LessWrong (Curated & Popular)

“Tracing the Thoughts of a Large Language Model” by Adam Jermyn

Mar 28, 2025

Adam Jermyn, author and AI enthusiast, dives deep into the fascinating realm of large language models like Claude. He uncovers how these models train themselves and develop unique problem-solving strategies. The discussion covers Claude's multilingual capabilities and how it constructs poetry with thoughtful rhymes. Jermyn also addresses its impressive reasoning and mental math skills, revealing the complexities behind its outputs. Lastly, he tackles issues like AI hallucinations and jailbreaking, highlighting the importance of understanding AI behavior.

22:18

Episode guests

Adam Jermyn

AI Summary

AI Chapters

Episode notes

Podcast summary created with Snipd AI

Quick takeaways

Claude's multilingual capabilities suggest a shared conceptual framework rather than language-specific processes, enhancing its understanding across languages.

The podcast highlights Claude's advanced planning in poetry generation, indicating its ability to think strategically rather than reactively.

Deep dives

Understanding Claude's Multilingual Capabilities

Claude's ability to communicate in multiple languages is explored, revealing that it may not operate with separate language-specific mechanisms for each language. Instead, there appears to be a shared conceptual space that allows Claude to understand and translate concepts across different languages. This is demonstrated through experiments where Claude processes and translates similar concepts, indicating that it can activate the same core features regardless of the language. This insight into Claude's multilingual processing suggests that it can leverage knowledge acquired in one language to apply that understanding when communicating in another, showcasing a sophisticated level of reasoning.

Intro

2min

Exploring the Inner Workings of a Multilingual AI Model

4min

Exploring Multilingual Conceptual Universality and Rhyming Mechanics

2min

Unpacking Claude's Reasoning and Math Skills

10min

Exploring AI Hallucinations and Jailbreak Strategies

4min

[This is our blog post on the papers, which can be found at https://transformer-circuits.pub/2025/attribution-graphs/biology.html and https://transformer-circuits.pub/2025/attribution-graphs/methods.html.]

Language models like Claude aren't programmed directly by humans—instead, they‘re trained on large amounts of data. During that training process, they learn their own strategies to solve problems. These strategies are encoded in the billions of computations a model performs for every word it writes. They arrive inscrutable to us, the model's developers. This means that we don’t understand how models do most of the things they do.

Knowing how models like Claude think would allow us to have a better understanding of their abilities, as well as help us ensure that they’re doing what we intend them to. For example:

Claude can speak dozens of languages. What language, if any, is it using "in its head"?
Claude writes text one word at a time. Is it only focusing on predicting the [...]

---

Outline:

(06:02) How is Claude multilingual?

(07:43) Does Claude plan its rhymes?

(09:58) Mental Math

(12:04) Are Claude's explanations always faithful?

(15:27) Multi-step Reasoning

(17:09) Hallucinations

(19:36) Jailbreaks

---

First published:
March 27th, 2025

Source:
https://www.lesswrong.com/posts/zsr4rWRASxwmgXfmq/tracing-the-thoughts-of-a-large-language-model

---

Narrated by TYPE III AUDIO.

---

Images from the article: