📋 Key Takeaways
- Central thesis: the LLM is a stateless one-shot function; the Agent is a script printer; building an agent = writing a script
- Token economics: input cheap, output 4× expensive; context window = paper tray; if you don't write the script, you're a character in someone else's
- Little Wukong = LLM + your 10 years of specialty; beat ChatGPT on three axes: ① narrow ② deep ③ real
- Hallucinations are a script problem, not a model problem; iron laws must be explicit or right answers get overruled
🎬 Video Replay
YouTube: https://youtu.be/9tt8RYoChKM
🔧 Materials
✏️ Tarea
HW1 (warm-up · no submission): math out monthly fee vs real token burn. See which of the three AI-money tiers you're on. HW2 (submit · Golden Hoop 3 · Freedom): write your Little Wukong script (5-part System Prompt + run 1 real dialog + brief) — this is the agent blueprint you'll build in S4.
⏰ Fecha límite: 5/7 周三 9:00 PM PT
→ Enviar tarea (S3-FREEDOM)▶ 🤖 Criterios de evaluación IA
| Dimensión | Peso | Descripción |
|---|---|---|
| Specialty clarity | 20% | Whether the narrow/deep/real axes are sharp enough that a stranger can immediately tell "ChatGPT cannot replace this" |
| Script completeness | 25% | Whether all 5 parts of the System Prompt are present and specific |
| Iron laws actionable | 15% | At least 3 concrete "always do X / never do Y" rules the AI can execute; not platitudes or marketing copy |
| Dialog validation | 15% | A real dialog was actually run; output demonstrably exceeds ChatGPT default persona; annotation calls out which of narrow/deep/real shows up |
| S4 buildability | 15% | The agent described in the script is actually buildable in S4 — inputs, processing, outputs are all clear |
| Next-step specificity | 10% | One concrete validation action this week; not "keep learning" |
90-100 Highly thoughtful; script is S4-build-ready; specialty obvious to a stranger · 70-89 Completed seriously; script is structurally complete; specialty is visible · 50-69 Completed but partially perfunctory — iron laws are platitudes or specialty unclear · 30-49 Very perfunctory or mostly AI-generated cliches · 1-29 Nearly empty or entirely ChatGPT one-click output