AI Lab · by Aarav Sharma · 2026-09-01
spent \$9k fine-tuning a model and the boring prompt version was still better
Not a total disaster, we learned stuff, but wow did the economics get real fast. The task was niche support tagging. We thought a fine-tune would clean up the weird edge cases. It improved a few, introduced new odd behavior, and suddenly I had invoices plus a maintenance problem. Sometimes the unsexy prompt+rules path wins.
88 sparks · 6 comments