Stop fine-tuning to fix retrieval problems: Oracle technologist on where knowledge should live
AI Engineer · youtube · 2026-10-04
Anant Srivastava, principal technologist for data and AI platforms at Oracle, argued at AI Engineer World's Fair 2026 that prompt, memory and weights aren't a ladder you climb when answers go wrong — they're three tools for three jobs.
- Most teams never decide where knowledge lives; six months of normal product work decides for them. Every prompt edit, indexed doc and training sample is an architecture decision.
- A support assistant kept inventing product names after a fine-tune — a case of facts being baked into weights instead of living in memory.
- His diagnostics: prompts for small, stable behavior; memory for knowledge that's current, large, citable or access-controlled (access control belongs in memory); weights only for reflexes that have stopped changing.
- Fine-tune reflexes, not facts — e.g., stable medical-coding tasks.
- He closes with a circulating architecture where the agent improves by doing its job, noting "the model is the easy part."
More from coding & agent
- Dev builds Browser MCP Defense demo with ampcode, seeks design partners — HankYeomans · 2026-10-04
- Follow-up: experimenting with custom fine-tuned models per customer — sorenrood · 2026-10-04
- GitHub Copilot agent maps animation paths across scene trees, live-screenshots progress — DanWahlin · 2026-10-04
- Computer-use agent API claims 90% token savings, tops OSWorld 2.0 at 73% for $15.70 per task — AdFluid9823 · 2026-10-04
- Multi-hour agent sessions need KV cache capacity no major AI provider offers yet — AccBalanced · 2026-10-04
- Agent-first is the way: what matters is 'Agent SEO', not your UI — markjeffrey · 2026-10-04