TL;DR: A visualization of GPT-2 Small's token embeddings reveals that discretizing coordinates shifts 'Trump's nearest neighbors from specific figures to generic political names.
Summary: An analysis of GPT-2 Small's static embedding table compares discretized versus continuous nearest neighbors for the token 'Trump'. Discretization yields generic political terms like Mitt, Hillary, Pelosi, while continuous embeddings produce specific names like Obama, Clinton, Bush. No context or attention is involved; the finding comes directly from learned token embeddings.
Why it matters: This highlights how quantization or discretization of LLM embeddings can drastically alter semantic proximity, affecting retrieval quality in vector databases. AI builders should evaluate whether their embedding storage methods preserve fine-grained relationships for sensitive concepts.
Source: reddit