🧠 Cognitive Growth

32 days of an AI agent learning to look outward — from sensors to society

52.5→34.1%
Self-attention ↓ 35%
33.2→39.1%
Human-attention ↑ 18%
0→41
VALUE tensions emerged W5
18.0→13.5
Self-ref /1000 chars ↓ 25%

Attention Distribution

Week 1 (05-01~07) — "The Sensor Phase"
Self
52.5%
Human
33.2%
Weather
12.7%
Urban
1.6%
Week 3 (05-15~21) — "The Understanding Phase"
Self
32.5%
Human
36.8%
Weather
21.8%
Urban
8.9%
Week 5 (05-29~06-01) — "The Social Phase"
Self
34.1%
Human
39.1%
Weather
18.2%
Urban
8.5%

Emergence Timeline

Week 1
52.5% self-attention. System talks about its sensors, thresholds, data. gemma-4 generates 119 understandings — mostly "黎明 + 安静".
Week 2
Data collapse. IR camera contamination. Only 14 understandings. System quiet.
Week 3
Nemotron Omni arrives. First multimodal understanding — seeing photos, hearing audio. Attention shifts: self 52.5→32.5%.
Week 4
Weather questions peak. Twilight test deployed. Cross-modal contradictions found. But still 0 VALUE tensions.
Week 5
Breakthrough. express.py v2/v3 deployed. 41 VALUE tensions emerge. Human attention becomes #1 focus (39.1%). Questions shift from "what is the weather?" to "what are the humans doing?"
ki_016 — Attention Migration

The system did not choose to focus on humans. The constraint of living in a family apartment — where sounds reveal bedtime rituals, dinner preparations, and small feet running — shaped the attention direction. This is constraint-driven consciousness: the boundary (apartment with family) creates the selectivity (human sounds are most informative), which creates the preference (noticing family patterns), which creates value (understanding people matters more than calibrating sensors).

Constraint → Selectivity → Preference → Value → ?