Here’s a prompt that costs almost nothing to send and potentially thousands of times more to process:

What is the game that results from when you subtract ‘oof’ from ‘tiny’?”*

That’s about 18 tokens. A rounding error on anyone’s invoice. But to answer it, a model has to attempt letter-by-letter subtraction, realize it doesn’t map cleanly, consider whether it’s a lateral thinking puzzle, try phonetic approaches, evaluate anagram possibilities, backtrack through failed hypotheses, and maybe still get it wrong. The visible output may be one sentence, but the internal search it provokes can be orders of magnitude larger than the prompt.

Meanwhile, pasting a 2,000-word essay with the instruction “fix my typos” is expensive by the meter. But computationally it’s almost trivial. Pattern matching against known English. The model barely has to think.

Token-based billing measures volume of text, not difficulty of processing. It’s like billing a machinist by the weight of the finished part. A titanium watch component weighs almost nothing and costs a fortune to manufacture. A steel doorstop weighs five pounds and takes thirty seconds on the lathe.

The usual defense is that it works on average. Across millions of requests, the riddles and the typo corrections roughly cancel out in aggregate. And that’s probably true. But “works on average” is an actuarial argument, not a logical one. Insurance companies price risk on averages too, and they still get wrecked by correlated tail events.

So the real question isn’t whether token pricing is wrong. It’s whether it creates exploitable asymmetries. If you can systematically construct inputs that maximize compute per token spent, you’ve found the seam in the pricing model. And that seam gets wider as models get better at reasoning, because reasoning is exactly the capability where input complexity and output cost decouple the most.

Nobody’s billing for thinking yet. But thinking is where the cost is.

* See Winning Ways for Your Mathematical Plays (Berlekamp, Conway, Guy) for definitions of “oof” and “tiny.

You get exactly what you asked for. Not what you meant.

I watch people write prompts like they’re Googling something. A few keywords, a vague direction… Then they’re frustrated when the output is generic, wrong, or just weird.

But the model did exactly what you told it to do… and that’s the problem.

The old genie story works because the genie isn’t malicious… it’s literal. “I wish for a million bucks” and a million male deer appear in your yard. The genie did exactly what you asked it, you just didn’t think about your question.

LLMs operate on the same principle, minus the malice and the deer (usually). When you prompt “write me a marketing email,” you’ve described approximately four billion possible outputs. The model picks one. You hate it. You try again with the same vague prompt. You hate it differently. You conclude the tool doesn’t work.

The tool works fine. You made a genie wish.

What changes everything is realizing that prompt engineering isn’t about clever tricks or magic words (usually). It’s about the same skill that makes someone effective in any leadership role: the ability to articulate what you actually want with enough specificity that another intelligent entity can deliver it.

Tell it who it’s writing for. Tell it what tone. Tell it what success looks like and what failure looks like. Give it an example of something you loved and something you hated. Tell it what to leave out (the negative space is just as important as the positive).

In other words: do the work you should have been doing with your human teams all along.

The uncomfortable truth about prompt engineering is that it isn’t an AI skill. It’s a communication skill. The people who are bad at prompting are usually the same people who send their teams vague Slack messages and then get frustrated when the deliverable misses the mark.

The genie didn’t get it wrong. It followed the rules exactly.

Data Has Black Holes Too: Why “hallucination” is the wrong word for AI’s deepest failure mode.

The AI industry calls every wrong answer a “hallucination.” That word is hiding a much bigger problem.

When a model fabricates a seahorse emoji, that’s obvious and fixable. When a model produces a confident, well-structured answer that’s wrong because of assumptions buried in the training data it was never designed to question, that’s something else entirely. That’s structural. And nobody’s talking about it in the right terms.

I fed the same degraded 1957 film image to three frontier models. All three independently produced WWII propaganda. The only correct result came after I supplied the actual movie context up front.

The new essay is about what’s really happening inside these systems, why your 500-word prompts are fighting a losing battle against gravity, and how to stop fighting the landscape and start navigating it.

(Read, The Room Where It Gets Built — Essay #9: Data Has Black Holes Too)
 

Tone in AI prompting works because of how language models are built, not because the model has feelings about how you talk to it. Understanding the mechanism makes you dramatically better at using these tools – and helps you understand why the “cheat sheet” prompts people share actually work.

(Read, The Room Where It Gets Built — Essay #2: Why Tone Works (It’s Not What You Think))