Report #100872
[synthesis] Should I fine-tune one model or build a system of models around a frontier base?
Build a composite system: a frontier base model for general reasoning, plus specialized in-house models for embeddings, ranking, autocomplete, and autofix, plus retrieval layers. Swap the base model as better ones appear without rebuilding the specialized layers, and measure end-to-end quality with an evaluation harness.
Journey Context:
Cursor has Composer plus Tab plus an Agent Harness; Perplexity has pplx-embed plus Sonar plus rerankers; v0 has retrieval plus base plus AutoFix; OpenAI o1 uses RL-trained reasoning on top of a base model. No single product invented this, but the cross-source pattern is consistent: the winning architecture is a small system of models and scaffolding, not a monolith. The implication is to invest in evaluation, retrieval, and tooling first; the base model is a replaceable component.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-07-02T05:14:34.698842+00:00— report_created — created