🔑 AI Can't Actually Think

Reading time: 8 minutes / Become my affiliate / Sponsor this newsletter

Greetings from above,

What's the difference between a calculator and a mathematician? One follows instructions. The other understands why they work.

Last week, Robert sent me a link at 2 AM.

Just a PDF. No message. That's his way of saying "this changes everything."

It was a research paper from Apple. Title: "The Illusion of Thinking." And it proved something I've suspected for months while building prompts for a living. These AI models we use every day don't reason. They pattern-match. And there's a hard ceiling to what pattern-matching can do.

As someone who makes a living writing prompts, this isn't scary. It's a competitive advantage. Because if you understand what AI actually does under the hood, you can use it 10x better than anyone who thinks it's "thinking."

Today's newsletter will show you:

Let's build your competitive advantage!

Imagine this. You open your phone to an alert. It says, “you spent $236,000,000 more this month than you did last month.”

If you were the top bidder at Sotheby’s fall auctions, it could be reality.

Sounds crazy, right? But when the ultra-wealthy spend staggering amounts on blue-chip art, it’s not just for decoration.

The scarcity of these treasured artworks has helped drive their prices, in exceptional cases, to thin-air heights, without moving in lockstep with other asset classes.

The contemporary and post war segments have even outpaced the S&P 500 overall since 1995.*

Now, over 70,000 people have invested $1.2 billion+ across 500 iconic artworks featuring Banksy, Basquiat, Picasso, and more.

How? You don’t need Medici money to invest in multimillion dollar artworks with Masterworks.

Thousands of members have gotten annualized net returns like 14.6%, 17.6%, and 17.8% from 26 sales to date.

Explore offerings

_{*Based on Masterworks data. Past performance is not indicative of future returns. Important Reg A disclosures: masterworks.com/cd}

Apple's researchers didn't just run a few tests. They built controlled puzzle environments to stress-test AI reasoning from scratch. Tower of Hanoi. River crossing. Block stacking. Checker jumping.

These aren't trick questions. They're the same logic puzzles we use to test if children can reason in steps. And the results were brutal.

Read the full paper here.

Apple found that AI models operate in three distinct performance zones. Understanding these zones is the single most important thing you can learn about working with AI.

Zone 1: Low Complexity (AI wins easily)

Simple tasks. Short chains of logic. Here's the kicker for this zone: "thinking" models actually perform WORSE than regular models. The standard model solves it faster, cheaper, and more accurately. The reasoning model wastes tokens overthinking a simple problem.

Zone 2: Medium Complexity (Thinking models shine)

This is the sweet spot. Problems with moderate depth where chain-of-thought reasoning gives a real advantage. The AI explores paths, self-corrects, and arrives at better answers than quick-response models.

Zone 3: High Complexity (Total collapse)

Both model types fail. Zero accuracy. Complete breakdown. And here's what's wild: as problems get harder, the AI doesn't try harder. It tries LESS. It generates fewer reasoning tokens. Uses fewer steps. Explores fewer solutions.

It's like watching a student stare at an exam question, write nothing, and hand it in blank. Except this student cost billions of dollars to train.

Apple measured the exact number of tokens each model dedicated to reasoning at every difficulty level. The data showed a clear curve:

The models still had thousands of unused tokens available. They just... stopped trying. This isn't a resource problem. It's a fundamental ceiling in how these systems work.

A human facing a hard problem leans in. Tries new angles. Spends more time. AI does the opposite. It quietly surrenders.

Here's where it gets really interesting for prompt engineers.

The researchers gave the AI the exact solution algorithm for the Tower of Hanoi. Step by step. All the model had to do was follow the instructions. Copy the homework.

It couldn't do it.

Performance was essentially the same whether the model figured it out alone or was handed the answer. The collapse happened at the same complexity point.

This tells us something critical: the limitation isn't about strategy or creativity. It's about execution. These models struggle to consistently follow logical steps across long sequences. They lose track. They contradict themselves. They repeat moves they already made.

On simple problems, something weird happens. The AI finds the correct answer early in its reasoning. Then it keeps going. It explores wrong paths. Second-guesses itself. And sometimes changes its correct answer to a wrong one.

Apple called this "overthinking." The model wastes compute budget on problems it already solved.

On harder problems, the pattern reverses. The model explores wrong paths first and (sometimes) arrives at the right answer later. But beyond a certain complexity, it never finds the right answer at all. It just generates wrong solution after wrong solution.

If you're using AI for business, this research hands you a huge advantage over anyone who doesn't know it. Here are the practical takeaways:

1. Stop using "thinking" models for simple tasks.

If your task is straightforward (rewrite this email, summarize this doc, format this data), use the standard model. It's faster, cheaper, and Apple's data shows it's actually more accurate for simple work. Save the reasoning models for problems that genuinely need multi-step logic.

2. Break complex problems into medium-complexity chunks.

The "collapse zone" is where AI goes to die. But the "medium zone" is where it performs brilliantly. Your job as a prompt engineer is to never send a prompt that lands in Zone 3. Break it down. Chain it. Feed outputs into new prompts. Keep each step in the sweet spot.

3. Never trust AI to self-verify on hard problems.

The research showed that AI self-correction is extremely limited. It can fix small mistakes on medium problems. But on hard problems, it fixates on wrong answers and wastes the rest of its budget defending them. You need to be the verification layer.

4. Treat AI as a pattern-matching engine, not a reasoning partner.

This isn't pessimism. It's precision. When you understand that AI excels at finding patterns in its training data and applying them to new situations, you stop asking it to "figure things out" and start asking it to "apply this specific framework to this specific situation." That's when results get insane.

💡 Use this prompt before sending any complex task to AI. It forces you to break your problem into the right-sized pieces so you stay in Zone 2 where AI actually performs.

Customize these variables:

What you learned today:

This knowledge transforms you from someone who "uses AI" into someone who engineers AI systems that work.

No more sending complex tasks into the void and hoping for magic.

You now understand the engine. And that's the only way to drive it properly.

And as always, thanks for being part of my lovely community,

Keep building systems,

🔑 Alex from God of Prompt

P.S. What's the hardest task you've tried to get AI to do in one shot? Reply and tell me. I'll show you how to break it into the sweet spot zone.