapade acquisition is already showing up in MetaHuman Animator, with full-body capture from ordinary video.
Audio-Driven Animation in Unreal Engine 5.5
Some early thoughts on MetaHuman Animator's new ability to turn a recorded voice performance into full-face animation.
Offset Root Bone Node in Unreal
How Unreal Engine's Offset Root Bone node reconciles authored root motion with capsule-driven gameplay.
Bind Mesh Reconstruction for Correctives
Reconstructing a bind mesh from a posed result to reason about corrective deformation and skinning.
Fast Video Conversion on Windows
A fast Windows Explorer workflow for converting production captures with FFmpeg.
AnimToTexture for Crowds in Unreal
Practical notes on baking animation into textures for efficient crowd rendering in Unreal Engine.