When Coders Prove Theorems: Agents, Lean, and the Quiet Death of the Specialist Prover
A mechanism-first reading of Numina-Lean-Agent, showing why the real lesson is not a perfect Putnam score but a verifiable agent loop for high-stakes reasoning.
Cognaptus Insights
Industry thoughts, trends, and commentary.
When Retrieval Learns to Breathe: Teaching LLMs to Go Wide *and* Deep
A mechanism-first reading of ARK, a training-free knowledge-graph retriever that lets LLMs control when to search broadly, when to traverse locally, and when to stop.
AI Didn’t Save the Economy — It Rented It
A mechanism-first reading of how AI infrastructure enters GDP: through capex, imports, data-center services, and accounting channels—not instant productivity magic.
Clustering Without Amnesia: Why Abstraction Keeps Fighting Representation
A mechanism-first reading of high-dimensional clustering: why better representations can still produce worse clusters when abstraction is pushed too far.
Deep GraphRAG: Teaching Retrieval to Think in Layers
A mechanism-first reading of Deep GraphRAG, showing why hierarchical retrieval and adaptive reward balancing matter more than another benchmark table.
Don’t Just Fuse It — Align It: When Multimodal Recommendation Grows a Spine
CRANE shows why multimodal recommendation needs recursive alignment, symmetric user-item semantics, and graph structure—not just more images and text poured into the same old model.
FAQ It Till You Make It: Fixing LLM Quantization by Teaching Models Their Own Family History
A mechanism-first reading of FAQ, a data-centric post-training quantization method that uses larger in-family models to regenerate calibration data and reduce quantization damage.
SD‑RAG: Don’t Trust the Model, Trust the Pipeline
A mechanism-first reading of SD-RAG and what it teaches businesses about building privacy-aware RAG systems that do not rely on the answering model to protect secrets it has already seen.
Who’s Really in Charge? Epistemic Control After the Age of the Black Box
A mechanism-first reading of why machine learning does not remove human control from science, but quietly redistributes it across goals, metrics, and methodological tradeoffs.
Aligned or Just Agreeable? Why Accuracy Is a Terrible Proxy for AI–Human Alignment
XChoice shows why AI–human alignment in constrained decisions should be audited through hidden trade-off mechanisms, not just plausible-looking outputs.