The experiment maps how Qwen3-1.7B represents conversations in its 2,048-dimensional residual stream, the shared workspace its layers update as they process tokens. It records final token states from 900 question-and-answer examples spanning 100 topics in English, Chinese, German, and Croatian, plus 90 English-prompted Python examples. The states are projected into three dimensions so token sequences can be seen as paths. A comparison with the model’s 151,936 token embeddings provides a view of its learned starting representations. The maps show English, German, and Croatian occupying overlapping neighborhoods, while Chinese appears more offset; code lies in the same broad region as the natural-language examples. These patterns suggest that representations reflect both subject matter and the language or form used to express it.
The visual patterns depend strongly on how the map is constructed. A global PCA dominated by vocabulary embeddings captures 42.4% of variation among conversation averages in its first three directions; fitting PCA to conversation averages instead captures 80.1% and makes some language differences clearer. The apparent thinness of the paths is also a projection effect: a plane explains 99.7% of variation visible in the three-dimensional view, but those axes capture only 17.6% of variation across the original token states. Language tracks the geometry strongly, while topic structure remains after accounting for language. Because language, script, tokenization, wording, and answer length vary together, the maps cannot identify which factor causes each separation, nor do they reveal a pure representation of meaning.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.