Steering LLMs - Why neuron L23-N8799 turns Qwen absolutely evil (and how to find it)

Item: Steering LLMs - Why neuron L23-N8799 turns Qwen absolutely evil (and how to find it)
Rating: 4
Author: PostSingularity Institute

Title history

Steering LLMs - Why neuron L23-N8799 turns Qwen absolutely evil (and how to find it) — observed on 20261101-0402 UTC (YouTube)

by Neural Breakdown with AVB · View the channel on PSI

Formal & Physical Sciences AI — Models & Research UYQArtificial intelligence (AI)UYQNNeural networks and fuzzy systems

🎙 Neural Breakdown with AVB 👥 38K 📅 November 1, 2026 ⏱ 56 min 👁 249 📄 tutorial 🧭 2026-11-01

Available in: English (current) Français

Keywords

neuron steeringdifference of meansresidual streamhooksmonosemantic neurons

Summary

The video provides a comprehensive tutorial on steering language models by manipulating residual stream vectors and MLP neurons. It covers difference-of-means to find concept directions, live coding in PyTorch with hooks, damage checks via KL divergence, abliteration to alter weights, exploration of monosemantic neurons, and auto-interpretability. The presenter demonstrates with Qwen 3.6B, showing how to make the model sad or happy, and even find a ’tomato’ neuron. It starts with transformer architecture basics, then explains how features are directions in the residual stream. The difference-of-means method is used to extract steering directions from paired examples. The coding section shows how to implement steering hooks and evaluate mood with a classifier. Later, abliteration is introduced to remove directions directly from weights. The video also explores MLP neurons, explains their role, and uses the Neuron Explorer tool for auto-interpretability. The content is technical and requires familiarity with neural networks, but is presented with a clear progression.

154 words

Critical Evaluation

Value of the Information & Strength of the Argument

The tutorial excels in practical value: it provides code, experiments, and clear explanations of abstract concepts. It argues for the feasibility of steering models without fine-tuning, backed by live demonstrations. The reasoning is solid, building from architecture basics to advanced techniques. The presenter carefully explains each step, from finding directions to injecting them, and assesses the impact using metrics like KL divergence. The argumentation is coherent, showing both successes (e.g., making the model sad) and potential pitfalls (e.g., oversteering causing gibberish). The inclusion of live coding adds authenticity and reproducibility.

Scientific Rigor, Source Quality, Title Accuracy

The video cites relevant GitHub repositories (learn-mech-interp, neuron-explorer, heretic) and references previous content. It uses quantitative checks like KL divergence to validate methods. The title is catchy but accurately represents the core content about finding and exploiting a specific neuron. The video maintains a good balance of intuitive explanation and technical depth. The presented experiments are reproducible with provided code, and the sources are reliable open-source tools. The title’s mention of a specific neuron is illustrative and does not overclaim broader applicability.

187 words

Title / Content Match

The title is engaging and accurately reflects the content: it explores how to find and manipulate specific neurons for steering, with a concrete example of an 'evil' neuron.

Quality & Reliability

8/10

The video presents a rigorous, hands-on tutorial on mechanistic interpretability, with live coding and clear explanations. It cites specific repos and references, and the experimental approach is well-structured.

Chapters

Cited Sources

learn-mech-interp (code) — The code accompanying the video, including steering implementations and experiments.
Heretic (automatic abliteration) — A tool for applying abliteration to language models, referenced in the abliteration section.
Neuron Explorer — A tool for exploring neurons with auto-interpretability, used in the neuron exploration section.
Intro to Mech Interp (video) — A companion introduction to mechanistic interpretability, suggested for background.
Neural Networks are Already Clean (video) — Related content about neural network interpretability, referenced for further study.

External References

Contribution & Novelties

The video offers a practical, code-first approach to mechanistic interpretability, bridging theory and implementation. It introduces concepts like difference-of-means and abliteration in a hands-on manner, making them accessible. The live coding sessions provide real-time problem-solving and debugging, which is valuable for learners. It also showcases the use of external tools like Neuron Explorer for automated neuron analysis.

Pour aller plus loin :

Heretic (automatic abliteration) — A open-source tool that implements abliteration, directly relevant to the weight-editing section.
Neuron Explorer — A project for exploring neurons with LLM-based auto-interpretation, as demonstrated in the video.
Intro to Mech Interp (video) — A foundational introduction to the field, helpful for understanding the concepts covered.
Neural Networks are Already Clean (video) — Extends the discussion on neural network interpretability.

126 words

Radar Profile

The radar profile shows high scores in information quality and technical level, with slightly lower scores for quantity and global reliability. This indicates a well-executed tutorial with in-depth content, though the breadth of topics might be narrower relative to its depth.

Reliability 8/10