What thousands of conversations look like inside a language model
TL;DR Highlight
Mapping Qwen3-1.7B states showed Chinese sitting apart from English, while German and Croatian overlapped it.
Who Should Read
An LLM application developer who wants to understand multilingual model internals by visualizing hidden states and embeddings
Core Mechanics
- 900 English QAs looked like a thin sheet
- In the 3D projection, a best-fit plane held about 99.3% of the variation in English token states.
- In the full 2,048 dimensions, the displayed 3D axes captured only 17.6%, and the best 2D plane 28.7%.
- Chinese sat apart; German and Croatian overlapped English
- English, Chinese, German and Croatian (900 examples each) were plotted in one fixed PCA frame.
- German and Croatian overlapped heavily with English, while Chinese was visibly offset.
- The author calls these overlapping regions with different centers and shapes, not separate language compartments.
- The PCA axes changed the picture
- Global PCA used 151,936 vocabulary embeddings plus 3,690 conversation averages, with vocabulary about 98% of fitting points.
- Conversation PCA captured 80.1% of conversation-average variation versus 42.4% for Global PCA.
- German and Croatian became more distinguishable.
- Code was near English but not nearest to it
- 90 Python examples appeared in the same broad neighborhood as English and other Latin-script data.
- By average distance, code was closer to Croatian and German than to English.
- Each coding example included an English question, so several properties were entangled.
- Bigger patches are not more knowledge
- Croatian answers had many more Qwen tokens than English, producing more points and lines.
- The author warns against reading a larger or denser patch as greater knowledge or capacity.
- Pinyin: closer on average, much worse at matching
- Converting Chinese to Pinyin cut average distance to English from 0.996 to 0.692.
- Finding the matching English case dropped from 471/900 with characters to 28/900 with Pinyin.
- Closer average position did not mean better alignment of shared content.
Community Reactions
- Opinion that including vocabulary embeddings hurt the PCA
- Commenters said vocabulary embeddings are disjoint from conversation embeddings and flatten the conversations into a plane.
- They suggested removing them might reveal clearer language structure.
- Pushback that last-layer results are expected
- One commenter said Chinese tokens following Chinese and Latin tokens following Latin is expected.
- They wanted to see translations in the middle layers.
- Opinion that the methodology is thin or under-explained
- Commenters asked why 3D, where the dimensionality reduction happened, and whether the plane means anything.
- One suggested UMAP or t-SNE as alternatives.
- Experience: length dominated a Qwen embedding map
- A commenter mapping Steam reviews with Qwen embeddings and UMAP found review length affected position more than content.
- They said most of this year's work went into fixing that artifact.
How to Apply
- Check text-length effects in your embedding map first
- When visualizing embeddings with UMAP or PCA, color points by text length.
- If clusters follow length rather than topic, resample to equalize lengths.
- Fit PCA on explicit data and compare two variants
- Build a version whose axes are fit on conversation averages only, not vocabulary or all tokens.
- Record explained variance for both plots.
- Evaluate multilingual retrieval by pair matching, not mean distance
- Build same-meaning multilingual pairs and measure how often the true match is found among candidates.
- Use average distance only as a secondary metric.
Try It
# Paste into Claude or ChatGPT (simplified version of the article's idea: review an analysis design without running code)
I want to visualize a multilingual embedding map with PCA. Check the setup below for interpretation risks.
Data: {{languages and sample sizes}}
Embedding model: {{model name}}
Reduction method: {{PCA/UMAP etc. and its input data}}
Answer:
1. Ways text length, script, and token count could contaminate the results
2. A pair-matching evaluation design to use instead of mean distance
3. A comparison plan for which data to fit the PCA axes on
Next step: pick one variable from the plan to control and check it on real data first.Terminology
Related Papers
Claude was trained on protein sequences at scale
Next-amino-acid tests on obscure proteins show only the latest Claude models predict them
Remembered as Brief: Stored Preferences SilentlyRemove Decision-Critical Information
A stored 'keep it short' preference silently drops warnings; listing must-know points first restores content.
All AI-generated code should be Open Source
A blog argues AI-generated code should be open source; most HN commenters disagreed.
AI, activity, and outcomes in software engineering
AI sped up writing code, but shipping speed did not rise as much, says a talk
LLMs Play Zachtronics' TIS-100
16 LLMs played TIS-100 and set 133 new Pareto frontier points across 18 puzzles
AI usage generates a fragile codebase
Generating all code with AI left shallow tests, complex code, and less human understanding, says one dev.
Accepted but Wrong: Silent Failures of LLM Code Repairs after API Migrations
One release-note edge input cut wrong LLM repairs passing the gate from 0.410 to 0.065
Qwen 3.8 27B on my laptop can interpret model weights, too
Local Qwen 3.8 27B identified 3 of 4 games from raw network weights alone