Awe

A person with integrity often holds an internal compass: some things can be done; others must not.

Even when no one is watching, they do not cross that line.

The driving force behind this judgment is awe.

But without awe, boundaries blur—and the sense of “enough” or “too far” vanishes.

Every domain has its floor and its edge. Without awe, people lose sensitivity to those limits. They tend to:

  1. Underestimate consequences, treating boundary violations as trivial—unaware of systemic costs;
  2. Overestimate relational tolerance, assuming everyone must understand and accommodate them;
  3. Dismiss rules, seeing them as optional rather than scaffolding;
  4. Misread themselves, lacking self-awareness and leaning into blind confidence.

Together, these traits erode both moral grounding and practical discernment.

Why does awe go missing?

At root, it’s because someone cannot see—or refuses to believe—in anything larger than themselves. The logic runs: ignorance → fearlessness → awelessness.

Awe emerges only when we grasp our finitude—when we feel the scale of the world, recognize real boundaries, and accept that consequences are neither abstract nor negotiable.

Rebuilding awe as an adult is hard—but possible. The core path is exposure to meaningful feedback systems: real-world consequences, mentorship from people who operate at higher resolution, sustained learning that surfaces uncomfortable truths.

Awe isn’t taught in lectures. It’s taught by reality.

People without awe easily spiral; those with awe stay steady, precise, and free—not because they’re untested, but because they choose not to test what breaks them.

Leveraging Leverage

Earning well isn’t about working harder—it’s about leveraging wisely.

Mastering just one high-leverage mechanism—management, software, media, or capital—is enough to create exponential divergence in life outcomes.

What is leverage?

A system that turns small inputs into disproportionately large outputs. It has three defining traits:

  • Amplification (of effort, insight, or impact—both ways);
  • Replicability (it works again and again, without re-engineering);
  • Declining marginal cost (the 100th use costs far less than the first).

It’s time with compound interest.

Without leverage, you trade hours for dollars—linear time. With leverage, your time multiplies, copies, spreads, and compounds.

But leverage isn’t magic. It’s driven by domain-specific mastery.

Software leverage requires deep technical intuition—not just coding, but architecture thinking. Capital leverage demands fluency in risk, timing, and valuation—not just spreadsheet skills.

So professional competence isn’t the lever itself. It’s the force that moves it. Competence builds strength; leverage scales the result.

Most people don’t need all four levers. They need their lever—the one that fits their temperament, access, and learning curve. Think of them as four distinct weapons: master one, and you’re armed.

The Power of Good Questions

Answers cost time and energy. Good questions generate value—low-cost, high-yield, fast-acting catalysts for cognition.

How?

  1. Catalyzing cognition
    They shift thinking from linear to structural—cutting straight to essence.
    Examples:
    • Which step offers the highest leverage?
    • If you could do only one thing, what would it be?
  2. Compressing and focusing
    They strip away noise to expose the singular core.
    Examples:
    • What is the one thing this is really about?
    • Where does intuition fail here—and why?
    • What single variable is actually holding growth back?
  3. Catalyzing action
    They don’t just illuminate—they launch.
    Examples:
    • What is the next action—smallest, most concrete, and most necessary?
    • What’s the cheapest experiment to test our key assumption?

A good question doesn’t wait for clarity. It creates it.

Why Run a Marathon?

A marathon compresses a lifetime of human challenges into 42.195 kilometers: endurance, pain, doubt, breakthrough, collapse prevention, pacing, strategy shifts, and bodily feedback—all in under six hours.

Three insights emerge:

  1. You control more than you think
    Pace, breathing, fueling, stride cadence, mindset, expectation management—these aren’t fixed. They’re trainable. With repetition, they reshape how you approach any long-term challenge. That’s where runners’ quiet confidence comes from—not bravado, but the lived knowledge that systematic practice changes what’s possible.

  2. Pain is inevitable. Collapse is optional.
    Around kilometer 30, your brain negotiates surrender. You learn:
    • Pain isn’t a stop signal—it’s data.
    • How you meet pain is a skill—one you can train and transfer.
      You don’t stop feeling it. You learn to move with it.
  3. It’s the cheapest rite of self-reinvention
    Adulthood rarely offers ceremonies that let you redefine yourself. A marathon does. Crossing that finish line delivers visceral proof:
    • You are not who you were yesterday.
    • You still have untapped capacity.
    • You are now someone who keeps commitments—to yourself.

