AI tools might be the most powerful discovery engines we’ve ever built.
Lately I’ve been thinking about something curious.
When developers use AI coding tools like OpenAI Codex or Claude Code, the workflow often looks like this:
idea → prompt → generated code → tweak the prompt → try again.
People sometimes call it vibecoding — a term Andrej Karpathy coined for the new kind of coding where “you fully give in to the vibes, embrace exponentials, and forget that the code even exists.”
What starts as a quick test often turns into an hour of exploration.
Why?
One possible explanation is that these tools create discovery loops.
Every prompt is a small experiment. Most results are ordinary. But occasionally the system produces something surprisingly good — a clever function, a better structure, an unexpected shortcut.
Those moments are small rewards.
They make you think: maybe the next prompt will be even better.
The same pattern appears even more clearly in image generators like Midjourney.
You write a prompt. You get a grid of images. You pick one. Generate variations. Repeat.
Most images are decent. Some are strange. Occasionally one is amazing.
And suddenly you’ve been experimenting for an hour.
This pattern isn’t new in the abstract. Others have called it the slot machine of generative AI — type, generate, refresh, hope. Behavioral psychology has a name for it too: a variable reward schedule, the most engagement-inducing reinforcement pattern there is. What’s new is the substrate: instead of pulling a lever, you’re producing code, design, and ideas.
What generative AI might really be doing is collapsing the distance between effort and reward.
Before AI:
idea → hours of work → result
With AI:
idea → prompt → result
Seconds.
When the feedback loop becomes that fast, exploration accelerates dramatically. Bret Victor argued years ago — in Inventing on Principle — that creators need an immediate connection to what they’re creating, and that any delay between thought and result kills a whole world of ideas. Generative AI just took that principle to a new medium.
So maybe generative AI tools are not just productivity tools.
They’re something closer to interactive discovery engines — systems that let us explore huge spaces of ideas, code, and images in real time.
Most of the time you find something ordinary.
But every now and then you discover something unexpectedly good.
And that possibility is what makes it so hard to stop.
References:
- Andrej Karpathy — original vibe coding tweet (Feb 2025)
- Bret Victor — Inventing on Principle (2012)
- Ruben Hassid — AI is a slot machine
- Andrei Zimin — Generative AI & Bret Victor’s Immediate Feedback Principle
- Harnessing generative AI: cognitive engagement and reward sensitivity through reinforcement theory — ScienceDirect, 2025
- Human Learning by Model Feedback: The Dynamics of Iterative Prompting with Midjourney — arXiv:2311.12131