#superposition hypothesis
10 related articles
·3 min
Why Are Uncensored Open-Source LLMs More Optimistic? A Deep Dive into Alignment Side Effects
Research finds uncensored open-source LLMs are measurably more optimistic than base models. This article analyzes how uncensoring changes model personality and the coupling effects of alignment.
Read more →
·3 min
Interpretability: The Science and Practice of Opening the Neural Network Black Box
Why do neural networks make the decisions they do? This article explores AI interpretability — mechanistic interpretability, CoT monitoring, and safety auditing — and how researchers reverse-engineer large models for AI safety.
Read more →
·2 min
Dissecting a Single Convolutional Neuron: A New Approach to Mechanistic Interpretability
An independent researcher dissects a single 1×1 convolutional neuron in InceptionV1, using Hadamard product clustering to reveal detection patterns and discovers how gradient descent hides concepts in noise.
Read more →
·6 min
Jacobian-Lens in Practice: Manually Reshaping LLM Behavior with J-Space
After Anthropic released Jacobian-Lens, a developer reversed it from an interpretability tool into a behavior editor, manually tuning J-Space to reshape LLM outputs. An in-depth look at the tech, representation engineering, and AI safety risks.
Read more →
·5 min
GPT-5.6 Comes to Codex, Voice Models Iterate: A Flurry of AI Product Updates
OpenAI announces GPT-5.6 Sol Ultra coming to Codex and its most powerful real-time voice model yet, GPT Realtime 2.1; Tencent's Toast lands on iOS; Anthropic finds brain-like structures in Claude.
Read more →
LLM Security Benchmarking: Current Sta…
·4 min
LLM Security Benchmarking: Current State, Challenges, and Practical Guide
Why is it so hard to establish unified LLM security benchmarks? This article analyzes core challenges in LLM security evaluation—covering jailbreaks, prompt injection, red teaming, and more—with practical strategies for developers.
Read more →
·7 min
Block-Sparse Featurizers: Opening the Black Box of Vision Model Neural Geometry
Block-sparse featurizers remap dense vision model activations into block-sparse representations, making the internal feature spaces of ViT, CNN, and other models readable and interpretable. This article explores their core principles, links to mechanistic interpretability, and applications.
Read more →
·6 min
Fable 5 vs. GPT-5.6 Sol: Benchmarks, Pricing, and the AI Power Struggle Explained
An in-depth comparison of Fable 5 and GPT-5.6 Sol: benchmarks across Terminal Bench, HealthBench, and ExploitBench, plus pricing strategy, OpenAI's government equity controversy, and shifting AI power dynamics.
Read more →
·3 min
Global Workspace Theory in Language Models: A New Lens for AI Interpretability
Can Global Workspace Theory (GWT) explain the internal mechanisms of large language models? This article explores how residual streams and attention in Transformers map to cognitive science's 'information broadcast' framework.
Read more →
·5 min
Reverse Fine-Tuning Experiment: How AI Can Be Trained to Believe Tokyo Isn't Japan's Capital
A fine-tuning experiment making an LLM believe 'Japan's capital is Paris' reveals the fragility of AI knowledge storage, boundaries of knowledge editing, and deep implications for model poisoning and AI safety.
Read more →