Simon Willison's coverage of this technique underscores a broader, second-order trend in applied AI: the shift from deterministic to probabilistic system design. The core insight—using embeddings to bridge generative output and structured data—is a pragmatic workaround for the brittleness of direct classification prompts. However, in our view, the method's success likely hinges on the quality and coverage of the embedding corpus; a sparse or biased tag set could render the 'closest match' meaningless.
It represents an interesting middle ground between fully open-ended generation and rigid taxonomy, but its real-world utility may be limited to domains where semantic closeness reliably indicates categorical correctness, which is far from universal.