You can’t compete with infinite compute. But you can find adjacent spaces where depth matters more than scale.
The Strategic Insight
I couldn’t compete with OpenAI, Anthropic, or Google on LLM capabilities. They have compute, talent, and capital I’ll never approach. The space is crowded, competitive, and moving at pace.
So I pivoted to something related but fundamentally different: not building better models, but better infrastructure for understanding what these models help us discover about ourselves.
Instead of asking “how do we make LLMs smarter?”, I asked: “what do our conversations with LLMs reveal about how we think, and how can we make that structure queryable, browseable, and conversable?”
This turned out to be a far less saturated space with deep intellectual problems and practical applications. And it maps naturally to a research program that can sustain itself over years, spawn diverse projects, and actually produce something useful.
The lesson: when you can’t compete on the main axis, find the orthogonal space where your particular skills and constraints become advantages.
The Cognitive MRI
I started by analyzing my own AI conversation logs—years of chats with ChatGPT, Claude, and other systems. Thousands of conversations spanning code, research, philosophy, health, projects.
Linear text hides structure. But when you construct semantic similarity networks from these conversations and analyze their topology, something fascinating emerges.
Method
- Embed conversations using standard language models
- Weight user inputs 2x more than AI responses (ablation studies showed 2:1 user:AI weighting maximizes modularity)
- Construct similarity graph with cosine similarity edge weights
- Apply threshold cutoff to keep meaningful connections (phase transition appears around θ ≈ 0.9)
- Identify communities through network clustering
The result: a map of how your knowledge exploration actually structures itself.
Key Finding: Heterogeneous Topology
Different kinds of conversations have fundamentally different network structures:
Programming and practical work: Tree-like hierarchies
- Linear problem-solving paths
- Branching from problem to solution
- Few cross-domain connections
- High average path length between any two conversations
Research and conceptual work: Small-world networks
- Hub-and-spoke structure
- Central concepts with many connections
- Bridge nodes linking distant domains
- Short paths between seemingly unrelated ideas
Intermediate domains: Hybrid structures between these extremes
This wasn’t predicted. The topology reveals the cognitive mode of each domain.
Bridge Nodes and the Giant Component
A few key conversations act as bridge nodes that hold the entire knowledge graph together. These aren’t necessarily the “most important” conversations by any subjective measure—they’re the ones that connect otherwise separate communities.
Remove these bridges, and your knowledge graph fragments into isolated clusters. They represent the conceptual linchpins of how you integrate different domains.
The Ablation Studies
We varied two parameters systematically:
User:AI weighting ratio: From 1:1 to 3:1, with 2:1 producing highest modularity Threshold cutoff θ: From 0.7 to 0.95, with phase transition around 0.9
This wasn’t arbitrary tuning—it revealed that there’s a principled way to extract community structure from conversational data. The phase transition suggests we’re hitting real structure, not just parameter artifacts.
What This Reveals
Complex networks science has spent decades proving one insight: topology reveals what reductionist approaches miss.
We’ve applied this to social networks, biological systems, infrastructure, citation graphs. Now we have a new data source that’s exploding in volume and richness: our conversations with AI systems.
These conversations are networks of thought—semantic connections, conceptual bridges, knowledge communities. And most tools don’t leverage this structure at all.
Current RAG Systems Miss the Graph
Typical retrieval-augmented generation:
- Convert query to embeddings
- Find nearest neighbors in vector space
- Return similar documents
- Done
This is nearest-neighbor search in a metric space. It doesn’t know:
- Which documents are strongly connected
- Where bridge nodes link distant communities
- Which documents act as hubs
- How knowledge clusters actually organize
We’re throwing away the graph structure.
The Vision: Queryable, Browseable, Conversable
What if tools actually understood their own network structure?
Queryable
Not just “find documents similar to this query” but:
- “What are the bridge concepts between these two topics?”
- “Which documents act as hubs in this domain?”
- “Show me the path connecting these distant ideas”
- “What communities exist in my knowledge base?”
The query language becomes richer because you’re navigating graph structure, not just searching vector space.
Browseable
Right now, browsing knowledge is either chaotic (no structure) or artificial (imposed hierarchies that don’t reflect actual relationships).
But if you surface the actual network topology:
- Show natural clusters and communities
- Highlight hub documents
- Reveal bridge concepts
- Display the actual semantic connectivity
Navigation mirrors how knowledge actually connects, not how you thought it should be organized when you filed it.
Conversable
An LLM responding to queries about your knowledge base could reason about topology:
“These three documents form a tight cluster because they explore X from different angles. This other document bridges to cluster Y through principle Z. Here’s how they all connect structurally.”
It’s not just summarizing content—it’s reasoning about the graph.
The Infrastructure: Complex-Net RAG
I’m building a Python package that generalizes these insights into a domain-specific language for network-augmented retrieval.
The DSL captures:
- Graph construction from arbitrary data sources (conversations, ebooks, bookmarks, emails, documents)
- Community detection with tunable parameters
- Bridge identification and path analysis
- Hub detection and centrality measures
- Conversational interface layered over graph structure
- Orchestration across multiple data sources via MCP
Apply it to:
- Your AI conversation history
- Your ebook collection
- Your browser bookmarks
- Your email archives
- Your personal documents
- Eventually: public data sources for reproducibility
Everything becomes part of one unified knowledge graph where the topology reveals structure you didn’t know existed.
Why Complex Networks Researchers Should Care
For decades, network science has analyzed structure in static datasets: social graphs from 2015, protein interactions, transportation networks.
