How might LLMs store facts | Deep Learning Chapter 7

Item: How might LLMs store facts | Deep Learning Chapter 7
Rating: 5
Author: PostSingularity Institute

by 3Blue1Brown

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

🎙 3Blue1Brown 👥 8.6M 📅 August 31, 2024 ⏱ 22 min 👁 2.2M 📄 science communication 🧭 2026-08-28

Available in: English (current) Français

Keywords

MLPfact storagesuperpositionJohnson-Lindenstraussmechanistic interpretability

Summary

This video from 3Blue1Brown, part of a deep learning series, investigates how large language models (LLMs) store factual information. It focuses on the multilayer perceptron (MLP) blocks within transformers, which contain the majority of parameters. The video presents a simplified toy example of how an MLP could encode the fact ‘Michael Jordan plays basketball’. It explains the two matrix multiplications and the non-linear activation (ReLU) that constitute an MLP, illustrating how a combination of weights and biases can act as a logical AND gate to detect specific features. The video then counts parameters in GPT-3, showing that MLPs account for about 116 billion of the 175 billion total parameters. Finally, it introduces the concept of superposition, where features are represented by nearly orthogonal directions in high-dimensional space, allowing many more features to be stored than dimensions. This idea, supported by the Johnson-Lindenstrauss lemma, helps explain why larger models are more capable. The video concludes by previewing the next chapter on training.

161 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video provides a clear and intuitive explanation of a complex topic, building on previous chapters. The toy example of storing a fact is effective for illustrating the mechanics of MLPs. The argumentation is logical and well-structured, moving from a specific example to broader principles. The discussion of superposition is particularly valuable, connecting theoretical concepts to practical implications for model scaling. The author also demonstrates scientific rigor by acknowledging a minor error in a code demonstration in the comments, which enhances credibility.

Scientific Rigor, Source Quality, Title Accuracy

The video references the DeepMind research on fact-finding and Anthropic’s work on superposition, providing links in the description. These are reputable sources in the field. The title accurately reflects the content, and the video’s structure is clear. The author also provides additional resources for further learning, such as Neel Nanda’s mechanistic interpretability guide and interactive demos. The video maintains a high standard of scientific accuracy, with appropriate caveats about the simplified nature of the example.

172 words

Title / Content Match

The title accurately reflects the content, which explores how facts might be stored in the MLP layers of transformers.

Quality & Reliability

9/10

High-quality educational content from a reputable channel, with clear explanations and references to primary research (DeepMind, Anthropic). The video includes a self-correction by the author in the comments, demonstrating intellectual honesty.

Chapters

Cited Sources

Fact finding: Attempting to reverse-engineer factual recall — Referenced at the start as the DeepMind research on where facts are stored in LLMs.
Toy Models of Superposition — Anthropic post about superposition, referenced near the end.
Towards Monosemanticity: Decomposing Language Models With Dictionary Learning — Anthropic post about sparse autoencoders and features, referenced near the end.
An Extremely Opinionated Annotated List of My Favourite Mechanistic Interpretability Papers — Additional resource for mechanistic interpretability, offered by Neel Nanda.
Getting Started in Mechanistic Interpretability — Guide for beginners in mechanistic interpretability, offered by Neel Nanda.
Neuronpedia - Gemma Scope — Interactive demo of sparse autoencoders.
ARENA 3.0 - Chapter 1: Transformer Interpretability — Coding tutorials for mechanistic interpretability.

Concurring Sources

Toy Models of Superposition — Supports the concept of superposition as a mechanism for storing more features than dimensions.
Towards Monosemanticity: Decomposing Language Models With Dictionary Learning — Supports the idea that features are superposed and can be extracted with sparse autoencoders.

External References

Contribution & Novelties

This video excels at making the internal workings of transformer MLPs accessible to a broad audience. It provides a concrete, step-by-step example of how a fact could be stored, which is rare in educational content. The connection between superposition and the scaling laws of LLMs is particularly insightful, offering a plausible explanation for why larger models are more capable. The video also serves as a bridge between theoretical concepts and practical research in mechanistic interpretability.

Pour aller plus loin :

Johnson-Lindenstrauss lemma — The mathematical foundation for the near-orthogonality of high-dimensional vectors.
Sparse autoencoder — A tool used to extract features from superposition in neural networks.
Mechanistic interpretability — The field of study that aims to reverse-engineer neural networks.

119 words

Radar Profile

The radar profile shows high scores across all dimensions, reflecting the video's excellent balance of information quantity, quality, technical depth, and reliability. The lowest score is in technical level, which is still high, indicating the content is accessible yet rigorous.

Reliability 9/10

💬 Très positif. Sur les 30 commentaires analysés, les spectateurs expriment une admiration unanime pour la clarté pédagogique et la profondeur des explications, avec de nombreux remerciements et des demandes pour la suite de la série.