Jab Petersen / AI Won’t Save You

Chapter 1 · When Answers Became Cheap

AI made answers cheap. Evidence is still expensive.

AI can produce the most productive-looking procrastination we have ever had. The fix is not using AI less. It is using it to get out of the chat and into reality.

On November 30, 2022, OpenAI published a short announcement called "Introducing ChatGPT." There was no robot on a stage. There was a text box. The launch post itself warned that the model could produce plausible-sounding but incorrect answers. The real shift was not that every answer was right. It was that a plausible first answer had become astonishingly cheap.

The bottleneck moved. Many of us kept behaving as if it had not.

The intelligence trap

Imagine Elias. He has been saying he wants to start a business for two years. On Saturday morning he asks an AI assistant for the best low-capital online businesses. Then which has the highest margin. Then a scorecard, competitor analyses, pricing, personas, risks and a twelve-month plan. By midnight he has a brand name, a landing page draft and a financial model.

Nothing has happened.

I know this loop because I have been inside it. When I started building websites for local businesses, I used AI to research more than a hundred companies. Each had a dossier, a price and a lead score. I had demo sites ready and three versions of the first email. I had not spoken to a single owner.

The easier it becomes to think about action, the easier it becomes to mistake thinking for action.

The documents are real. The analysis may even be good. But reality has not voted yet.

Langley had answers too

In 1898 the US War Department gave Samuel Langley, secretary of the Smithsonian, $50,000 to build a manned flying machine. The Wright brothers ran a bicycle shop in Dayton. They were not anti-theory. They were unusually aggressive about letting reality correct theory: gliders, crashes, their own wind tunnel, new measurements, back to the dunes.

On October 9, 1903, after one of Langley's failed tests, The New York Times wrote that a real flying machine might take "from one million to ten million years." Sixty-nine days later, Orville flew for twelve seconds.

A model is built from what has already been written. When you try something that has not been done yet, its best answer is often the old consensus. Use what the machine knows. Then go and find out what nobody has written down yet.

Climb the evidence ladder

Not all information has equal weight. When you want to know whether something is true in the world, there are levels:

  1. Imagination. You think people would want it.
  2. Information. Research suggests the problem exists. AI is extraordinary here.
  3. Conversation. People tell you what they think. Better, but "I would use that" costs nothing to say.
  4. Behavior. They reply, send files, ask a follow-up question.
  5. Commitment. They pay, refer you, or stay after the first project.

Each step is more expensive and more informative. The danger of cheap intelligence is that levels 1 and 2 become so detailed they feel like level 5. A thirty-page AI report can create the sensation of evidence without any behavioral evidence at all.

Keep an Evidence Ledger with three columns: what you believe, what you actually observed, and which level of the ladder it came from. Write the level, not how confident you feel. An interview is level 3. A payment is level 5. That is much harder to fool yourself with.

Ask AI to get you out of AI

Elias's first prompt was What business should I start? A better one:

I believe independent accountants lose time keeping up with regulatory changes. Identify the three assumptions that would kill this business if they are false. For each, design the fastest real-world test I can run in the next forty-eight hours. Prefer tests that require behavior rather than opinions.

Now the model is not extending the conversation. It is helping him exit it. He talks to three accountants. One says the summaries are already free. Another says the real pain is explaining changes to anxious clients. The idea is already changing. A bad idea dying on Tuesday is worth more than a beautiful plan surviving until December.

AI is not replacing reality. It is reducing the time between encounters with reality.

The 10-10-1 rule

When a problem matters:

  • 10 minutes: define the real question. Not "research marketing" but "which of these two customer groups has a problem I can verify this week?"
  • 10 minutes: use intelligence aggressively. Find assumptions, design tests, make the opposing case.
  • 1 external action: create evidence. Send the message. Call the customer. Publish the page. Ask for the sale.

The numbers are not sacred. The rule exists to stop one failure mode: unlimited intelligence consumption with zero exposure to reality.

When to stop asking

At the end of a week, compare how much information you consumed with how much evidence you created. Six hours of tutorials, forty saved prompts and three strategy documents with zero conversations is a bad ratio. One hour of AI preparing ten customer calls is a good one.

When another answer is unlikely to change what you do next, stop asking and act.

And at the end of every AI session, ask one question: What can I know only by doing?