Now we have:
A new data source: LLM conversations generating networks of thought at scale, continuously
A new capability: LLMs can interpret what network structure means in ways pure algorithms cannot
A new scope: We can apply network analysis across interconnected data sources simultaneously through unified interfaces
This isn’t a departure from complex networks thinking. It’s the fulfillment of it.
We’re finally building systems that understand networks the way network science understands them.
The Meta-Point: Research Strategy
This research direction emerged from strategic necessity:
- Identified a crowded space (LLM capabilities) where I couldn’t compete
- Found adjacent territory (understanding conversation structure) that was less saturated
- Brought existing expertise (one networking class + mathematical maturity) to a new domain
- Produced publishable results quickly (Complex Networks 2025, December in New York)
- Opened a research program that can support multiple theses and extensions
This is strategic research positioning: find where your constraints become advantages.
Why This Space Works
Low barrier to entry: You don’t need to be a networks expert from day one. You need intellectual curiosity and willingness to learn tools as needed.
High generativity: Once you have the infrastructure, endless research questions emerge:
- Methodological improvements
- Application to new domains
- Systems and architecture papers
- Cognitive science angles
- Tool development
Practical usefulness: It solves a real problem (making personal knowledge actually useful) that will only grow more pressing as we accumulate more AI-mediated knowledge.
Sustainable: Not chasing trends—building foundational infrastructure that supports years of work.
The Pitch to Network Scientists
We have all this data—mountains of LLM conversations, accumulated knowledge in diverse formats, interconnected information sources.
We want to make it queryable, browseable, and conversable.
All three benefit from complex networks, but also from the attitude complex networks brings: topology reveals what reductionism misses.
You’ve been collecting data and analyzing it through network lenses for decades. Now we have:
A new and growing data source (AI-mediated knowledge exploration)
An opportunity to layer conversational components (we can talk to our multimodal data through LLMs)
Much smarter self-organizing, searchable, browseable infrastructure (with LLMs and network analysis working together)
Orchestration across many sources (through protocols like MCP)
The historical insight of complex networks—structure matters—applies directly to how we should build knowledge tools.
One Networking Class, One Publication
I took one networking class. It resulted in a peer-reviewed publication and a talk at Complex Networks 2025.
This isn’t about being brilliant. It’s about:
- Asking the right questions in a less-saturated space
- Bringing complementary skills (programming, statistics, mathematical thinking)
- Working strategically rather than competing on compute
- Building infrastructure that supports long-term programs
The activation energy for productive research in this space is lower than people think.
The bottleneck isn’t prerequisite knowledge—it’s intellectual curiosity and problem-solving ability.
The Larger Context
This work exists in a particular context I should acknowledge: I have stage 4 cancer, uncertain time horizons, and recurring treatment cycles.
Cancer doesn’t change the intellectual work—it clarifies what’s worth doing.
Strategic positioning matters more when time is uncertain. Finding sustainable research directions matters more. Building infrastructure that others can continue matters more.
The cognitive MRI project documents my own knowledge exploration during compressed timelines. The networks map real intellectual work under real constraints.
It’s fitting: network analysis of thought, while time runs out.
What Makes This Different
Most AI research focuses on model capabilities: bigger, faster, smarter.
This focuses on infrastructure for understanding: what do conversations reveal, how do we make that structure useful, how do we build tools that leverage topology rather than ignoring it.
It’s the difference between:
- Building better search engines vs. understanding how knowledge organizes
- Optimizing retrieval vs. revealing structure
- Responding to queries vs. reasoning about topology
- Scaling compute vs. scaling comprehension
Both are valuable. The second is vastly less crowded.
For Graduate Students
If you’re looking for research directions:
Don’t compete on the main axis everyone else is competing on. Find the orthogonal space where your particular constraints become advantages.
Look for problems that are:
- Adjacent to hyped areas but less saturated
- Solvable with your existing skills + learnable tools
- Generative of multiple follow-up questions
- Practically useful beyond academic novelty
- Sustainable over multi-year timelines
Personal knowledge graphs + complex networks + LLMs is one such space. There are others.
The key is strategic positioning: where can you actually contribute something novel without infinite compute or decades of specialization?
The Next Steps
The complex-net RAG package will:
- Provide a DSL for graph-augmented retrieval
- Work across diverse data types
- Enable queryable/browseable/conversable interfaces
- Orchestrate multiple sources through MCP
- Support both personal and public data for reproducibility
The research program will explore:
- Methodological refinements (weighting schemes, threshold selection)
- Applications to new domains (emails, documents, multimodal data)
- Cognitive science questions (what do these structures reveal about thinking?)
- Systems questions (how to scale, how to integrate)
- Tool development (making this usable by others)
And most importantly: it creates a training ground where students can enter the space quickly, contribute meaningfully, and build their own research directions.
The Bottom Line
Strategic insight: Can’t compete on compute? Find adjacent spaces where depth matters more than scale.
Technical finding: Conversation networks have heterogeneous topology that reveals cognitive structure.
Vision: Build infrastructure that makes knowledge queryable, browseable, and conversable through network-aware tools.
Broader point: The next frontier of complex networks isn’t just analyzing networks in data—it’s building tools that understand the network structure they operate on.
This is work for decades. It’s generative, practical, and intellectually rich.
And it’s a reminder: when the main path is crowded, look for the orthogonal space where your constraints become advantages.
Related Work
- Cognitive MRI: Accepted to Complex Networks 2025 — The conference paper
- Mathematical Elegance in Software Design — On building with structure
- Why I Build Comprehensively in Open Source — Infrastructure that outlasts you
Complex Networks 2025, New York, December. Revealing thought topology hidden in conversation.
Discussion