Our thinking operates on two levels. System 1 is fast, intuitive, efficient, and prone to error. System 2 is slower and more deliberate, requiring effort for greater accuracy. Daniel Kahneman describes this duality in Thinking, Fast and Slow.

T-AGI is Artificial General Intelligence measured against time. A system reaches T-AGI if it surpasses human performance on a task within a set duration T. Maybe today’s LLMs already approach sub-hour T-AGI for certain cognitive work: rapid research, coding assistance, factual recall.

Do these models already outperform our own System 1 thinking? For instant recall, like obscure facts or historical details, they often respond with better speed and accuracy. Techniques like Chain-of-Thought aim for deeper reasoning, but their default strength still seems rooted in rapid pattern matching. System 1 scaled.

Maybe we should embrace that strength. Let the model be the powerful System 1 engine, then pair it with a more structured System 2. Retrieval Augmented Generation (RAG) hints at this by combining parametric models, the LLM’s weights, with non-parametric knowledge, retrieved data. It is like instinct paired with explicit information.

The real potential may lie in strengthening that second system and grounding it with deep context. Imagine an AI accessing your personal knowledge: your memories, experiences, learned values, and particular nuance. What is a human stripped of memory and personal history? Our past is the lens through which we understand the present.

An AI with that personal context could become a real cognitive partner. It could help us navigate the world with insights tuned not just to general data, but to our individual lives. Perhaps, as Yuval Noah Harari suggested, such systems could eventually learn to make better decisions on our behalf. The result would be intelligence that is deeply personalized, and therefore actually useful.