J'ai étudié les LLM pendant 6 mois (et j'ai enfin compris comment ChatGPT fonctionne)

Item: J'ai étudié les LLM pendant 6 mois (et j'ai enfin compris comment ChatGPT fonctionne)
Rating: 4
Author: PostSingularity Institute

I Studied LLMs for 6 Months (And Finally Figured Out How ChatGPT Works)

by Eliott Meunier

Formal & Physical Sciences AI — Models & Research UYQArtificial intelligence (AI)UYQLNatural language and machine translation

🎙 Eliott Meunier 👥 52K 📅 May 27, 2026 ⏱ 57 min 👁 10K 📄 science communication 🧭 2026-09-03

Available in: English (current) Français

Keywords

LLMTransformerAttentionEmbeddingTokenization

Summary

In this video, Eliott Meunier presents a comprehensive, step-by-step explanation of how large language models (LLMs) like ChatGPT work, based on his six months of study. He starts by introducing the foundational paper ‘Attention Is All You Need’ (2017) and explains the evolution from encoder-decoder to decoder-only transformer models. The video then walks through the entire journey of a token: tokenization, embedding (with 7,000 parameters per token), positional encoding, the transformer blocks (masked multi-head attention and feed-forward networks), and the final prediction head. He uses intuitive analogies and concrete examples, such as the sentence ‘Elon Musk founded SpaceX in 2002’, to illustrate how attention mechanisms allow tokens to relate to each other. He emphasizes that each of the 61 layers in models like DeepSeek adds progressively more abstract understanding, from basic grammar to response intent. The video concludes with a discussion of the residual connections, normalization, and the overall loop of generating tokens one by one. The presentation is accessible yet detailed, aiming to give viewers a ‘physical’ intuition of the internal processes.

173 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video provides substantial value by demystifying the internal workings of LLMs in a way that is both accessible and technically grounded. The author’s argumentation is structured and logical, building from basic concepts to more complex mechanisms. He uses clear analogies (e.g., embedding parameters as descriptive features) and concrete examples to illustrate abstract ideas. The explanation of attention mechanisms, including query, key, and value, is particularly effective. The author also acknowledges the limitations of his simplifications and points to the original paper for deeper understanding. Overall, the argumentation is solid and persuasive, effectively conveying the ‘magic’ of LLMs while maintaining scientific accuracy.

Scientific Rigor, Source Quality, Title Accuracy

The video demonstrates a good level of scientific rigor, primarily relying on the seminal paper ‘Attention Is All You Need’ (arXiv:1706.03762), which is correctly cited in the description. The author’s explanations align with the concepts presented in that paper, and he accurately describes the architecture of transformer models. The title accurately reflects the content, as the video indeed presents a comprehensive study of LLMs. The author’s background in knowledge management and his six-month study lend credibility to his explanations. While the video is not a formal academic source, it serves as an excellent educational resource. The description also includes a link to the paper, which is a positive sign for source transparency.

229 words

Title / Content Match

The title accurately reflects the content: the author shares his six-month study of LLMs, explaining how ChatGPT works in a comprehensive, accessible manner.

Quality & Reliability

7/10

The video provides a detailed, step-by-step explanation of LLM internals, grounded in the original 'Attention Is All You Need' paper. The author demonstrates a solid understanding of the concepts, though some simplifications and analogies may lack precision. The content is educational and generally accurate, but not peer-reviewed.

Chapters

Cited Sources

Attention Is All You Need — The foundational paper introducing the Transformer architecture, cited as the basis for the video's explanation.

Concurring Sources

Attention Is All You Need — The video's explanations are consistent with the concepts and architecture described in this paper.

Contribution & Novelties

The video’s original contribution lies in its pedagogical approach: it translates complex mathematical concepts into intuitive, almost physical analogies, making the internal workings of LLMs accessible to a broader audience. It bridges the gap between oversimplified explanations and dense academic papers, providing a middle ground that is both informative and engaging.

Pour aller plus loin :

Transformer (machine learning model) — A comprehensive overview of the Transformer architecture, including attention mechanisms.
Word embedding — Explains the concept of embedding words into vector spaces, which is central to the video’s explanation.
Attention Is All You Need (arXiv) — The original paper for deeper technical details.

104 words

Radar Profile

The radar profile shows a balanced performance across all dimensions, with slightly higher scores in information quantity and technical level, reflecting the video's comprehensive and detailed nature. The quality and reliability scores are also strong, indicating a trustworthy educational resource.

Reliability 7/10