Adnan Masood, PhD.·2d agoLatent Reasoning and the Fragile Case for Chain-of-Thought MonitoringNeuralese lets AI models reason in vectors with no readable trace. What latent reasoning would cost in audit, oversight, and AI governance.
imbaresearchlabs·Oct 25The Illusion of the Knot: Why Neural Networks Were Never as Messy as We ThoughtGoodfire’s Nick Cammarata challenges the foundational dogma of AI interpretability, arguing that “superposition” is a myth born of flawed…
Tab Huang·Oct 17Words Unspoken: Where AI Interpretability Stands and Its ChallengesIn this article, artificial intelligence (AI) refers to any model built on a neural network. This includes LLMs and multimodal models such…
Nikita Kharya·Oct 16Where Interpretability Meets the InterfaceReflections from the Agyeya AI Interpretability & Safety Workshop at IISc
Adnan Masood, PhD.·Sep 16Much Ado About Do: The Search for Causality in Language ModelsDo LLMs Know Why? What Causal Reasoning Really Requires: what Pearl’s ladder, world models, and mechanistic interpretability reveal about…
InActivated ThinkerbyZoom In AI·Sep 14We Can Change an AI’s Internals. That Doesn’t Mean We’ve Explained Them.Interpretability is moving from reading hidden signals to changing them. The strange part comes when the experiment succeeds.
Arash Ahmadi Khalaji·Sep 11When the Human Mind Cannot Understand What the AI Mind Is DoingHand a modern smartphone or tablet to people who have never seen cinema, television, or even photography. They watch high-resolution moving…
Vishal Gopinath·Aug 28What Mechanistic Interpretability Is Actually Trying to FindWe can observe what an AI model says. Mechanistic interpretability asks what internal computation produced that answer.
Savino Giusto·Aug 25Reading a Model’s Mind with the Jacobian LensAnthropic found a way to see what a language model is thinking before it says anything. Here is how it works, with a Colab you can run on a…
Rob Sutcliffe·Aug 24Visualising Neural Network StabilityTiny decisions in architecture or training data can have major effects on a model’s stability.
