STOP UPGRADING THE MODEL: ADD RELATIONSHIPS AND STATE INSTEAD
Two new posts argue that explicit relationships and state often beat bigger models for recommendations and AI companions. A Hurix piece makes the case that kno...
Two new posts argue that explicit relationships and state often beat bigger models for recommendations and AI companions.
A Hurix piece makes the case that knowledge graphs outperform vector-only search for learning recommendations by encoding prerequisites and concept links, not just proximity in embedding space Hurix. That shift turns a pile of content into a map of concepts, which changes how paths, gaps, and remediation are handled.
A dev.to writeup shows a cheaper model with a product-layer relationship architecture behaved more like a person than a stronger model without state or boundaries DEV. The signal: model strength helps, but structure, memory, and pacing rules often matter more.
Upgrading architecture (graphs, state, boundaries) may yield bigger gains than upgrading the model tier.
Better structure reduces drift, backtracking, and unsafe responses without expensive fine-tuning.
-
terminal
A/B: vector-only recommender vs. a small knowledge graph of skills and prerequisites; measure backtracks, completion rate, and dwell time.
-
terminal
Companion agent: add a finite-state relationship model with boundaries and cooldowns; compare refusal rates and session quality vs. a stronger stateless model.
Legacy codebase integration strategies...
- 01.
Add a sidecar graph over existing content and events; start by extracting entities, prerequisites, and outcome edges from logs.
- 02.
Introduce a lightweight state store for agents (relationship stage, boundaries, last escalations) before touching model configs.
Fresh architecture paradigms...
- 01.
Design a domain graph first-class: entities, prerequisite edges, and progression policies alongside embeddings.
- 02.
Separate long-term user state from chat context; route policies through a rules layer before LLM calls.
Get daily CONVERSATIONAL-AGENTS + SDLC updates.
- Practical tactics you can ship tomorrow
- Tooling, workflows, and architecture notes
- One short email each weekday