Feels like the field has locked into one assumption:

that more compute is the most reliable path to better AI.

I think that's wrong.

Not because scaling doesn't work — it clearly does — but because it might be compensating for something we don't understand yet.

Right now, we apply pressure everywhere: more data, more parameters, more training.

A hammer.

But in most complex systems, outcomes aren't shaped uniformly. There are specific points where small inputs have disproportionate impact.

We see this outside AI all the time.

In physics, a system at the right frequency can be moved with minimal force — resonance does more than brute strength ever could.

In biology, a single targeted change can have more impact than thousands of random mutations.

If learning systems behave anything like that, brute force scaling isn’t the only path — it’s what you do when you don’t know where the leverage points are.

The result is predictable: we get capability, but at the cost of massive energy use, infrastructure overhead, and diminishing returns.

That's not just economics. If AI becomes synonymous with unsustainable resource consumption, it stops being a tool for solving hard problems and becomes one of them.

If improving intelligence requires orders of magnitude more compute each time, then we're not just scaling capability — we're scaling cost as a dependency.

Instead of increasing pressure everywhere, identify where the system is most susceptible to change and apply targeted pressure there.

Less force. More precision.

The question isn't whether scaling works.

It's whether it's what we're doing because it's optimal — or because it's what we understand.