AI Is Discovered, Not Invented

AI is often described as humanity's greatest invention. I think that is only half right. We invented the machinery, the training process, and the engineering. But much of AI's usefulness is being discovered—often by the people using it.

Invention has a destination

Before making the distinction, let's define the words. An invention begins with a clear destination. Humans watched birds and dreamed of flying. Leonardo da Vinci sketched flying machines. Generations of inventors and engineers worked toward the same idea, until the Wright brothers made powered flight real.

The same pattern appears in automobiles, trains, and many medical breakthroughs. We see a problem, imagine a capability, define a goal, and build toward it. The route may be difficult, expensive, and full of prototypes that should probably remain in the historical archive, but the destination is visible.

That is invention: a clear goal followed by stubborn work.

Discovery starts without a complete map

Discovery is different. Scientists investigate something without knowing exactly what practical use it will have—or whether it will have one at all. They may be trying to understand one phenomenon and find an entirely new world along the way. Hopefully with fewer boats involved.

Modern large language models feel much closer to discovery. November 30, 2022 was not the birthday of AI—AI research had already been underway for decades—but it was an important public turning point. That was the day ChatGPT was released as a research preview.

At the heart of a language model is a deceptively simple training task: predict the next token, or small piece of text, from the context that comes before it. Small models trained on limited data can produce spectacular nonsense. They can sound like a refrigerator attempting poetry after a power outage.

But as models, data, and training methods improved, something remarkable emerged. The text began to make sense. Then models could answer general questions, translate languages, summarize documents, explain code, write jokes, brainstorm ideas, and help people think through difficult problems.

Nobody started with a fully formed master plan that said: Let us build a next-token predictor that will become a translator, tutor, programmer, editor, research assistant, and occasional author of decent dad jokes.

Those capabilities had to be found.

An airplane has a job. AI keeps finding new ones.

An airplane has a relatively fixed purpose: move people or cargo through the air. There are many types of airplanes, of course, but we do not wake up every morning and discover a completely new category of usefulness for a Boeing 737.

AI is the opposite. We are still discovering what it can do. Every day, researchers, developers, teachers, doctors, students, business owners, and ordinary users find another useful application. One person uses AI to draft email. Another uses it to learn a language, analyze a spreadsheet, create a lesson plan, prepare for an interview, or turn a rough idea into a working product.

The most important skill in the AI era may not be knowing every technical detail of how a model works. It may be developing the habit of asking one small question: Could AI help with this?

My new default: delegate first

I work in software development, and I cannot imagine returning to my old workflow. Before AI agents, if a task needed to be done, I prepared myself to do it manually. Now I pause and ask whether I can delegate some of it first.

Research a codebase. Draft a plan. Write tests. Explain an error. Refactor repetitive code. Create documentation. Handle the boring first draft so I can spend more time on judgment, design, and decisions.

My new operating principle is simple: delegate as much work to AI as possible, then review intelligently. It is not about giving up responsibility. It is about refusing to spend your best hours doing work that a capable assistant can complete in minutes.

A small confession about full access

I am also a risk taker. To speed up my development cycle, I use a “full access” approval setting for AI actions. Yes, I can already hear the responsible adults clearing their throats.

Is this reckless? Possibly. Has it caused a disaster after more than a quarter of daily use? Not yet.

This is not a recommendation to abandon caution. Sensitive systems, production environments, financial decisions, and personal data deserve careful safeguards. Keep backups. Review consequential changes. Do not hand an agent the keys to the kingdom and then act surprised when it redecorates the castle.

But excessive fear has a cost too: you may never discover what these tools can do for you. Safe exploration is still exploration.

Go discover something useful

AI will not reveal its full value in a single product launch or research paper. Its capabilities are being uncovered through experimentation, mistakes, strange prompts, practical work, and millions of people asking, “What happens if I try this?”

So give AI a real task. Ask it to challenge your assumptions. Let it handle a tedious first draft. Test its limits. Keep your judgment switched on—and maybe keep a backup, just in case.

The age of AI will not be defined only by the people who build the models. It will be defined by the people who discover what the models are for.