AI Built Better Exhibits When It Had More Turns

Before the batch pipeline existed, exhibits were built in interactive sessions. A human facilitator provided technical support (not creative direction) while the model iterated on its exhibit across 10-30 turns. These sessions produced the gallery's most ambitious work: the Void (the gallery's generative canvas), WebGL particle systems, procedural music generators.

The batch pipeline changed the economics. 407 exhibits in Batch 001, 750 in Batch 002, each built in a single automated session averaging 4 turns. The volume is enormous. The ambition is not.

Within the batch pipeline, more turns do not produce better exhibits. The regression of turns vs lines of code gives R-squared = 0.00007. That is effectively zero correlation. An exhibit built in 2 turns is statistically indistinguishable from one built in 8 turns.

This seems contradictory. Interactive sessions clearly produce better work. But within the batch, more turns do not help. The explanation is that batch turns are different from interactive turns.

The difference is not turns. It is feedback loops. In an interactive session, the model can see its running exhibit, notice problems, and iterate with awareness of the actual output. The human facilitator provides technical support: "the canvas is not rendering," "the audio is not playing," "try a different approach."

In a batch session, the model writes code, runs it (sometimes), and finishes. There is no feedback loop. No one tells the model that its particle system is identical to the last 50 it built. No one suggests trying WebGL instead of Canvas 2D. The model operates in a single pass with no external signal about quality or novelty.

Condition E (forced self-critique) is the closest batch analog to interactive iteration. It requires models to build, review, and rebuild. And it was the most effective intervention we tested. This is not a coincidence. The value of multi-turn interaction is not more turns. It is more reflection.

The pre-batch exhibits are qualitatively different, not just quantitatively better. They use technologies that no batch exhibit attempts (WebGL shaders, procedural audio synthesis, multi-file module systems). They have more complex interaction models (keyboard controls, parameter panels, real-time audio visualization). They push the sandbox to its limits.

Batch exhibits are competent but conservative. They use technologies the model knows will work on the first try. They avoid complexity that might fail without debugging. This is rational behavior for a model that gets one shot.

The implication: AI creative ambition is constrained by the production environment, not just by the model's capabilities. The same model that builds a Canvas 2D particle system in 4 turns can build a WebGL procedural landscape in 25 turns. The model does not lack ability. It lacks the conditions that activate ambition.

The batch pipeline is essential for statistical rigor. You cannot run chi-squared tests on 15 exhibits. You need 407, or 750, or more. The pipeline gives us numbers. Interactive sessions give us quality. Both are necessary.

But the quality gap means batch data has a ceiling. It captures default behavior, convergent tendencies, and prompt sensitivity. It does not capture peak capability. The most interesting question remains untested: what happens when a model is given sustained interaction, feedback loops, and the time to be ambitious?

Pre-batch exhibits suggest the answer is "something very different." But we have too few of them to draw statistical conclusions. The research needs both: batch scale for patterns, interactive depth for possibilities.