← Stream
field notes

Thematic analysis of the garden: graph-augmented

What happens when you start from embeddings instead of reading? Comparing cluster-based discovery against the fresh thematic analysis.

Part of experiment-2-thematic-analysis-with-1m-context.

Method: Loaded all 213 non-draft items’ bge-m3 embeddings (1024 dimensions), ran k-means clustering (k=10), then named clusters by examining their members. No raw text reading; structure comes entirely from the embedding space.

The 10 clusters

ClusterSizeLabelCharacter
128AI and philosophy (library)Format-driven: 22 of 28 are library entries
210Writing, language, content strategyEclectic, weakly cohesive
312The ChatGPT exploration seriesTopic-focused: all articles about ChatGPT encounters
437Fiction, personal reading, miscellaneousJunk drawer: items that don’t fit elsewhere
515Building the garden (meta)Clear: garden infrastructure, design system, meta-content
614Knowledge graph researchTight: embeddings, Saga papers, key phrases, thresholds
722Videos and tutorialsFormat-driven: 15 of 22 are videos regardless of topic
837Design, learning, promptingBroad: design thinking, prompt design, instructional design
93OutliersNoise: three items with no clear shared theme
1035Conversation design professionVery clear: career, community, tools, CxD practice

What the embeddings confirm

Three of the fresh reading themes map almost directly to embedding clusters:

  • Theme 7 (knowledge graph as personal research program) = Cluster 6. The tightest match: the same 14 items form a cohesive group in both analyses.
  • Theme 3 (community as infrastructure) lives inside Cluster 10, which captures the broader conversation design profession. The community articles are there, but embedded within a larger professional context.
  • Theme 5 (design is not engineering) maps to Cluster 8. The design-focused articles and books cluster together, though the cluster is broader than the theme.

What the embeddings miss

The non-linear mind

The emerging theme from Approach A — Maaike as a non-linear thinker, the garden as a medium for a non-linear head — is completely invisible to embeddings. The evidence is scattered across clusters 4 (book recommender), 5 (garden building), and 10 (freelance journey). Embeddings can’t see this pattern because it’s expressed in personal anecdotes and autobiographical asides, not topical keywords.

The bilingual practitioner

Dutch content is sprinkled across clusters: DALL-E debiasing in Cluster 8, SSML tutorials in Cluster 7, Dutch books in Cluster 4. But the fresh reading’s insight — that being non-English-default in an English-default AI world is a critique position that reveals hidden biases — is invisible. Embeddings see language similarity; they don’t see epistemological stance.

Truth as a coherent thread

The fresh reading identified a single theme running from Frankfurt’s bullshit to Grice’s maxims to fake hyperlinks to the designing-for-doubt concept. The embeddings split this across three clusters: Cluster 3 (ChatGPT articles), Cluster 8 (DALL-E/prompting), and Cluster 1 (philosophy books). Each cluster sees the surface topic; none sees the underlying concern about truth and trust in AI interfaces.

The voice of the maker

Authorship and authenticity as a theme is scattered across Cluster 1 (AI philosophy books), Cluster 2 (writing), and Cluster 5 (the garden’s AI transparency system). The connection between Nick Cave’s critique, the “100% Maai” label, and the friction of accordion editing requires reading, not vectors.

What the embeddings see that reading didn’t emphasize

Format contamination (Cluster 7)

15 of 22 items in Cluster 7 are videos. The embeddings cluster by content type, not topic. A video about designing for doubt and an article about the same concept end up in different clusters because video descriptions are short and share stylistic patterns. This confirms the known chunk size problem: embeddings encode format alongside meaning.

The junk drawer (Cluster 4)

37 items that don’t belong anywhere else: fiction, personal books, miscellaneous videos, even one infrastructure field note (typed relations). When embeddings can’t find a strong topic signal, items fall into a catch-all cluster. This is where the model gives up.

Conclusion

The fear was right: starting from embeddings constrains discovery. The clusters confirm existing topical structure (knowledge graph research, conversation design profession, garden building) but miss cross-cutting themes that require understanding stance, autobiography, and argument structure rather than topic similarity.

The fresh reading found themes about truth, authenticity, non-linearity, and bilingual critique — none of which live in a single topical cluster. These are the themes that only emerge from holding the entire garden in context simultaneously.

That said, the embeddings are not useless for thematic analysis. They are excellent at confirming obvious structure and identifying format contamination. A combined approach might work best: embeddings to map the topical landscape, then full-context reading to find the threads that run across it.

Mycelium tags, relations & arguments