Mini AGI can be created in two ways: one large model with a breakthrough that can take the current transformer to the next level. Or multi-models with different abilities that work to achieve user tasks and break into multiple steps.

The second option is more achievable in the short term, for this, I create a framework that uses graph theory to navigate the AI models as multiple tasks and handle the relations between different models.

This example shows how to create a data science teaching assistant using open-source Gemma and stable diffusion: https://www.kaggle.com/code/jaguar00/data-science-teaching-a...

This article describes the concept in detail: https://towardsdatascience.com/graph-theory-to-harmonize-mod...