13 KiB
13 KiB
Autonomous System Comparison
V1 (Current) vs V2 (Proposed)
┌─────────────────────────────────────────────────────────────────────┐
│ V1 SYSTEM (Current) │
└─────────────────────────────────────────────────────────────────────┘
⏰ Timer (every 15 min)
│
├──> 🎲 Random roll (10% chance)
│ │
│ ├──> ❌ No action (90% of time)
│ │
│ └──> ✅ Take action
│ │
│ ├──> 🎲 Random pick: general/engage/tweet
│ │
│ └──> 🤖 Call LLM to generate content
│
└──> ⏰ Wait 15 min, repeat
Problems:
❌ No awareness of channel state
❌ Might speak to empty room
❌ Might interrupt active conversation
❌ Mood doesn't affect timing/frequency
❌ Wastes 90% of timer ticks
┌─────────────────────────────────────────────────────────────────────┐
│ V2 SYSTEM (Proposed) │
└─────────────────────────────────────────────────────────────────────┘
📨 Events (messages, presence, status)
│
├──> 📊 Update Context Signals (lightweight, no LLM)
│ │
│ ├─> Message count (5 min, 1 hour)
│ ├─> Conversation momentum
│ ├─> User presence changes
│ ├─> Time since last action
│ └─> Current mood profile
│
└──> 🧠 Decision Engine (simple math, no LLM)
│
├──> Check thresholds:
│ ├─> Conversation momentum > X?
│ ├─> Messages since appearance > Y?
│ ├─> Time since last action > Z?
│ ├─> Mood energy/sociability score?
│ └─> User events detected?
│
├──> ❌ No action (most of the time)
│
└──> ✅ Take action (when context is right)
│
├──> 🎯 Pick action based on context
│ ├─> High momentum → join conversation
│ ├─> User activity → engage user
│ ├─> FOMO triggered → general message
│ ├─> Long silence → break silence
│ └─> Quiet + curious → share tweet
│
└──> 🤖 Call LLM to generate content
Benefits:
✅ Context-aware decisions
✅ Mood influences behavior
✅ Responds to social cues
✅ No wasted cycles
✅ Zero LLM calls for decisions
┌─────────────────────────────────────────────────────────────────────┐
│ MOOD INFLUENCE EXAMPLE │
└─────────────────────────────────────────────────────────────────────┘
Bubbly Miku (energy: 0.9, sociability: 0.95, impulsiveness: 0.8)
┌─────────────────────────────────────────────────────────┐
│ Channel Activity Timeline │
├─────────────────────────────────────────────────────────┤
│ [5 messages] ────────> Miku joins! (low threshold) │
│ [quiet 20 min] ─────> "Anyone here? 🫧" │
└─────────────────────────────────────────────────────────┘
Shy Miku (energy: 0.4, sociability: 0.2, impulsiveness: 0.2)
┌─────────────────────────────────────────────────────────┐
│ Channel Activity Timeline │
├─────────────────────────────────────────────────────────┤
│ [5 messages] ────────> ... (waits) │
│ [15 messages] ───────> ... (still hesitant) │
│ [40 messages] ───────> "Um... hi 👉👈" (finally joins) │
│ [quiet 2 hours] ─────> ... (doesn't break silence) │
└─────────────────────────────────────────────────────────┘
┌─────────────────────────────────────────────────────────────────────┐
│ RESOURCE USAGE COMPARISON │
└─────────────────────────────────────────────────────────────────────┘
V1 System (per hour):
┌──────────────────────────────────────────────────┐
│ Timer checks: 4 (every 15 min) │
│ Actions taken: ~0.4 (10% of 4) │
│ LLM calls: ~0.4 (only when action taken) │
│ Wasted cycles: 3.6 (90% of time) │
│ Context awareness: 0 🚫 │
└──────────────────────────────────────────────────┘
V2 System (per hour, typical server):
┌──────────────────────────────────────────────────┐
│ Message events: ~50 (passive tracking) │
│ Presence events: ~10 (passive tracking) │
│ Decision checks: ~60 (lightweight math) │
│ Actions taken: ~0.5-2 (context-dependent) │
│ LLM calls: ~0.5-2 (only when action taken) │
│ Wasted cycles: 0 ✅ │
│ Context awareness: Real-time 🎯 │
└──────────────────────────────────────────────────┘
Key Difference:
V1: Blind random chance, no context
V2: Smart decisions, full context, same LLM usage
┌─────────────────────────────────────────────────────────────────────┐
│ DECISION FLOW EXAMPLE │
└─────────────────────────────────────────────────────────────────────┘
Scenario: Active gaming chat, Miku is "excited" mood
1. Message arrives: "Just beat that boss!"
└─> Engine: track_message() → momentum = 0.7
2. Check decision:
┌────────────────────────────────────────────┐
│ conversation_momentum = 0.7 │
│ threshold (excited) = 0.6 * (2-0.9) = 0.66 │
│ 0.7 > 0.66 ✅ │
│ │
│ messages_since_appearance = 8 │
│ 8 >= 5 ✅ │
│ │
│ time_since_last_action = 450s │
│ 450 > 300 ✅ │
│ │
│ random() < impulsiveness (0.9) │
│ 0.43 < 0.9 ✅ │
│ │
│ DECISION: join_conversation ✅ │
└────────────────────────────────────────────┘
3. Execute action:
└─> Call existing miku_detect_and_join_conversation_for_server()
└─> LLM generates contextual response
└─> "Wahaha! That boss was tough! What did you think of the music? 🎵✨"
4. Record action:
└─> Reset messages_since_appearance = 0
└─> Update time_since_last_action
┌─────────────────────────────────────────────────────────────────────┐
│ MIGRATION PATH │
└─────────────────────────────────────────────────────────────────────┘
Phase 1: Install V2 (parallel)
┌──────────────────────────────────────────────┐
│ Keep V1 scheduler running │
│ Add V2 event hooks │
│ V2 tracks context but doesn't act │
│ Monitor logs to verify tracking works │
└──────────────────────────────────────────────┘
Phase 2: Test V2 (one server)
┌──────────────────────────────────────────────┐
│ Enable V2 for test server │
│ Disable V1 for that server │
│ Observe behavior for 24 hours │
│ Tune thresholds if needed │
└──────────────────────────────────────────────┘
Phase 3: Full rollout
┌──────────────────────────────────────────────┐
│ Switch all servers to V2 │
│ Remove V1 scheduler code │
│ Keep V1 code as fallback │
└──────────────────────────────────────────────┘
Phase 4: Enhance (future)
┌──────────────────────────────────────────────┐
│ Add topic tracking │
│ Add user affinity │
│ Add sentiment signals │
│ ML-based threshold tuning │
└──────────────────────────────────────────────┘