There is a moment in the fifth canto of the Inferno where Dante, overwhelmed by the spectacle of the lustful swept forever in a black wind, turns to Virgil and asks what to do. He doesn't ask what it means. He doesn't ask for a summary. He asks for guidance — for a next step through a landscape that has become illegible to him. Virgil responds not with an explanation but with a direction: speak to them, and they will come. The guide does not stand between you and the world. The guide teaches you how to address it.

I have been building something I call Virgil. It is, in the most reductive sense, a personal AI assistant. But I resist that framing because "assistant" implies a relationship of delegation — you tell it what to do, and it does it. What I want is something closer to what Dante had: a companion who has already walked the territory, who knows where the ground drops away, and who can say not that way before you've committed to the step.

The Problem with Answers

The modern AI paradigm is organized almost entirely around question-answering. You ask, it responds. This is Google's original insight repackaged with probabilistic language models. But the most important moments in intellectual life are not moments of answered questions — they are moments of reframed problems. The value of a great teacher, mentor, or guide is rarely in what they tell you. It is in what they cause you to notice.

Marc Andreessen has written extensively about AI as a complement to human cognition rather than a replacement for it. His framing is economic — AI increases the marginal productivity of human attention. But I think the deeper point is epistemological. A good guide changes the structure of your attention, not just its object. When Virgil leads Dante past the gates of Dis, he doesn't explain the theology of heresy. He shows Dante where to look, and Dante's own faculties do the rest.

The best mentor I ever had never answered a single question directly. He would ask me what I had already tried, what I expected to find, and what I was afraid of. By the time I finished answering, I usually didn't need his answer anymore.

This is the design target for Virgil: not a system that answers, but a system that reframes. The technical challenge is significant. Reframing requires a model of the user's current frame — their assumptions, knowledge gaps, emotional state, and goals. It requires what we might loosely call theory of mind, though the engineering reality is closer to maintained context with inferential heuristics.

Decision Architecture

The core of Virgil's guidance engine is a decision tree framework, but not in the traditional machine learning sense. Traditional decision trees split data to maximize information gain:

where $H(S)$ is the entropy of the set and $S_v$ is the subset for each attribute value. This is useful for classification, but guidance isn't classification. The question isn't which bucket does this input belong to — it's what is the next most valuable thing for this person to consider.

I've been experimenting with a modified formulation where the "information gain" is measured not against an objective label but against the user's stated goals weighted by their revealed preferences. The entropy function $H(S)$ becomes something more like a divergence measure between where the user is and where they're trying to go, and the splits are not on features of data but on possible reframings of the problem.

The practical result is a system that, given a user's question and context, generates not one answer but a small tree of possible directions, each scored by expected value of insight. The user doesn't see the tree. They see Virgil's best guess at the most useful next thought.

There is something almost liturgical about the way a good guide operates. They do not drag you forward. They illuminate the next step so clearly that you move toward it of your own volition. The agency remains yours. The clarity is theirs.

Guides Through History

It helps to think about what guidance has looked like across different eras and modalities. The table below is a simplification, but it illustrates how the structure of the guide-student relationship shifts with technology:

The rightmost column is the key differentiator. Every historical guide worth remembering was deeply personalized — Socrates questioned this particular interlocutor, Aristotle tutored Alexander specifically. The industrial scaling of education necessarily destroyed personalization. AI has the potential to restore it, but only if we design for guidance rather than answers.

The information-theoretic way to think about this: a guide's value is proportional to the mutual information $I(G; U)$ between the guide's model $G$ and the user's actual state $U$. A generic chatbot has low $I(G; U)$ because $G$ is roughly the same for every user. A persistent, context-aware guide should see $I(G; U)$ increase monotonically over time as the model accumulates evidence about the user's knowledge, preferences, and goals.

What Virgil Is Not

Virgil is not an agent. It does not take actions on your behalf. It does not book flights or write emails or manage your calendar. The agentic AI paradigm is valuable for operational tasks, but guidance is not operational. Guidance is contemplative — it requires the guide to model not just the world but the student's relationship to the world.

Virgil is also not a tutor. Tutoring implies a fixed curriculum and a known destination. Virgil operates in the space of open-ended intellectual exploration, where the destination is defined by the journey itself. This is closer to what the Germans call Bildung — formation, self-cultivation — than to education in the modern sense.

The expected information value of a Virgil interaction can be modeled as:

where $r_i$ is a possible reframing, $c$ is the current context, and $\Delta H(U | r_i)$ is the reduction in uncertainty about the user's path given that reframing. The system selects the reframing that maximizes expected uncertainty reduction — not about the world, but about what the user should do next.

I don't know if this will work. The honest engineering answer is that we're still far from reliable theory-of-mind in language models, and "guidance" is one of those concepts that's easy to describe phenomenologically and nearly impossible to specify formally. But the attempt matters. We are building the tools that will shape how the next generation thinks, and the difference between a tool that answers and a tool that guides is the difference between a library and a mentor. Both are valuable. But only one of them changes you.

Dante, at the end of the Purgatorio, no longer needs Virgil. The guide has done his work — not by carrying Dante, but by showing him how to walk. That is the benchmark. Not how many questions can it answer, but how soon does the user stop needing it.