Decision model vs. LLM on real app: 10x cheaper but 10x more code complexity
slakmehl · reddit · 2026-10-03
The author added a natural language interface to their Europe trip planner tripsnek and compared a conventional LLM intent extractor against the "decision model" (jev):
- Yes/no and pick-from-list parameters worked extremely well out of the box (travel mode, pace, interests).
- Picking from very large sets (specific cities, sights, date ranges) was much harder, requiring complex regexes and text munging.
- The resulting pipeline matched or exceeded a well-prompted LLM at roughly an order of magnitude lower cost—but with an order of magnitude more code complexity.
- The author actually prefers debuggable, improvable code over an inscrutable prompt black box.
Try it at tripsnek.com/describe with a checkbox to drill down into every question and answer; full write-up at tripsnek.com/lab/jev-trip-intent.
More from Apps
- Synthesia's first use case for conversational avatars: sales roleplay training — mattturck · 2026-10-03
- Sam Altman calls dot his favorite OpenAI product yet, praising daily improvement — sama · 2026-10-03
- Apple Tightens macOS Full Disk Access, Citing New Risks From AI Agents — TechCrunch AI · 2026-10-03
- Maket 2.0 launches: AI home design platform turns prompts into floor plans and 3D views — testingcatalog · 2026-10-03
- ChatGPT users annoyed by pushy Work mode prompts: 'never ask me this again' — tsunami_forever · 2026-10-03
- Enterprise user's verdict on Copilot: Microsoft had the best agent surfaces and just pasted chatbots everywhere — YoungDito · 2026-10-03