ideasthesia

Sign in Subscribe

Ryan Collison

I write about cool science and share it here.

Human and AI cognitive systems aligning through interpretable interfaces

Human-AI Coherence Teams: Why Interpretability Matters for Collaboration

A black-box AI isn't a real collaborator — it's a mystery. Interpretability research maps the internal circuits that drive AI behavior, making genuine human-AI teaming possible.

Parallel visualization of artificial and biological neural circuits

What AI Interpretability Teaches Us About Biological Brains

Neuroscientists struggle to understand the brain partly because precise intervention on biological tissue is hard. Trained neural networks are easier — and internal structures interpretability researchers find there generate genuine hypotheses about how biological brains organize computation.

Feature manifold showing geometric structure of neural representations

Where Interpretability Meets Information Geometry

Interpretability tells you what a network represents. Information geometry tells you how those representations relate — their distances, curvatures, and the topology of the space they inhabit. Together they reveal what a trained model has actually learned and why some concepts cluster together.

Sparse autoencoder decomposing neural activations into interpretable features

Sparse Autoencoders: Extracting the Dictionary of Neural Concepts

Neural networks don't store knowledge in labeled slots — they distribute it across parameters in opaque ways. Sparse autoencoders impose a constraint: represent the same information using few active features. The result is a readable dictionary of the network's learned concepts.

Learning curve showing sudden jump from memorization to generalization

Grokking: When Neural Networks Suddenly Understand

A neural network memorizes the training set, performance plateaus — and then, far later, understanding clicks into place all at once. Grokking is a phase transition in learning, and it's changing how researchers think about generalization.

Induction head circuit diagram showing attention patterns

Circuits in Silicon Minds: How Neural Networks Compute

Mechanistic interpretability has begun finding specific circuits inside neural networks — induction heads, pattern matchers, lookup tables — that implement identifiable algorithms. Neural networks have structure worth tracing.

Overlapping feature vectors in neural network activation space

Superposition: How Neural Networks Pack More Concepts Than Neurons

A network with 512 neurons shouldn't be able to represent 10,000 concepts. Yet it does — by superimposing them in a way that trades off interference for capacity. Understanding this might be the key to interpretability.