Small model, right tools, closed loop

A clip of Andrej Karpathy has been going around this week: agents aren't magic, they're distillation at scale. The real formula is smaller than the hype — a small model, the right tools, and a closed loop. The 99.99% of LLM capacity everyone is paying for is mostly wasted on garbage data. Give a small model the right context, the right tool, and a way to learn from what happens next, and the capability becomes terrifying.

“

Agents aren't magic. They're distillation at scale. Small model + right tools + closed loop = terrifying capability.

Distillation at scale

For a student, distillation means the model doesn't need to know everything. It needs to know the right things: the syllabus, the last three mistakes, the topic that keeps showing up, the exact phrasing the teacher uses. The rest is noise. The more you strip away, the better the answer gets.

That is why a raw frontier chatbot is often the wrong tool for homework. It answers correctly, but it answers in a vacuum. It has no idea which concept the student is actually weak on, no memory of the last explanation, and no stake in whether the student still understands it tomorrow. The context is the distillation.

The right tools

The clip's second term is the right tools. For a student, the tools are not the model. The tools are the loop that turns a question into long-term understanding:

• A syllabus-aware prompt envelope, so the answer stays inside the class's actual scope instead of drifting into a generic lecture.
• A memory graph that tracks every weak spot, every correction, and every concept that needs to come back later.
• A savings meter that routes the question to the cheapest model that can actually handle it, in real time.

The router is the tool. The memory is the tool. The savings display is the tool. The model is just the engine they point at the right problem.

The closed loop

The third term is the closed loop. A student asks a question, gets an explanation, tries a quiz, gets a correction, and sees the same concept again when they are about to forget it. Every step feeds back into the memory graph. The loop is what makes the model useful; without it, the model is just another answer box.

“

The terrifying capability is not the model. It's the loop that turns one answer into a month of better answers.

Why the routing matters

Karpathy's clip also explains why the savings meter is not a gimmick. If 99.99% of LLM capacity is wasted on garbage data, then the most expensive model is almost always the wrong choice. The right choice is the smallest model that has enough context to be right. Route every question there, and the student gets the same answer for a fraction of the cost.

That is the honest economics. The model layer is a commodity. The routing, the memory, and the closed loop are the moat. The clip is about agents in general. The student version is the same principle applied to one person's learning graph.