AI Isn't Artificial Intelligence. It's a Bicycle for the Mind.

In a well-known interview, Steve Jobs described a study measuring the efficiency of locomotion across species — how much energy each one burns to move a given distance. Humans, on foot, weren't particularly impressive. Near the middle of the pack. But put a human on a bicycle, and suddenly we become the most efficient species measured, by a wide margin.

His point wasn't about bicycles. It was that tools amplify human capability — sometimes so dramatically that the amplified version barely resembles the unamplified one. That's about as good an analogy for AI as exists.

Every real tool follows the same pattern

• Excel didn't replace accountants. It amplified their analytical range.
• Photoshop didn't replace designers. It amplified their iteration speed.
• CAD didn't replace engineers. It amplified their ability to explore designs.
• Autopilot didn't replace pilots. It amplified their operating envelope.

AI is another amplifier in that same lineage. The people who adopt it early don't become smarter. They become more leveraged. The gap it opens isn't intelligence — it's output per hour of applied judgment. (I unpacked that same "judgment, not execution" idea in Autopilot Didn't Replace Pilots.)

But there's a twist software has never had before

Here's where the analogy needs an addendum, because AI's economics genuinely differ from every tool that came before it — including the bicycle.

Traditional SaaS scales beautifully. Build the software once, and the 100th customer costs almost nothing more to serve than the first. Netflix doesn't pay more because you rewatched a show. Salesforce doesn't pay more because you logged another opportunity. Revenue climbs; marginal cost stays nearly flat.

AI doesn't work that way. Every conversation consumes compute. Every generation has a cost. Every token produced has a marginal price attached to it — the first user and the millionth user both cost the company money, in a way traditional software fundamentally does not. Instead of revenue climbing while cost stays flat, AI businesses often see both climbing together, nearly in lockstep. That's exactly why I think we'll all end up tracking something like the Token Expenditure Index.

That's precisely why AI companies obsess over caching, prompt optimization, smaller specialized models, request routing, batching, distillation, and inference optimization. It's not micromanagement — it's a completely different business model wearing software's clothing.

A quick primer: what are you actually paying for?

If you're a product leader adopting AI, you don't need to derive backpropagation equations, but a few concepts are worth actually knowing:

• Architecture — the blueprint (e.g., the Transformer). Fixed, published, mostly shared across the industry.
• Weights — everything the model learned during training: billions or trillions of adjusted parameters that encode the accumulated pattern of essentially the internet's books, code, and conversations, compressed into numbers. When someone says "the Llama weights," they mean the learned intelligence itself — not the code that runs it.
• Prompt — your instructions to the model at run time.
• Context window — its working memory for the current conversation.
• Open vs. closed weights — whether that learned intelligence is published for anyone to run, or kept behind an API.
• Fine-tuning / distillation / quantization — ways of specializing or shrinking a model for a specific job or budget.
• Inference — the act of actually running the model to produce an answer; this is the step that costs money every single time.

A rough analogy: the architecture is like Photoshop's executable; the weights are like the accumulated skill baked into a master editor's muscle memory; the prompt is your creative brief. You don't need to write Photoshop's source code to use it well — but you get much further if you understand what's actually happening when you click "generate."

The conclusion that follows

The winners in this next wave won't necessarily have the smartest AI. They'll have the lowest cost per useful answer. Amazon didn't win retail through the flashiest storefront — it won through operational efficiency. AI companies, and the businesses that build on top of them, may win the same way: not with the biggest model, but with the most efficient system wrapped around it.

AI is a bicycle for the mind. Just remember — unlike a bicycle, this one has a metered toll on every mile.