I Was Being Cheap With Intelligence

Yesterday I was watching some Y Combinator videos, and what struck me wasn’t just the advice. It was the perception field they seem to live inside.

There is this engineering euphoria around them. A sense that, yes, a lot of jobs are probably going to be displaced, but also: look what we can build now. Look how fast this can go. Look how much more leverage a single person can have.

They are not just using AI as a tool. They are trying to make everyone else build inside that same assumption.

One idea that stuck with me was “token maxing”: using as many tokens as possible. Letting the model run. Letting agents explore. Not trying to squeeze intelligence through the smallest possible pipe.

And it made me realize something slightly embarrassing.

I have still been using AI like intelligence is scarce.

I have been careful with tokens. Conservative with runs. Trying to be efficient before I have even let the system breathe. Some part of me is still emotionally pricing this stuff like human labor, when the actual cost is fractionally cheap compared to a person’s time.

I should not be such a cheap bastard with intelligence.

If intelligence is becoming abundant, then the bottleneck shifts. It is not “can I get something to do the work?” It is aim. Taste. Clarity of direction. Review. Iteration. The ability to know what problem is worth pointing all this energy at.

That changes the feeling of work.

Maybe my job is less to sit down and grind through every step. Maybe my job is to sit, have a thought, aim it properly, and then put it into massive disproportionate action through agents.

That action does not always feel like action. It can feel like watching. Reviewing. Redirecting. Running another loop. Asking for five versions. Asking a critic to attack the best one. Asking another agent to turn it into a system. Asking another to make it clearer. Asking another to find what I am missing.

It feels strangely passive compared to old work.

But the output can be much larger.

This is the part I am still metabolizing: volume is no longer coupled to time in the same way. You can run more variations, more reviews, more drafts, more tests, more parallel attempts. Not perfectly. Not magically. But enough that the old instinct to conserve effort starts to become a liability.

The scarce thing is not the tokens.

The scarce thing is knowing what loop to run.

There is still a deep kind of clarity required. Maybe deeper than before. You need to understand the problem. You need to know what good looks like. You need to know when the output is fake, thin, generic, or over-smoothed. AI does not remove taste. It makes taste more exposed.

But maybe clarity also arrives through messy building. Run the bad version. Test it. Let it break. Ask for alternatives. Compare them. Do more of what works. Make the loop better.

This is what skills and agents make possible. Small processes. Repeatable workflows. Little systems that turn a thought into drafts, critique, images, posts, experiments, and decisions.

And the weird thing is how little permission is needed.

You can just start making them.

So now I am thinking about ROI per token. Not just cost per token, but return per token. Which loops produce more clarity, better taste, better output, better business results? Which uses of intelligence compound?

Because if the return is high enough, then token maxing is not waste. It is how you stop starving the process.

A different world is coming.

Actually, it is already here.

And it is mad.