What the Longest Escalator in the Western Hemisphere Teaches Us About AI

If you live in the DC area, you know about the Wheaton Metro station. It’s home to the longest single-span escalator in the Western…

What the Longest Escalator in the Western Hemisphere Teaches Us About AI

If you live in the DC area, you know about the Wheaton Metro station. It’s home to the longest single-span escalator in the Western Hemisphere.

The escalator is 230 feet long, and the ride takes nearly three minutes. When you stand at the bottom and look up, the exit is just a small dot of light. It feels more like entering a sci-fi portal than taking public transit.

The main thing about the Wheaton escalator is that once you get on, you’re committed.

If you realize halfway up that you forgot your keys, you can’t turn around. You can’t run back down because it’s too fast. You can’t jump off since it’s a sheer drop. You have to ride the steel stairs all the way to the top, turn around, and then ride them back down.

This is a lot like how Large Language Models (LLMs) work. It also explains why your ChatGPT conversations sometimes go off track.

The Physics of “Token Inertia”

We tend to treat AI like a human in a conversation. If a human starts telling a boring story, you can interrupt and say, “Skip to the good part!” and they will pivot immediately.

AI doesn’t work that way. Instead, it runs on inertia.

An LLM is basically a prediction engine. It looks at the words that came before to decide what comes next.

  • If the last 50 words were angry, the AI expects the next word to be angry too.
  • If the last 50 words were dry and academic, the AI thinks the next word should match that style.

If you give a vague prompt and let the AI go on for three paragraphs in a boring, corporate tone, it builds up a lot of momentum. It has already stepped onto the escalator.

Even if you interrupt and say, "Be funnier!", the AI often struggles. Why? It looks at its own previous output, that big wall of boring text, and uses it as a guide for what to do next. It’s fighting its own momentum.

The “Step On” Moment

In my book, The Art of Whispering to Machines, I talk about the importance of the System Prompt. Here are some examples to illustrate: If you want a formal tone, try "Write an executive summary about AI technology innovations." For a humorous approach, use "Pretend you are a stand-up comedian discussing AI development mishaps." To evoke a poetic style, you might say, "Compose a sonnet about the future of artificial intelligence." These varied prompts can serve as useful templates for prompt engineers to experiment with.

Think of the System Prompt as the moment you’re standing at the bottom of the escalator. This is the only time you have full control over where you’ll end up.

  • Bad Prompt: “Write a blog post about coffee.”
  • Result: You’ve just stepped onto the “Generic Wikipedia Voice” escalator. It’s a slow, three-minute ride to Boredomville, and you can’t fix it halfway up.
  • Good Prompt: “Act as a caffeinated hipster who is obsessed with bean origin. Write a blog post about coffee.”
  • Result: You’ve just stepped onto the “High Energy” escalator. Now the momentum is working in your favor.

Practical Advice: Don’t Fight the Stairs

The next time you’re chatting with an AI and it starts “hallucinating” by making things up or gets stuck in a logic loop, don’t try to argue with it.

If you argue, you’re just adding more “bad tokens” to the context. It’s like trying to run down the up escalator. Instead:

  1. Hit “Stop Generating” right away. That’s like getting off the stairs.
  2. Edit your original prompt. This is like going back to the bottom.
  3. Start a new chat. That’s like finding a different escalator.

Don’t try to change the ride once it’s started. Fix your approach before you begin. However, there are times when some models are more adaptable or when quick midstream corrections can occasionally yield better results. Acknowledge these possibilities, but remember they are the exception rather than the rule.

By Joshua McDonald on January 19, 2026.

Canonical link

Exported from Medium on August 26, 2026.