We have been talking a lot about AGI these days, to the point that AGI is not seeing anymore as a bad or naive/dillusional term and accepted even in serious business and academic communities.

However it’s hard to trace a path of previous and future development in the field.

I’m not a big fan of everything that Musk says but I believe he’s right about this idea of deconstructing everything to its core fundamental principles.

Also, building a framework to build on top is really important.

What’s AGI? What are the main ideas? What will be the main components of a v0? What are we missing? What are the road blocks? What are the short and long term products/deliverables that can come from it while we still don’t have it but could be important to justify and fund its research?

I plan to review, study, discuss, prototype, interview and monitor the development of AGI, facing it as an engineering problem like any other, reducing it in smaller problems in order to solve (or die trying 😅) it.

Pre-AGI Ideas

• Transformers
• LLM
• Generative Models
• World Models
• Mamba

Existing Theories and Ideas

• LeCun’s energy based models that optimize in a conceptual/embedding space (JEPA)
• Leggs’ non-episodic LLMs
• Ilya’s scale is all you need (fully learned pipeline)

Components of a AGI

Based on neuroscience, behavioral psychology and AI.

• World model representation
• Long non-episodic memory