The first thing you learn in juggling is not juggling.
You learn to throw one ball from one hand to the other and catch it cleanly. You do this until the throw is consistent - same height every time, same arc, same landing spot in the opposite hand. You do not move on until that is automatic.
This is not the warm-up. This is the foundation. The three-ball cascade is just that throw, repeated in rhythm, with two more balls layered on top. If the single throw is off, adding balls does not fix it. It only makes the problem worse faster.
The temptation to skip ahead
Every learner wants to add balls before they are ready.
It feels slow to practice one ball when the goal is three. It looks like nothing is happening. The skill you actually want is visible on the other side of a step that feels too basic to take seriously.
This happens in AI as well.
People reach for the most powerful model, the most complex architecture, the most ambitious agent design - before they have spent time with a single prompt. Before they understand what the model is actually doing. Before they have any mental model of what “input” and “output” mean in practice.
The sophisticated thing is to start simple. Not because you will stay there, but because everything else is built on top of it.
What the brain is doing
There is a reason juggling starts with one ball and not three.
Studies from Oxford and Regensburg documented what happens in the brain during a period of juggling practice: grey matter increases in areas responsible for processing visual motion and coordinating action. The brain physically changes. But it only changes in response to the right level of challenge - enough to require attention, not so much that it cannot form reliable patterns.
One ball is the right level to start with. Two is a step. Three is the reward for having made the earlier steps automatic.
Machine learning follows the same curve. Your mental model of how a model processes language - the intuition you build from experimenting with basic prompts, from seeing what it gets right and what it misses - that is grey matter. That is a structural change in how you think. It cannot be shortcut by reading a technical paper. It has to come from doing.
Starting with one ball in AI
The equivalent of one ball is calling a model with a clear prompt and observing the output.
Not building a pipeline. Not configuring agents. Not deploying infrastructure. Just: here is a question, here is the response, what is actually happening here?
This is the throw-and-catch. Get it reliable before you add anything else.
The model is not magic. It is also not simple. The best way to build an accurate mental model of what it does is to use it - repeatedly, on small tasks, with enough attention to notice what surprises you and why. That noticing is the foundation. It makes every more complex thing you build later more intentional and less fragile.
The sophisticated thing is to start simple. Not because you will stay there, but because everything else is built on top of it.
The cascade is three throws, not one trick
When juggling three balls finally clicks, it does not feel like a new skill. It feels like what was always there, expressed at the right tempo.
The pattern was already in the single throw. The timing was already in the two-ball practice. Adding the third ball was just the moment when the foundation became visible as a complete thing.
Building with AI has a similar structure. The agent that handles a real workload - the system that adapts to new inputs and recovers from failures - is not a different kind of thing from the first prompt you got working. It is the same understanding, extended and trusted and given more to do.
Start with one ball. Make the throw clean. The rest will follow.
Related reading: The Loop That Rewires You - on myelination and why repetition is not optional. Adding the Fourth Ball - on what happens when you add scale to a foundation that is not ready.