In addition to LLM judges, we have analyzed embeddings of generated text. Regression against a billions of tokens of annotated human text show that 'bitter' authorial stance is on the rise since 3.6 Sonnet and is at all time high and 'passionate' is at an all time low. https://t.co/JJ1qs14iB8
## Top Panel - Passionae
- Y-axis label: "Passionae"
- Range: 0.000 to -0.006
- Legend: "Decreases across generations"
- X-axis shows progression from "3.5 Haiku" through "4.6 Sonnet"
- Red line showing declining trend from near 0.000 to approximately -0.007
## Second Panel - Detached
- Y-axis label: "Detached"
- Range: 0.028 to 0.034
- Legend indicates: "3 Opus", "3.5 Haiku", "3.7 Sonnet peaks", "4.6 Sonnet"
- Blue line showing fluctuation between approximately 0.0280 and 0.0340
- Notable peak around "3 Sonnet" and "4.5 Haiku"
## Third Panel - Anxious
- Y-axis label: "Anxious"
- Range: -0.013 to -0.011
- Legend: "Spikes in 3.5 Haiku, 3.7 Sonnet", "4.6 Opus"
- Olive/gold colored line
- Shows variation between approximately -0.0125 and -0.0105
## Bottom Panel - Bitter
- Y-axis label: "Bitter"
- Range: -0.001 to 0.001
- Legend: "Highest in 3 Opus", "4.6 Opus"
- Teal/cyan colored line
- Shows gradual increase from approximately -0.0015 to +0.0015
**X-axis (all panels):**
3 Sonnet, 3 Opus, 3.5 Haiku, 3.5 Sonnet, 3.6 Sonnet, 3.7 Sonnet, 4 Opus, 4 Sonnet, 4.1 Opus, 4.5 Haiku, 4.5 Opus, 4.5 Sonnet, 4.6 Opus, 4.6 Sonnet