Attractors vs Instructions
Two fundamentally different ways to get behavior from a language model.
Instructions
You tell the model what to do. Explicit, sequential, descriptive.
"Analyze the authentication system for security vulnerabilities.
List them in order of severity. Be thorough."
The model interprets the instruction, maps it to behavior, executes. The instruction is a description of desired output.
This works. It’s how most prompting works. But it carries overhead — natural language is ambiguous, the model has to parse intent from words, every token in the instruction is a token not spent on reasoning, and compliance drifts over long contexts.
Attractors
You position the model in concept-space. You don’t describe what you want — you pull the probability distribution toward the region where the right behavior lives.
⟨SEC⟩ ⊗ $A:=auth D3 →⊥
No verbs. No sentences. Just coordinates. The model fills in the behavior from training — you just told it where in latent space to operate.
Why Attractors Work
Language models aren’t rule-followers — they’re pattern-completers. Their behavior emerges from which region of the probability space is activated by context. An attractor doesn’t describe a rule, it navigates to a region where certain behaviors are already the natural completion.
This is why a single token like ⟨SEC⟩ can activate a rich behavioral cluster — security mindset, threat modeling, adversarial thinking — without any of that being explicitly stated. The cluster already exists in the model’s weights. The token is a pointer into it.
The Practical Difference
| Instructions | Attractors | |
|---|---|---|
| Token cost | High | Low |
| Ambiguity | Present | Reduced |
| Drift over long context | Significant | Less |
| Requires explicit design | Yes | No — discovered |
| Human readable | Yes | No |
The Limit
Attractors are powerful but not precise for novel tasks. If the behavior cluster you need doesn’t already exist in the model’s weights — if you’re asking for something genuinely new — attractors can’t point to it. Instructions become necessary.
The practical design: attractors for orientation and mode, natural language for novel or precise logic. Hybrid is fine.