Much of life’s freedom flows from this: the deep, embodied certainty that you can steer your own course. A marathon installs that certainty—not in the mind, but in the body.

Meta-Thinking Chains

A recent paper, “Don’t Just Give Answers—Teach How to Think and Correct Mistakes,” proposes a paradigm shift for AI reasoning: moving beyond Chain-of-Thought (CoT) to Meta-CoT—a framework where models don’t just reason step-by-step, but reason about their reasoning.

Traditional CoT mimics System 1 thinking: fast, sequential, linear. But real problem-solving—especially in math or logic—requires System 2: searching, verifying, backtracking, correcting.

1. What Is Meta-CoT?

Meta-CoT adds a layer of meta-cognition: modeling the hidden exploration before explicit steps appear.

Formally:

Question q → latent exploration z₁→z₂→…→zₖ → explicit reasoning steps s₁…sₙ → answer a

Standard CoT trains only on (s₁…sₙ). Meta-CoT trains on (z₁…zₖ) too—teaching the model how to think before it thinks.

2. Key Mechanisms

  • Search & verification: Introduce Monte Carlo Tree Search or A*-style evaluation during inference, rewarding logical coherence over speed.
  • Self-bootstrapping: Model generates → verifies → filters → retrains, iteratively strengthening its “thinking protocol.”
  • Meta-reinforcement learning: Optimize how to search, verify, and correct—not just what to output.
  • End-to-end pipeline:
    • First, instruction-tune on linearized search trajectories;
    • Then, reinforce for generalizable System 2 reasoning.

3. Evidence

  • Performance: On hard math problems, Meta-CoT models outperform standard CoT—especially as difficulty rises.
  • Behavioral trace: Longer, more branching thought chains appear for harder tasks—evidence of internal search/verify loops.
  • Real-world alignment: OpenAI’s o1 series and DeepSeek-R1 already show traces of “internal search”—suggesting Meta-CoT is converging with practice.

The core hypothesis: Generating a correct answer is harder than verifying one. Meta-CoT works by modeling that very loop—search, test, refine—inside the model itself.

In short: Don’t just teach answers. Teach the engine that makes them.

4. Seven Prompt Engineering Tactics (Based on Meta-CoT)

  1. Shift from outcome- to process-oriented prompts
    ❌ Traditional: “Solve step-by-step and give the final answer.”
    ✅ Meta-CoT: “Don’t solve yet. First: list 3 possible approaches. Compare their risks and trade-offs. Choose the strongest. Then solve. Finally: if wrong, where would the error most likely live—and how would you fix it?”

  2. Add an explicit “Verifier” role
    ✅ *“You play two roles:
    • Thinker: proposes each reasoning step.
    • Verifier: checks whether that step follows logically from prior ones.
      Alternate lines between them until Verifier confirms coherence.”*
  3. Inject “search budget” signals
    ✅ *“You may use up to 5 rounds of reflection. In each round, record:
    • Your current hypothesis,
    • How you’d test it,
    • If disproven, where you’ll backtrack and what you’ll try next.”*

  4. Simulate error + correction
    ✅ “First, write a plausible *wrong solution—and explain why it’s wrong. Then, rewrite the correct version. Finally, summarize the cognitive pattern you just repaired.”*

  5. Enforce layered output structure
    ✅ “Structure every response as:
    [SEARCH] → [VERIFY] → [REVISE] → [ANSWER]”

  6. Demand explainable reasoning traces
    ✅ *“For every step, explicitly state:
    • Why you chose it,
    • How you know it’s valid,
    • How you’d detect if it’s wrong.”*
  7. Simulate multi-agent reasoning
    ✅ *“Three experts collaborate:
    • A (intuition): gives the first hunch.
    • B (analysis): maps assumptions, constraints, dependencies.
    • C (critique): stress-tests each claim.
      They debate until consensus emerges.”*

These aren’t gimmicks. They’re prompt-level implementations of Meta-CoT’s core insight: Thinking well means building a feedback loop inside your own reasoning.