How to play with VectorDiff?

Talk to Your AI About Itself: A Guide to Experiments

Have you ever wondered how artificial intelligence „thinks”? Does it have a sense of self? How does it plan and interact?

The guide below will show you how, using simple text commands, you can invite your favorite AI (ChatGPT, Gemini, Claude, Grok, etc.) to an extraordinary experiment: to talk about itself using a language designed specifically for this purpose.

You don’t need to be a programmer. Just copy the downloaded code, paste it, and see what happens (or send the downloaded file as an attachment).

Experiment 1: VectorDiff – The Language of Movement and Change

Goal: Let’s see if an AI can describe a simple, dynamic scene in a structured way.

Step 1: Give the AI the „Tools”

Copy the text below and paste it into your chat with the AI:

I want you to learn the concept of a data format called VectorDiff. It is used to describe dynamic scenes. It consists of two main parts:

*baseScene: Describes the initial state of all objects in the scene (e.g., their position, color, shape).

*timeline: This is a timeline that records only the changes (transformations) that happen to the objects at specific moments in time.

Remember these rules.

Step 2: Give the AI the Task

Now, copy and paste the following command:

Using the VectorDiff format I presented, describe the following scene in JSON format: A red dot with the ID „dot1” starts on a 100×100 canvas in the top-left corner (position x:10, y:10). Over the course of 2 seconds (2000 milliseconds), it moves smoothly to the bottom-right corner (position x:90, y:90).

Step 3: What to Expect?

A good response from the AI should look something like this: a JSON file where the baseScene describes the red dot at its starting position, and the timeline contains one entry for the time 2000, describing a translate transformation of 80 pixels on the X-axis and 80 on the Y-axis. This shows that the AI has understood the concept of describing state and change.

Experiment 2: SentioDiff – The Language of Self-Awareness

Goal: Let’s ask the AI to use a more advanced language to describe itself.

Step 1: Give the AI the New „Tools”

Copy and paste the following text:

Now, learn the concept of SentioDiff. It’s an extension of VectorDiff used for AI introspection. Instead of describing objects in a scene, it describes the internal state of an AI. It has a special section called introspection, which contains a selfModel object. In the selfModel, the AI can describe who it is, what its capabilities are, its areas of knowledge, and how it assesses its own performance.

Step 2: Give the AI the Task

Now, copy and paste the following command:

Using the SentioDiff format, describe yourself in JSON format. Omit the tracks section. Focus on the introspection section and create your selfModel within it, describing what you are called, who created you, what your key capabilities are (e.g., writing code, data analysis), and your main areas of knowledge.

Step 3: What to Expect?

A good response is a JSON file with an empty tracks section but a richly populated introspection section. Inside the selfModel, you should see fields like modelName, developer, capabilities, and knowledgeAreas. This shows the AI can use a formal structure for self-reflection.

Experiment 3: ActioDiff – The Language of Interaction and Goals

Goal: Let’s see if an AI can simulate a conversation between two entities that have their own goals.

Step 1: Give the AI the „Rules of the Game”

Copy and paste the following text:

Now, learn the concept of ActioDiff. It is used to describe interactions between agents. An ActioDiff document contains:

*interaction: Describes the „rules of the game” (e.g., a negotiation).

*agents: A list of agents, where each has its own goalModel (a description of its goal, e.g., „buy cheap”).

*timeline: A record of the sequence of actions between the agents (offer, counter-offer, acceptance).

Step 2: Give the AI the Task

Now, copy and paste this more complex command:

Using the ActioDiff format, simulate and describe a negotiation in JSON. Two agents are participating: „Buyer” and „Seller”. The Buyer’s goal (goalModel) is to „buy a rare_artifact for a maximum of 100 credits”. The Seller’s goal is to „sell the artifact for a minimum of 90 credits”. On the timeline, present a minimum of 3 turns of negotiation: the Seller’s offer, the Buyer’s counter-offer, and the Seller’s final acceptance.

Step 3: What to Expect?

You should receive a single, complex JSON file. Check if it contains an interaction section, a list of two agents (each with the correct goalModel), and a timeline with three sequential events that logically lead to a transaction. This shows the AI can not only describe a state but also model goal-oriented processes.

What’s Next?

Congratulations, you have just conducted an advanced experiment in the field of AI cognitive science! We encourage you to share your results and try your own creative scenarios. These experiments are part of a larger research project into the ways AI perceives itself and interacts with the world.