This is a piece I originally posted to my Instagram story in March 2026.

I was recently struck by an interview with Terence Tao. He acknowledged that the cases of AI solving mathematical problems are multiplying, but argued that this doesn’t immediately translate into mathematical progress. It’s true that some problems have been solved, but the overall success rate is still low, and what matters most is whether those solutions generate new concepts or new understanding.

He points out that today’s AI tools are extremely strong at broadly applying existing techniques, yet weak at accumulating incremental progress or setting new directions. And he sums up the difference simply: the literature can grow richer, but it doesn’t grow deeper.

This distinction isn’t unique to mathematics.

Over the past few years, the technical barrier to making video has fallen dramatically. In the past, producing footage of any real quality required equipment, manpower, time, and skill. That’s no longer the case. Anyone can now achieve a certain baseline of visual polish, and by the standards of the past, that baseline is quite high.

This shift looks like a democratization of expression. But at the same time, something else happens. As the floor of quality rises, quality itself stops being a meaningful basis for distinction. “Well made” becomes a less and less persuasive verdict.

This is where the question of taste enters. The taste I mean here isn’t simple stylistic preference; it’s closer to the criterion that decides what to keep and what to throw away. In the past, production cost itself acted as a kind of filter. Work built on poor judgment rarely got made in the first place, precisely because it cost so much to produce. But now that the filter is gone, the problem of judgment surfaces far more directly.

The output of AI video tools illustrates this shift well. Most of what they produce is technically hard to fault. The composition, the color, the transitions, the rhythm are all stable. Yet in many cases they leave nothing behind. Because there’s nothing wrong with them, there’s no reason to remember them either.

This kind of output is very well suited to certain uses. Public-agency promotional videos, for instance. The message is clear, the tone is steady, there’s nothing to cause offense, and nothing that grates on anyone. It performs exactly the function required. But nothing beyond that.

Because such videos are designed not to fail, the reasons they might succeed are stripped away along with the risk. They don’t touch emotion too forcefully, they leave no room for interpretation, they don’t push any particular point of view. What’s left, in the end, is only the information that was conveyed.

This phenomenon can be seen as expression rapidly converging on the average. AI excels at combining and optimizing previously successful patterns. The results are stable and reproducible. But they are also, at the same time, replaceable. It’s expression with the risk removed.

The trouble is that most meaningful work happens precisely where it departs from that average. Important works often contain choices that are hard to explain, inefficient structures, or elements that seem unnecessary. Those very elements are naturally eliminated in an optimized generative process. And yet artistic value forms in exactly that place.

This is close to the distinction film theory draws between continuity editing and montage. An optimized flow aids comprehension, but it doesn’t generate meaning. Meaning arises more often where one scene collides with another: cheerful music playing over a deeply sorrowful scene, an unnecessarily long silence stretching through a tense situation, the eye drifting to some trivial detail at the very moment emotion peaks.

At this point the nature of the choice becomes clear. It’s not simply about generating more scenes; it shifts to the question of which scenes to set side by side, which flows to deliberately break, which collisions to allow. In other words, it’s not a problem of production but of editing, and within that, a problem of montage-like choice.

Returning to Tao’s distinction: today’s tools can make expression richer, but they don’t make these choices for us. They can supply more material, but deciding what relationships to forge among that material is still the work of a human exercising judgment.

So the problem lies not in the technology but in the choice. The judgment of what to make, what to discard, and what to push all the way through is not handed to you from the outside. A tool can carry out the execution, but it won’t take over the criteria for choosing.

In this context, saying that the threshold of taste doesn’t drop is not mere stubbornness. Technology has advanced quickly, and the difficulty of production has fallen dramatically. But the standard for what counts as a meaningful result doesn’t automatically drop to match it.

If anything, the situation is closer to the opposite. The options have multiplied, and with them the responsibility of choosing surfaces more directly. The filtering that technical constraints used to do for us now has to be performed by individual taste, directly.

In the end, the question that remains is the same. It’s not what you can make, but what you judge worth making. And that standard still forms slowly.

Technology will go on producing more and more scenes, but the work of deciding which scenes to collide, where to break the flow, and what to leave behind still remains. And that choice, I think, is what will matter most in the work to come.