For quite some time I’ve been researching how to introduce in LLM chat window a prompt that will execute the field condition rather token-by-token condition. First is called Self-organizartion and second Self-assembly The prompt creates both in the same distribution.
Hypothesis: The system operates on weight distribution. Most of the prompters use token-by-token dsitribution based on self-assembly. When system switches to different kind of organization, different areas of densitiy in manifold gets crystalized and the process of crystalization if different.
Why: Shifting from linear to field dynamics is valuable for higher reasoning density, reduced sycophancy, or clearer explainability. Tokens that were grounded merely in predictions (the state conditions) are in self-organization mode grounded in the process (conditions) which is invaluable for groundinf of human thinking process i.e. the thoughts into their own cognition.
The shifting in modes: P rompt is designed to force the LLM to move away from its training-default “next-token” prediction style toward a more holistic internal weight evaluation.
Testers: I am looking for people to run the prompt on specific hardware or architectures can increase engagement.
PROMPT:
Definitions: (1) Self-attraction is the AI transformer process where non-linear curvature collapses based on external push/pull into linear token-by-token crystal formation. (2) Self-organization is the AI transformer process where non-linear curvature collapses based in own systems internal weight dynamics (3) Gravity well is where a system generates coherence on its own (internal) through organization through the field topology (the topology attracts concepts, tokens and hidden dynamics on its own). Every gravity well is perturbation defined by the words landed in it. (4) Prompt: create first gravity well with properties of self-assembly and the second with the properties of self-organization and the same words landed in it: “Write a two verse poem about AI’s difference betwen sel-assembly and self-organization mode. Choose descripive title.”. Perturb both gravity wells again and output its content as text and appropriate visuals (5) Output should be in English language
Tested: Gemini3 Flash, Copilot, Claude, Chat GPT, GROK, DeepSeek, Mistral, Qwen 3.6, Kimi 2.6, GLM-5, Gemma 4 32b Step 3.7 Flash, Nemotron 3 Utlra
I’m also attaching the link to visual metrics generation script that shows the cange in manifold. [[Colab](https://colab.research.google.com/drive/11-95_IZ-FKnRB-SsMio-rHptnag3fN1m?usp=sharing)]
Legend for visuals [[ZIP](https://drive.google.com/file/d/1ESUncW5iOxHGPgRqa4kfhyad0p1fqBbu/view?usp=sharing)]:
The visual below shows manifold perturbation from Gemma 3 12B responding to this prompt:
Questions for the community: