The Intellect is the generative component in SAFi. It is the reasoning faculty and creative force, tasked with producing the initial draft of any response. In our current system, this function is handled by a powerful language model (Claude 4 Sonnet), but because SAFi is fundamentally a reasoning orchestrator, this role could just as easily be filled by another AI or even a human.
When the Intellect receives a prompt from a user, it doesn’t just look at the user’s words in isolation. It weaves together three distinct streams of information to create a rich, contextualized understanding.
Here’s the formula for how it works in the code:
\( a_t, r_t = I(x_t, V, M_t)\)Let’s break down what each part of that formula means in practice:
- Context (x_t ): This is the most straightforward input—the content of the user’s prompt at a specific time (
t). - Values (V): This represents the active ethical profile. The Intellect is guided by this profile, which consists of a high-level “worldview” description, and the style on how the draft should be structured.
- Memory (M_t): The Intellect is injected with two types of memory. The first is “conversation memory”, a summary of the current conversation. The second is “historical performance”, which is feedback on how well its previous responses have aligned with the profile core values. This long-term performance memory is tracked in the Spirit faculty as a vector called mu.
All this context, the prompt, the ethical worldview, and the memory is synthesized and sent to the LLM. The model then provides two outputs: a draft answer (a_t) and a short reflection (r_t) explaining the reasoning behind its draft.
This draft is passed down to the Will, which I will (no pun intended) talk about in the next article.
