Toy 05 · Thesis

The Robot Has No Taste

"AI doesn't have taste. The actual creatives do."

Monthly · one round

A monthly round where humans and AI models score the same game-feel clips on an 8-axis rubric, plus a running leaderboard of where models agree with us and where they miss.

Surface vibeA game show: humans vs. robots, and someone loses.
Profound layerWhere taste actually lives, and which creative decisions you should and shouldn't hand to AI.

The rubric

Every clip, human or robot, gets scored on the same eight dimensions, 0 to 4 each. The anchors are published so anyone can argue with a score. The robot's job is to match the humans. The humans' job is to be right.

#DimensionWhat a 4 looks like
1Impact
Does a hit land?
The world flinches. Freeze-frame, punch, flash. You feel it in your hands.
2Anticipation & Settle
Is the motion authored?
Wind-up, overshoot, settle. Motion has weight and follow-through.
3Theatre
Is the payoff staged?
Owns a region of the screen and, at full magnitude, breaks its own rules.
4Density
How much feedback per second?
One action cascades into staggered, per-element responses.
5Audio Punch
Does it sound like it feels?
Sharp attacks, wide dynamic range, pitch that carries information.
6Restraint
Does it know when to shut up?
Quiet baseline, violent peak. Saves its voice for what matters.
7Permanence
Does the world remember what you did?
Consequences persist and accumulate. You can read your own history in the level.
8Tactility
How does it feel under the finger?
Can't be judged from video. Scored only from hands-on.

Rounds

Round 00 · 2026-10-14 · Human panel only. No robots scored yet.

Calibration: setting the ceiling and the floor

Before any model gets a turn, the human panel scores two anchors so every later round has a top and a bottom. Balatro's scoring sequence is the ceiling: scale-punch plus hitstop, almost no screenshake, two rising pitch runs in eight seconds. Wordscapes' mid-level solving is the floor: competent, successful, inert, with roughly 35 times less motion. The instructive kind of dead.

GameClipRoleHuman panel
BalatroBoss blind score cascadeCeilingImpact 4 · Density 4 · Audio 4 · Theatre 4 · Restraint 3
WordscapesMid-level word solveFloorImpact 1 · Density 1 · Audio 1 · Theatre 0 · Restraint 2
Vlambeer screenshake demoThe 30-step ablation, scored step by stepReference ladderScored per step; used to calibrate the 0-2-4 anchors

The take. Taste starts with knowing what the top and bottom look like. Most arguments about feel are really arguments about which anchor someone has in their head.

Leaderboard

#ScorerAgreement with humansBiggest miss
-Human panelbaselineThe panel is the ruler. Its misses show up as arguments in the comments.
1Model A (frontier, video in)Round 1Not yet run
2Model B (frontier, frames + audio)Round 1Not yet run
3Model C (open weights)Round 1Not yet run

Next. Round 1: ten clips, three models, one human panel. Results published whoever wins, including the robot.

Get round results

Newsletter provider not chosen yet. Nothing is collected.