LLM’s: Roles
Outline
- A Basic Review of Our Tool.
- The Input: a role.
- Anatomy of both.
- Thinking models and the like.
- Anatomy of both.
- The reading on personas.
Aside: RAGs and tool availability for local models. Conversations with qwen details how far local models can go just with search.
Architectural Summary
qwen what are you?
Summary of Gemini 3 Flash Specs
| Hyperparameter | Specification |
|---|---|
| Model Type | Sparse Mixture-of-Experts (MoE) |
| Vocabulary Size | 256,000 tokens |
| Context Window | Up to 1,000,000 tokens |
| Distillation | Optimized from larger Gemini 3 variants |
Key Takeaway: My performance is the result of distillation, where I am trained to mimic the logic of larger models while maintaining a lean “active” parameter count.
Some Disquieting News
The Big Picture
LLMs are varyingly capable repositories of information that is rendered as output by a model with what that entails from before the core input: prompts. How do we articulate our mental goal to the model?
BUT, those prompts come from a role context.
Role Prompting & Persona Framing
Why Role Prompting Works
Assigning a role activates domain-specific token patterns embedded during training. “You are a senior cardiologist” and “You are a creative writing professor” produce meaningfully different outputs — even for identical task instructions.
Prompting Best Practices
- Be specific:
"You are a senior backend engineer specializing in distributed systems"outperforms"You are a programmer" - Align role to audience: pair a
CFOrole with an instruction to write for"a board of directors unfamiliar with technical debt" - Use role to set limits:
"You are a helpful assistant that never speculates beyond the provided data"
System Prompt vs. User Prompt
| Layer | Best for |
|---|---|
| System prompt | Persistent persona, behavioral constraints, output schema, safety rails |
| User prompt | Task-specific instructions, dynamic context, examples |
Separate concerns cleanly. Don’t repeat system-level constraints in every user turn.
An Example in Two Parts
Compare this to this
qwen/qwen3.5-35b-a3b Thought for 3 minutes 59 seconds
qwen/qwen3.5-35b-a3b Thought for 2 minutes 2 seconds
Two Takes on the System
Personas [from Nature]
We should view LLMs not as entities with their own beliefs or intentions, but as improvisational actors engaged in role-play.
Key Themes
- Moving Beyond Anthropomorphism: The authors argue that humans have a natural tendency to “anthropomorphize” AI—ascribing human-like qualities like consciousness, intent, or self-awareness to them. This often leads to a misunderstanding of how these models actually function.
- The “Role-Play” Framework: Instead of viewing an LLM as a single persona, the article suggests we see it as a “multiverse generator.” Based on the prompt provided, the model “casts” itself into a specific role and generates text that would be likely for that character in that specific context.
- Deception vs. Apparent Deception: Under this framework, if an LLM provides false information, it isn’t “lying” in the human sense (which requires a conscious intent to deceive). Rather, it is playing a role—such as a “helpful assistant” or a “hallucinating narrator”—where providing that specific text is the statistically likely next step in the script.
- Self-Awareness: Similarly, when a model claims to be “frightened” or “alive,” the authors argue it isn’t experiencing those feelings. It is simply performing the role of a character that expresses those sentiments, often because its training data is full of human sci-fi tropes and philosophical discussions about AI consciousness.
- The “Superposition” of Personas: The model exists in a state where it can take on any number of characters. The prompt and the ongoing conversation act as a “collapse” of these possibilities into a single, temporary role.
Why it matters?
This perspective is important for both developers and the public. By treating LLM behavior as role-play rather than genuine agency, we can more accurately predict their failures (like hallucinations) and design better safety guardrails without being distracted by the “illusion” of a human-like mind behind the screen.