How I Use AI to Learn Things

Main ideas / lessons

• AI can optimize learning by improving how instruction is delivered and how a learner’s mental effort is used.

• Current education is inefficient because it uses:

  • Many-to-many matching between learners and teaching sources (teachers/books/courses).
  • Two inefficiencies:
    1. One outlet → many learners: a single “one-size” teaching arc and set of explanations can’t be optimal for any particular learner because optimal teaching depends on the learner’s current understanding and their path to the goal.
    2. One learner → many outlets: learning across multiple interfaces/sources/styles creates overhead—especially loss of trust, where the brain hesitates to fully accept facts from unfamiliar/less reliable outlets.

• Ideal scenario: one-to-one in both directions

  • Ideally, one teacher per learner and one learner per teaching path—but AI should act as the single interface that aggregates many perspectives rather than reducing them.

• Trust must be “engineered,” not built over time

  • Instead of gradually earning trust, an AI system should rely on:
    • reliable verification and fact-checking (avoid hallucinations)
    • making the learning process easier when reliability is clear

• Central principles of the system

  • Optimized teaching: tailor explanations and the learning arc to the learner’s current understanding.
  • Optimized allocation of mental resources: reduce cognitive overhead devoted to “logistics” (planning what to learn, finding resources, verifying facts), so the struggle is concentrated in the material itself, not in the process.

---

Method / process described

A) Overall workflow phases

1. Phase 1: Probe (measure current understanding)

   • Use a quiz tool with graded multiple-choice questions.
   • Start broadly and narrow down (described as similar to binary search on the dependencies/prerequisites).
   • Output: a detailed map of the learner’s understanding.

2. Phase 2: Plan (reason and construct an individualized teaching plan)

   • The system reasons about how to teach that specific learner the target concepts.
   • It also runs fact-checking / verification sub-agents.
   • It presents the resulting plan as a Mermaid graph/diagram.
   • The diagram serves two purposes:
     • helps the learner understand what’s coming
     • forces the AI to explicitly reason rather than “wing it”

3. Phase 3: Teach (instruction + ongoing feedback)

   • The teaching style is intentionally customizable to the learner’s philosophy and preferences.
   • Key component regardless of style: feedback loop
     • The AI quizzes periodically to confirm understanding and prevent self-deception.
     • Feedback keeps the system calibrated to the learner’s evolving understanding.
     • Practice/application helps “lock in” new understanding.

B) Design choices to support the methodology

• Move cognitive load from process → content

  • The system handles logistics, including: planning/order, fact verification, and figuring out what to learn.
  • The learner’s effort focuses on difficult concepts.

• Use specialized sub-agents

  • Visualization agents create/verify visuals (e.g., rendered diagrams/SVGs).
  • They check whether visuals are correct by reviewing generated images.

• Control teaching “pace”

  • During teaching, the narrator prefers the AI to proceed slowly, step-by-step, rather than rushing through explanations.

---

Demo walkthrough (what was actually done)

• The narrator demonstrates implementing the system in a local “learning directory” with an Obsidian-based workflow.

Tools / setup mentioned

• Obsidian used as the learning interface / UI.
• An “MD log extension” links Markdown files to the session so session content appears with LaTeX rendering and persistent artifacts per learning session.
• “teach skill,” a quiz extension, visualization-related skills, and sub-agents.

Learning target chosen for the demo

• Differential forms, inspired by an earlier discussion about Maxwell’s equations expressed using differential forms.

Observed probe behavior

• The system asks multiple questions, including ones about:
  • line integrals (example: force field along a curve)
  • divergence/flux
  • Stokes theorem
  • Faraday’s law
  • special relativity transformation of electric/magnetic fields
• The narrator notes this was a long probing phase due to limited prior knowledge.

Observed planning behavior

• The AI produced a plan diagram (Mermaid/graph).
• It planned to reach advanced topics such as generalized Stokes before the final target.

Observed teaching behavior

• The system generates visuals (SVGs) and embeds them in Obsidian.
• It teaches by walking down a dependency tree (introduced concepts like co-vectors, then covector fields, wedge products, etc.).
• It quizzes the narrator after key reasoning steps.
• A few operational issues occurred (an “overloaded”/agent error while generating visuals), but the demo continued.

Wrap-up lesson from the narrator

The approach worked largely because the system implements a previously described teaching style that shapes: 1. the overall learning arc (path from current understanding to goal) 2. the individual explanation steps

---

Speakers / sources featured (at end)

• Speaker: The video narrator (author describing their own system and demo)
• Source referenced: 3Blue1Brown (referenced as a familiar, trustworthy explanation source)
• Source referenced: A random user on X (Twitter) (contrasted with 3Blue1Brown for trust effects)
• Framework/tool referenced: Obsidian (used as the UI for logging/rendering)
• Model referenced: “Kimik 3” (named as the AI used for teaching in the demo)
• Math concept references (non-speaker): Maxwell’s equations, differential forms, Stokes theorem, generalized Stokes, line integrals, divergence/flux, Faraday’s law, special relativity, co-vectors, wedge product, generalized differential forms.