Ethan C. Jackson, PhD

This morning I read A Definition of AGI. I appreciate this contribution. I think it is useful to propose a clear framework for comparing AI to human intelligence, and to distinguish “AGI” from other forms of AI progression, such as Recursive AI or Superintelligence. Here I’m recording my immediate reflections after reading the paper.

RAG and agentic scaffolds: would they help close the gap?

The paper clearly identifies long term memory, and particularly experiential memory, as a fundamental gap between current models and its definition of AGI. It is then argued that retrieval-augmented generation (RAG) is “not a substitute for the holistic, integrated memory” required for models to achieve long-term contextual understanding. In my mind this reads as: RAG is not a substitute for a proper hippocampus. And while I have long been convinced that, indeed, LLMs are missing and would greatly benefit from such a deeply integrated mechanism for experiential memory, I also think that we shouldn’t be so hasty to dismiss RAG (or at least a broader interpretation of what RAG can mean) from further experimentation.

I’ll be a little more direct. In the paper, we’re presented with results that show GPT-5 outperforming GPT-4 towards AGI, but where GPT-5 still has significant gaps to close are in the dimensions related to memory. What I would love to have seen next would be the question: “To what degree can the memory gap be closed by popular techniques such as RAG?” And I would add to that: there are many variation of RAG including ones that go well beyond document retrieval, and several tools/articles/resources have been published describing how semantic and episodic memory systems can be built on top of the same underlying vector databases and retrieval techniques as basic RAG systems. And as LLMs are improving in their ability to reason and use tools, I think it would be reasonable to at least measure how well an externalized experiential memory system could help close the gap between the paper’s results and the AGI target.

More simply: what would the results look like for a memory-enabled agent using the same models?

Where do goals and motivations fit in?

Next, after reading this paper, I find myself still unsure how to think about the relationship between an AI system’s intelligence and its motivations, or its broader behaviour. Perhaps it was by design to strictly define AGI in relation to an established psychometric framework but, to me, this doesn’t describe anything about the characteristics of autonomy or agency that we should be tracking. Perhaps an AI agent’s goals and motivations are separate from its intelligence? No, I really doubt that. Perhaps social behaviour falls more neatly into the category of Self-Sustaining AI? To me it feels like there is an entirely missing lens which needs to consider AI systems and agents as autonomous, socially accountable actors. Do we think of AGI as being necessarily capable of setting and acting upon its own goals, or are we restricting the discussion to performance on assigned tasks? Does it make sense to measure proficiency, absent of motivation, or other behaviour-determining factors?

I don’t think I’m articulating a critique of the paper, but rather I’m highlighting in my own reflection that I’m still not sure where I draw the line between intelligence and agency. I think these concepts are likely to be deeply intertwined as AI progresses. Perhaps AGI is reasonably defined in terms of versatility and proficiency alone. But I don’t think this will help us to answer what the most salient questions are. As AI systems become more intelligent and autonomous, what will shape their goals and motivations? Who will have the ability (or not) to steer them? What algorithmic and architectural choices will affect these?