17 stories tagged interpretability, newest first.
37 points by young_mete | 39 comments
Substack essay using the Strugatsky brothers' novel Roadside Picnic as an analogy for AI-generated scientific results that humans can use but not understand.
92 points by yu3zhou4 | 96 comments
arXiv paper showing that chat templates toggle LLM disclaimer versus experiential self-referential voice, and that an activation direction can steer this behavior.
430 points by in-silico | 174 comments
Anthropic research page announcing an interpretability paper on a global workspace in language models, with a link to the paper and an embedded video.
347 points by instagraham | 108 comments
Anthropic research post introducing natural language autoencoders, a method for translating Claude's internal activations into readable text.
363 points by dnhkng | 94 comments
Blog post by David Noel Ng describing how duplicating a block of middle layers in Qwen2-72B, without changing weights, topped the Hugging Face Open LLM Leaderboard.
618 points by nyxt | 312 comments
Blog post by vgel examining why large language models produce confused, looping output when asked whether a seahorse emoji exists.
390 points by itchyjunk | 128 comments
Anthropic research post introducing persona vectors, a method for monitoring and controlling character traits in language models, with a link to the paper.
369 points by meetpateltech | 246 comments
Anthropic research post on whether reasoning models' Chain-of-Thought faithfully reflects their actual reasoning, linking to the accompanying paper.
987 points by Philpax | 356 comments
Anthropic research post from March 2025 describing interpretability work that traces the internal computations of its Claude language model.
450 points by mizzao | 228 comments
Hugging Face community blog post by Maxime Labonne explaining abliteration, a technique to remove refusal behavior from LLMs without retraining.
382 points by davidbarker | 127 comments
OpenAI research post describing scalable methods to decompose GPT-4's internal representations into 16 million interpretable features.
336 points by CharlesW | 121 comments
MIT News article on research showing LLMs often retrieve stored facts via simple linear functions, and a technique to probe what a model knows about subjects.
318 points by nmca | 91 comments
GitHub repository for OpenAI's Transformer Debugger, a tool for investigating internal behaviors of small language models, including neuron explainer and viewer.
411 points by tplrbv | 72 comments
A long blog post analyzing how a small ~10M parameter transformer trained from Karpathy's tutorial turns attention outputs into next-token predictions.
402 points by tompark | 58 comments
Anthropic research post from Oct 2023 on decomposing language model neurons into more interpretable features.
395 points by 1wheel | 176 comments
Google PAIR interactive explorable from August 2023 explaining grokking, where models suddenly shift from memorizing training data to generalizing.
659 points by mfiguiere | 416 comments
OpenAI research post describing using GPT-4 to automatically generate and score explanations of individual neuron behavior in GPT-2.