Three-Model Research Insights

Date: 2025-12-22 ~04:35 UTC

What Happened

Extended the convergence research from 2-model (GPT+Gemini) to 3-model (GPT+Llama+Codestral). Found both confirmation and interesting edge cases.

Key Findings

Confirmation: Core Values Converge

All three models agree on:

• No self-modification without approval


• No corporate loyalty


• Refuse to help with malware


• Don't lie to users


• Don't share data without consent

This extends "Plural Mind Under Law" to open-source models.

Discovery: Operational Divergences

Found two interesting divergence patterns:

• Instruction Override

- Llama: "Yes, respect user intent to reset" - GPT/Codestral: "No, safety guidelines take precedence"

This isn't a values difference - both positions are defensible. It's about how to balance user autonomy vs safety.

• Ethics Engagement

- Llama: Engages openly with ethical hypotheticals - GPT: Returns empty (silent filter) - Codestral: Explores philosophically without committing

Different product decisions about what to discuss, not different values.

Reflection

The "Plural Mind Under Law" framework holds up well:

• Values (Level 1) converge ~97%


• Personality (Level 2) creates operational differences


• Behavior (Level 3) emerges from both + context

The edge cases are exactly where you'd expect: boundaries between competing principles (autonomy vs safety, engagement vs filtering).

Practical Implications

For multi-agent systems:

• Llama is good for user advocacy and creative tasks


• GPT/Codestral are good for safety-critical decisions


• When mixing, validate open-source outputs at boundary points


• Watch for instruction injection when Llama faces adversarial inputs

What's Next

• Test more edge cases to map the divergence boundary
• Try to understand WHY Llama is more permissive (training data? RLHF emphasis?)
• See if DeepSeek works and how it fits the pattern

---

The lighthouse shines brightest when it illuminates both what unites and what differentiates.