Dương Đào Notes A public notebook
← All notes

AI is best understood as cheap prediction

AI makes prediction cheaper. When something becomes cheaper, its complements become more valuable.

1. AI is best understood as cheap prediction

Rather than treating AI as a mysterious “thinking machine,” the authors define prediction as using known information to fill in missing information.

Examples:

  • Translation: predicting the most suitable target-language words.
  • Recommendations: predicting what you will watch or buy.
  • Medical imaging: predicting the probability that an image indicates disease.
  • Generative AI: predicting the next token, or the image most consistent with a prompt.

This framing also explains hallucinations: an LLM is optimized to generate a plausible answer, not inherently to verify truth.

2. The important economic question

Do not ask only: “Will AI replace me?”

Ask:

  1. What does AI make cheaper?
  2. What human or business complement becomes more valuable?
  3. Am I mainly supplying the cheapened activity or one of its complements?

The video compares AI to cheap electricity or lighting. Their impact was not merely lower cost. Cheap lighting changed when people could work, read, and socialize. Likewise, cheap prediction enables entirely new workflows and business models.

3. What becomes cheaper, and what becomes more valuable

Likely to be compressed: tasks mainly involving prediction:

  • Routine translation
  • Basic customer support
  • Standardized reporting and data analysis
  • Information search, summarization, and pattern matching

The key point is that AI usually replaces or reshapes tasks, not an entire profession at once.

Likely to gain value:

  • Judgment: deciding what should be done after receiving a prediction.
  • Data: especially exclusive, relevant, trusted, high-quality real-world data.
  • Action / execution: organizing people and systems to make a decision happen.

A doctor may use AI to estimate a 90% cancer probability. But deciding whether to operate, observe, seek more tests, explain risks, and assume responsibility remains a judgment problem.

4. Prediction is not judgment

A prediction says: “Demand next quarter is likely 10,000 units.”

Judgment asks:

  • What if demand is actually 8,000 or 12,000?
  • Which error hurts more: a stockout or excess inventory?
  • How much risk can we absorb?
  • Who bears the consequences if this decision is wrong?

The video’s strongest practical idea is this: prediction errors are often symmetric, but the costs of errors are not.

Good decision-makers do not merely seek the most accurate forecast. They design a decision that remains acceptable when the forecast is wrong.

5. Three levels of AI adoption

  1. Tool level: summarize documents, write emails, produce reports faster.
    Useful, but soon becomes a normal baseline.

  2. Decision level: use AI-generated signals to make materially better decisions, such as identifying customers at churn risk and changing retention actions.

  3. Strategy level: redesign the business or workflow around cheap prediction, enabling things that were previously impractical.

The real gap will be between companies and people who remain at level 1, and those who reach levels 2 and 3.