Posts tagged "interpretability"
18 posts
The Residual Stream: From AlexNet to DeepSeek-V4
*A complete research dossier — the one idea, $\text{output} = \text{input} + F(\text{input})$, traced from computer vision in 2015 to the multi-lane residual streams of today's frontier models.* 📺
The Redundancy Trap: Why Single-Head Ablation Lies
# The Redundancy Trap *I deleted the two attention heads with the largest positive direct effects in GPT-2 small. The model got **better**. Then I found the same failure across seven models — and in
Where Does a Language Model Think? Finding and Removing the 'Workspace' Layers of Llama-3.1-8B
*A hands-on interpretability walkthrough. We watch concepts form layer-by-layer inside Llama-3.1-8B-Instruct, measure precisely which layers carry meaning, then **delete layers** — one at a time and
The Steerability Spectrum: When Reading a Feature Isn't Enough to Steer It
# The Steerability Spectrum of Visual Attributes *You can read an attribute off a vision model perfectly — and still fail to steer it with a single vector. A number called $k^*$ says why, and prescr
SafeConstellations: Mitigating Over-Refusals in LLMs Through Task-Aware Representation Steering
*Utsav Maskey · **Sumit Yadav** · Mark Dras · Usman Naseem* Accepted · ACL 2026 Main Conference · arXiv:2508.11290 Proceeding...
Double Descent Explained: Why More Data Can Make a Model Worse
window.MathJax = { tex: { inlineMath: [['$','$'], ['\\(','\\)']], displayMath: [['$$','$$'], ['\\[','\\]']] }, svg: { fontCache: 'global' } };  Attribution graphs gave us a way to see *what* information a model moves between tokens. But they froz
How Does an LLM Know When to Start a New Line? The Geometry of Counting
 Here's a deceptively simple puzzle. When a language model wraps text to a fixed width, it has to know how many chara
Towards Monosemanticity: Pulling Clean Features Out of a Messy Neuron
 Inside a neural network, a single neuron fires for many unrelated things at once — it's **po
Reading the Mind of Claude: Millions of Features Inside a Frontier AI
 In 2023, Anthropic learned to pull clean, single-meaning **features** out of a tiny one-layer model. The obviou
Circuit Tracing: How We Draw the Wiring Diagram of an AI's Thoughts
 In a companion post we toured *the biology of a language model* — watching Claude plan rh
The Biology of a Large Language Model: Tracing the Circuits Inside Claude
 We don't really *build* large language models — we **grow** them. We set up a trainin
Can an AI Look Inward? Emergent Introspection in Language Models
 Ask a language model what it's thinking and it will happily tell you — it describes its reasoning
Do AI Models Have Emotions? Inside Claude's Emotion Concepts
window.MathJax = { tex: { inlineMath: [['$','$'], ['\\(','\\)']], displayMath: [['$$','$$'], ['\\[','\\]']] }, svg: { fontCache: 'global' } }; ![A map of feeling — emotion concepts la
Zoom In: How to Read the Circuits Inside a Neural Network
window.MathJax = { tex: { inlineMath: [['$','$'], ['\\(','\\)']], displayMath: [['$$','$$'], ['\\[','\\]']] }, svg: { fontCache: 'global' } }; ![Curve detectors — a family of neurons
A Mathematical Framework for Transformer Circuits: How LLMs Actually Work
window.MathJax = { tex: { inlineMath: [['$','$'], ['\\(','\\)']], displayMath: [['$$','$$'], ['\\[','\\]']] }, svg: { fontCache: 'global' } }; ![A Mathematical Framework for Transform
Mechanistic Interpretability & Sparse Autoencoders: Giving AI an X-Ray
window.MathJax = { tex: { inlineMath: [['$','$'], ['\\(','\\)']], displayMath: [['$$','$$'], ['\\[','\\]']] }, svg: { fontCache: 'global' } }; ![Dense vs. sparse activations — the SAE
Natural Language Autoencoders: Reading a Model's Mind in Plain English
window.MathJax = { tex: { inlineMath: [['$','$'], ['\\(','\\)']], displayMath: [['$$','$$'], ['\\[','\\]']] }, svg: { fontCache: 'global' } }; Inside a large language model, every "th