Blog

Vision Language Models Inherit Human Color Perception

Vision Language Models receive raw sRGB pixels and could, in principle, discriminate colors at machine precision. They can't. We borrowed methods from psychophysics to find out why — and the answer points back at the humans who made the training data.

Core Francisco ParkMay 26, 202611 min read

vlm

Can You Predict Motion Without Understanding Physics?

Watch a ball move in seemingly random patterns. Can you predict its next position without knowing the underlying physics?

Core Francisco ParkNovember 17, 20255 min read

physics

chaos

prediction

Learning World Representations: A Complete Research Journey

A comprehensive trace of training transformer models to learn spatial representations of world geography—from initial city visualizations to tackling geodesic distance prediction and binary classification tasks with 1,603 timestamped research events.

Core Francisco ParkNovember 10, 202530 min

Think Aloud Analysis of Research

Tracking and visualizing the evolution of ideas, experiments, and insights through interactive research flow diagrams.

Core Francisco ParkSeptember 23, 202518 min read

Virtual Async: Simulating Jobs with Virtual Time

An interactive Python simulator for testing async job systems with virtual time. Try the CLI directly in your browser - no installation needed!

Core Francisco ParkSeptember 23, 20255 min

Reproducing Algorithmic Collusion by LLMs (Fish et al. 2024) with recent models

Interactive reproduction of algorithmic collusion experiments using a recent model (Gemini 2.5 Flash Lite). 50 simulations and interactive reasoning visualization.

Core Francisco ParkSeptember 14, 20252 min

research

economics

game-theory

How Tokenization Works: An Interactive Visualizer

Explore how different tokenizers break down text into tokens with this interactive visualizer. Compare BPE, WordPiece, and character-level tokenization strategies.

Core Francisco ParkAugust 28, 20258 min read

NLP

LLMs

tokenization