TL;DREvaluating LLMs is hard — benchmarks give numbers, but "vibes" and task-specific tests matter too.
Unlike a classifier with one accuracy number, LLMs do open-ended tasks where "good" is fuzzy. Standardized benchmarks give comparable scores but get gamed and saturate; human preference ("vibes") is subjective; the real measure is performance on your task. Best practice: build a small, representative eval set for your use case and track it as you change prompts/models.
Key points
Open-ended output makes "good" fuzzy
Benchmarks: comparable but gameable/saturating
Human preference is subjective
Build a task-specific eval set
Common mistakes
Trusting a leaderboard number for your task
No eval set → flying blind on changes
Optimizing for benchmarks, not real use
Try it: Explain why a high benchmark score may not predict success on your task.
Example code
<!doctype html><html><head><meta charset="utf-8"></head>
<body style="background:#06040d;color:#e6e0ff;font-family:monospace;padding:20px"><pre>leaderboard ≠ your task
→ build a small eval set of YOUR cases
→ track it across prompt/model changes</pre></body></html>